Source: mcp_surface.md #

View this page as Markdown

On this page

Generated automatically from the published contract sources.

Source path: mcp_surface.md.

This source leaves out the rows and paragraphs that name an action only platform staff can call.

Contract text #

The MCP Surface — Contract #

Purpose #

This file is the contract for the Turn Zero Cloud MCP server's surface: how a customer's AI coding tool connects and authenticates, which kinds of capability surface as tools, resources, and prompts, and what the launch surface is. It settles the launch surface on the launch definition's own rule — the scenarios drive, the API guarantees (CHI-L0-02). It states the surface of a *client*: every capability here is the one API's (API-L0-01), reached through the MCP server acting as the connected identity (API-L0-02), and nothing here is MCP-only.

The launch surface's enumeration has one home: the catalog at schemas/mcp_surface.json beside this document — capability by capability, each with its form and its authority tier — held coherent by a test in the platform's suite. This document states the rules the catalog obeys and deliberately repeats none of its rows.

What this document does not cover: the API's own resource contract and versioning (management_api_contract.md); the pending action's mechanics (API-L0-07 owns them; this surface only reaches them); and the super-admin's capability inventory (MAPI-09's grant rows). Nor does it cover the launch skill set, whose rows are the contract's in schemas/skills.json under API-L0-10's derivation rule.

Requirements are numbered MCP-nn.

The server and the connection #

MCP-01 (decided): The Turn Zero Cloud MCP server is one remote server, hosted and operated by Turn Zero Cloud, reached over the Model Context Protocol's streamable HTTP transport at an address the customer adds to their tool. One connection reaches the products the account holds that the reached address lists (MCP-14). The address alone selects that set; no product choice is read from the connection or its query. It needs no local install, no sidecar, and nothing on the customer's machine beyond their own AI coding tool.

The estate's configured origins serve four MCP addresses: its own origin, tzCloud's Hosting address, tzTools, and the operators' address (MCP-02). Hosting lists tzCloud and the shared account tools; tzTools lists tzTrack, tzMesh and the shared account tools. The company's origin lists every customer tool and no operators' tool, to staff as to any other account. The operators' address lists its own tools under MCP-02's admission. An address is served only where its origin setting is set. Every address declares the same server name and description and answers the same handshake.

The server declares itself turnzero at every address; that is its protocol identity. Published install stanzas use turnzero for the company's origin, turnzero-hosting for tzCloud's Hosting address, and turnzero-tools for tzTools. No install stanza is published for the operators' address. A client prefixes a server's tools with the name it is bound under, so the stanza name is the word a customer's assistant reads on every action. The declared name spells the company's name, Turn Zero (azure:AZD-08), since the server fronts the products the account holds (MCP-14). It identifies neither one product, the host, nor the workspace's internal code name. The declared and stanza names use lowercase letters, digits, and hyphens. Some clients' tool-name grammars exclude dots, so names with domain punctuation are not addressable everywhere the product is installed.

Its handshake also gives the server's description, AI backend: host, talk, track, the short line a host shows beside the name in its connector list. The description names no product and no requirement, and is the same on every estate.

What a customer's own tool stores as the label for the connection is that customer's to choose and no requirement's to impose. This statement fixes the server's declared name and each published stanza's name, which are what Turn Zero controls here. Each stays apart from the name of every other MCP server Turn Zero publishes, each product's door and tzTeam's server among them. Each is a stable identifier on MCP-08's terms. A later change therefore reaches every affected address and published stanza in one release that serves the new name, never a quiet edit.

Which protocol revision the server speaks, and how it negotiates with older clients, is versioned call configuration beneath this contract, not a statement in it. Every request at the MCP route is answered as JSON. Those answers declare charset=utf-8 in their Content-Type, so a client that applies a default charset to a bare media type decodes the server's UTF-8 bytes as written.

MCP-02 (decided; conforms: ACC-04): The authentication binding is OAuth 2.1 authorization code with PKCE, discoverable by the standard authorization-server metadata, with dynamic client registration so a customer's tool connects without a preconfigured client identity. The authorization moment is the sign-in page (WEB-L0-12): the builder realm's own surface (ACS-L0-01), served on the control plane's origin (ACS-L0-04), and the one browser interruption of CHI-L0-03, which the computer login page opens too, as a command-initiated case of that moment (WEB-L0-25). There a first sign-in creates the account where the builder realm's creation policy admits it (ACB-L0-77; ACS-L0-07) and is refused by name where it does not. The token the flow issues is the builder realm's session (ACS-L0-11), the credential the authority tiers bind to (API-L0-06). The connected session acts as that identity and nothing else, and disconnecting or signing out revokes the session rather than orphaning it.

A client_id no registration names, one the client sweep deleted or a development bake start dropped, is refused invalid_client. The token endpoint answers 401 before it reads the grant, so a host that clears its stored client on that refusal registers again. The authorize route answers the person's browser a page that says to remove the tool's sign-in for this server and connect again.

A minted token (API-L0-05) connects the same server as its bearer with no browser moment, the sign-in flow never run for it. The connected session then acts as the token's own bounded identity under the same challenge. The platform credential connects nothing here and draws the standing challenge (SEC-L0-07).

The authorized computer's token the turnzero command keeps (PLD-L0-104) is minted under the session this sign-in opened, by the computer login page's mint under the browser session or by the tool's mint_token call (MAPI-14; WEB-L0-25). Each is bound to the command's challenge and spent by its wire exchange (MAPI-23). The command holds no sign-in client of its own. login opens the browser on the platform's own page, and login --device prints a link to it for a person to open on any device; the page's cookieless arrival is the sign-in page returning to it.

Deployment configuration fixes one canonical authorization-server origin through OAUTH_ORIGIN and one protected-service resource identifier per MCP address; neither is inferred from a raw Host header or widened by an alias. The deployment template sets the origin from its trusted public API origin. An explicitly configured origin is validated before database migration or provisioning; an empty or malformed value refuses startup. Startup also refuses a missing OAUTH_ORIGIN when PGHOST has a nonempty value. Without OAUTH_ORIGIN or that deployment setting, a local executable selects the loopback origin from its actual bound listener, including an assigned ephemeral port. An application factory with no listener defaults to http://127.0.0.1:8080 unless its caller supplies the trusted origin.

Each set origin serves its own /mcp resource: OAUTH_ORIGIN, TZCLOUD_API_ORIGIN for Hosting, TZTOOLS_ORIGIN for tzTools, and TZOP_ORIGIN for the operators' address. The reached host is the lower-case host the origin lock verified, matched exactly against those settings. A prefix, suffix or trailing dot matches no address. Two address settings naming one host, or one naming OAUTH_ORIGIN's host, refuse startup. On each address's host, protected-resource metadata names that address's resource, the canonical authorization-server issuer, and the scopes tzcloud then offline_access. Sign-in stays on the authorization-server origin.

Changing the origin makes prior-origin builder OAuth access and refresh grants ineligible and requires new OAuth authorization; other credential kinds keep their own validity rules, and an existing valid browser session may authorize the new grant without another human sign-in. Each address's identifier names the same Turn Zero Cloud protected service: the MCP endpoint and the API, documentation and data-plane surfaces already admitting this bearer under ACS-L0-11; it grants no access to another service. Protected-resource metadata names the reached address's resource and the canonical issuer; authorization-server metadata repeats that issuer and its canonical endpoints. Both publish the supported OAuth scopes. Generated callback and return URLs whose destination is the control plane use its configured origin; the client's redirect remains its exact registered URI.

Authorization, code exchange, access-token records and refresh-family records preserve the resource, issuer, client and granted scopes. A supplied resource must match exactly; omission selects the configured service at authorization and retains the recorded resource at exchange or renewal, never a wildcard. The scope names and default are versioned call configuration. OAuth scopes are separate from authority tiers and application grants and cannot create either. The shared credential resolver rejects unbound or foreign OAuth access credentials before dispatch; a protected call with a valid OAuth credential lacking its required scope answers insufficient_scope with HTTP 403.

This paragraph qualifies the paragraph above's exact match for the plane's own resource. Wherever a supplied resource is compared with the plane's own configured resource, a value equal to it but for a query holding products alone reads as that resource and is recorded as it. That reading is no alias: the identifier recorded, advertised, and bound stays the one configured resource.

An explicitly supplied address resource matches exactly and carries no query; another value is refused invalid_target. Omission at authorization still selects the plane's own resource. Exchange and renewal retain the recorded resource, and naming another is refused invalid_target. The shared credential resolver admits a session bound to any configured MCP address on the REST and data routes, at renewal, and at the front's product-token exchange.

At /mcp alone, a session's resource must equal the reached address's resource, the company's origin included. A mismatch answers the standing 401 authentication_required challenge with invalid_token, the same body as a presented credential resolving to nothing, and the reached address's metadata. Every 401 and the 403 insufficient_scope on that route name the reached address's metadata. A minted token carries no resource and meets no comparison.

The operators' address admits only a credential holding a grant that a catalog row naming tzop carries as grant or names in admits, except device_sign_in. A row with no mark admits nobody by this rule. Every other credential meets the same 401 before any tool runs. Sign-in for that resource refuses an account holding none of those grants with access_denied and its state, minting no code and writing no tool_authorized event. The exchange reads the grants again and refuses invalid_grant if the account no longer holds one.

A session-resource request's scope is read the same way at authorization, exchange and renewal. An omitted or empty scope selects the documented default at authorization and retains the grant thereafter. An earlier name of the session scope, turnzero_cloud or appcloud, is read as tzcloud, while a grant recorded under one stays refused. A scope of another resource the plane serves, tzmesh or tztrack, is left out of the grant where the session's scope is also named (RFC 6749 section 3.3). So a client asking for the authorization server's whole advertised list is granted tzcloud alone, and the token answer names it. A scope this plane never served, an attempted broadening, or a product's scope alone is refused invalid_scope.

The scope offline_access is admitted beside a granted scope on every resource the authorization server serves, at authorization, code exchange, renewal and device authorization. It is left out of the grant, binding, refresh record and session, and never appears in a token answer's scope. Named alone, it is refused invalid_scope because it leaves nothing granted. Authorization-server metadata lists tzcloud, then tzmesh and tztrack where served, then offline_access. The plane's protected-resource metadata lists tzcloud and offline_access, in that order. The challenge's scope="tzcloud" and MCP-10's refresh-token rules stand unchanged.

Every MCP request the plane admits, presenting no credential or one resolving to no account, initialize included, answers 401 and a WWW-Authenticate challenge naming the protected-resource metadata and applicable scopes. An invalid presented credential also carries invalid_token and its error description; missing credentials carry none; the wire's library rows challenge as the context reads do (MAPI-11). The challenge's JSON body states in its detail the same case the header distinguishes: a connection that presented no credential, which reconnects with sign-in or starts the OAuth flow, or a presented credential that resolved to no account, which signs in again. Each detail says nothing was done, names the person's sign-in and then the call made again, and names turnzero login, the turnzero command's own sign-in (PLD-L0-104). Neither adds OAuth error information to the header, and the wire route's refusal states the same two texts.

This paragraph qualifies the paragraph above for a presented bearer of a grant's form that resolves to no credential. The challenge's body and the wire route's refusal state a third text in that case: the platform serves no grant of that value. The text names both ways to a fresh grant: the tool call whose answer carried it, made again, and the application's own request for one, made again (MAPI-03). The name, the status, and the header stand as the paragraph above states them for a presented credential.

Every body the server's 401 challenge answers carries help, the address of its refusal's row on the refusals page (MAPI-10): authentication_required, end_user_credential_not_admitted, transfer_grant_not_admitted, secret_grant_not_admitted, egress_key_not_admitted, and app_credential_not_admitted. Browser-only approval and all other credential-kind boundaries remain MAPI-03 and API-L0-07's.

Registration serves public clients with token_endpoint_auth_method=none, each naming at least one redirect. A client whose grant_types is the device-code grant alone may name none, and so never authorizes at /oauth/authorize. Authorization accepts only an exact registered HTTPS redirect or HTTP loopback redirect, with no URI credentials or fragment; code exchange requires S256 PKCE. An additive migration retains old rows as explicitly unbound. It may backfill only where recorded issuance provenance proves the binding, and production cutover otherwise requires the developer-approved one-time reauthentication disposition; an unbound row never means all audiences. The authorization-server metadata adds revocation_endpoint, the issuer's /oauth/revoke, and revocation_endpoint_auth_methods_supported with none alone. Public clients present no secret.

POST /oauth/revoke takes JSON or a form-encoded body carrying token, with client_id and token_type_hint optional. A named client must be registered, or the endpoint refuses 400 invalid_client before reading the token; without a name it reads the client from the token's row. Whoever holds the connection's refresh or access credential can end that connection. The hint orders the two lookups and never refuses. A current or redeemed refresh credential ends its family while that family has a live row, however many renewals the stored copy lags. A live builder session's access credential names its connection through the session's OAuth binding and its row's family. Without either, the session carries no connection.

Ending its own connection answers 200 with contract_version and revoked true; API-L0-21 states the end. An unknown or ended token, another client's under the client named, a browser session's value, and no token all answer the same 200 body, with revoked false. This other-client answer departs from RFC 7009 section 2.1, as the end-user realm's does (ACS-L0-14). A platform credential's prefix alone draws 400 unsupported_token_type, without checking validity: minted and product tokens, end-user sessions, transfer grants and egress keys.

Anyone may register a client, so a registration is bounded. It names at most ten redirects. Each is an HTTPS redirect, or an HTTP redirect on a loopback address, localhost, [::1], or an address in 127.0.0.0/8, at any port (RFC 8252). None carries URI credentials, a fragment, or a wildcard in its host. A registration outside these bounds is refused invalid_client_metadata and stores nothing.

No redirect names a host the plane serves. That is a hostname equal to, or under, the registrable domain of the authorization server's origin, or of the tzTrack origin where the plane names one. It matches at any scheme or port and with any trailing dots, so www. and the service hosts match as the apex does. Two kinds of hostname are excepted: the tzmesh origin's and each TZMESH_SITE_ORIGINS origin's exact hostname, and an application's hostname under the serving suffix. On a loopback development origin the rule is that origin exactly. Authorization reads these bounds again over the redirects a registration stored. Where the registration falls outside them it refuses invalid_request at any of its redirects, so one stored before the bounds no longer authorizes.

A source address registers at most twenty clients within the hour, and the next is refused rate_capped. Each registration writes one info line to the control plane's record naming the client, the hosts its redirects name, and its registered name where it gave one, and never the source address.

A token the retired first-party client turnzero-blueprint was issued has no expiry date. It ends at revoke_token, after 30 days unused (PLD-L0-67), at sign out everywhere, at a passkey change's revert (ACS-L0-10), and at the account's suspension (API-L0-12), the last three each counting it in exchange_tokens_revoked.

Thirty-one paragraphs, this one and the thirty after it, qualify five rules above for the protected resources of Turn Zero's own products. The five are the token the opening paragraph names, the scopes the metadata publishes, the default scope and the separation of scopes from grants, registration's exact redirect match, and the binding's one grant, which the device grant joins for a client with no browser. Each product is one row of the plane's product table, and the table holds two rows, tzmesh and tztrack (each member's own requirements). A row holds the product's technical name, which is its profile name, its OAuth scope, its signing realm, and the word its token prefix carries. It also holds the setting that names the product's origin, and the paths the accounts service serves for the product.

Where the plane's setting TZMESH_ORIGIN names the tzmesh origin, the tzmesh resource is that origin's /mcp, and the authorization-server metadata advertises the scope tzmesh beside tzcloud. The tztrack resource is the origin TZTRACK_ORIGIN names, its /mcp; its scope is tztrack and its exchange is the tzTrack paragraphs'. The metadata lists tzcloud first, then tzmesh and tztrack where served, and offline_access last. Where a product's setting is absent the plane names no origin for it: the metadata stands as it stood, and a request naming the product's resource is refused invalid_target as any foreign resource is. The plane's origin and its configured MCP-address hosts serve their own resource metadata; each product's origin serves its own.

A registered client asks for the tzmesh resource by naming it in resource at /oauth/authorize, with the scope tzmesh or none, or a builder's application with tzmesh.application, a scope the metadata does not advertise. A request also naming tzcloud or offline_access is granted the tzmesh scopes alone at authorization and exchange, as a session-resource request is granted the session's (RFC 6749 section 3.3); another scope is refused invalid_scope. After the sign-in moment a tool's request, one naming the scope tzmesh or none, holds no tzMesh page, and the paragraph below states its completion. The tzmesh site's own sign-in client holds no page and a builder's application one line at its first sign-in alone, as the paragraphs on each state, and a prefetch holds nothing (WEB-L0-24).

Where a tool's redirect is not a loopback address, its request meets the one consent screen a tool's sign-in shows (CHI-L0-03), whose Allow access completes it. Where its redirect is a loopback address, it completes at the sign-in's completion, or at the allow-access screen's Allow access where the browser's session signed in within WEB-L0-16's window (CHI-L0-03). At a tool's completion the plane mints the single-use code for the connection's tzmesh token and redirects it with the state to the client's registered redirect. The one screen names the account by its sign-in address and the tool by its redirect's hostname with no port, its line naming the tzMesh messages alone and none the applications, and its Cancel returns access_denied and the state and mints nothing. The Terms of Service's section 15, Each product, carries the agreement that the tool may send and read the person's tzMesh messages.

A tool's completion and the exchange read the account's profiles through the plane's one rule (ACB-L0-76). It answers tzmesh for each member of the invite group while the plane's setting invite_only_products names tzmesh, for every account while it does not, and for an operator account always. In these paragraphs an account in the group is one the rule answers tzmesh for. An account outside the group is returned to its client with access_denied and the state at the completion, with no description and no page, and is refused invalid_grant at the exchange; neither answer names the product.

The token endpoint exchanges that code for one signed tzmesh token at the level admin under the client's name, issued to the client, and never for a session (MAPI-14; MAPI-03 states the signed form). In the same act it revokes the client's own unrevoked tzmesh token of that label, and every other holder's token of the label stands. The answer carries the value once, with token_type, scope, level, label, expires_at, origin, resource, token_id, and replaced, the identities it revoked, under Cache-Control: no-store. It carries no refresh token and opens no connection (MCP-10). The exchange reads the generation again after the mint, and a token minted while it moved is ended, its end written as revoke_token's line, and never answered. The token expires 180 days after its mint and ends as API-L0-05 states for it.

The plane also serves the device authorization grant (RFC 8628) for its own resource and each configured customer MCP address, whose exchange answers the builder realm's session. It serves the grant for the tzmesh and tztrack resources too, each exchanging for a signed product token. A registered public client posts client_id, resource, and scope to /oauth/device_authorization. A session resource reads scope as a session-resource request does. A product resource admits its scope or none; a request also naming tzcloud or offline_access gets the product's scope alone, as at /oauth/authorize. The answer carries device_code, user_code, verification_uri, the plane's /oauth/device, verification_uri_complete, expires_in 600, and interval 5. Another resource, the operators' resource or none is refused invalid_target with no product word; another scope is refused invalid_scope.

The user code is eight characters from an alphabet of twenty consonants, printed in two hyphenated groups. The metadata advertises the endpoint and the grant on every estate, since the plane's own resource is always served; a product's resource is served where the plane names its origin.

The verification page is the plane's own sign-in moment for the request the code names, and holds the resource's rule after it. For a session resource that is the page alone, naming the platform and no product, and for tztrack the page alone, its gate read at the exchange. For tzmesh it is the group check (ACB-L0-76). The page shows the code and the client's registered name and offers Allow access and Cancel alone, Allow access taking no click until the pointer has moved over the page or a short delay has passed (CHI-L0-03). The page and its posts answer only the browser whose sign-in held them, no other site may frame them, and a prefetch holds nothing.

For each admitted resource, Allow access writes the approval onto the request, and Cancel denies it, a page that refuses the account writing nothing but offering Cancel; an unanswered request expires. A code the page does not know is told apart from no other. The token endpoint answers the poll as MAPI-14 states. The device code stands in PKCE's place: only the asking client holds it, and it is answered once, the request's row leaving at that answer; the daily pass deletes an expired one (PLD-L0-67).

For a session resource the exchange issues the builder realm's session with its refresh token, as the authorization code's exchange does (MCP-10; ACS-L0-11). The session's binding carries the bounding grant device_sign_in, which every row of its refresh family keeps through rotation and its grace. The session holds no super_admin, even for an operator's account, since that grant lifts every bound (MAPI-09). The bound admits it to the rows whose admits mark names the grant, the rows marked access: anonymous, and the forwarded rows MCP-14 states (MAPI-16). Every other row refuses it, an unmarked row device_sign_in_bounded and a grant-marked row grant_required. So a sign-in a script was tricked into deploys nothing, exports nothing, reads no storage area, mints no wider token, and destroys nothing without the person's approval in the browser (MCP-07).

The accounts service publishes each product's key set at /accounts/v0/<product>/keys, every key with its stamps, and its revocation list at /accounts/v0/<product>/revocations, each anonymous under the admission window (PLD-L0-67). The tzmesh pair is /accounts/v0/tzmesh/keys and /accounts/v0/tzmesh/revocations, and the tztrack pair is /accounts/v0/tztrack/keys and /accounts/v0/tztrack/revocations. Each accounts process that starts while the plane names a product's origin and that product's key set holds no key mints a key into it, so the set is served ahead of the first token; the read writes nothing. A set holding any key is read as it stands.

The list answers each revoked token's identity and revocation stamp for the product, an account row for each account the product was turned off for or whose tokens the operator ended, and the account's standing rows (API-L0-12). It answers them in one revoked array ordered by the stamp then the identity, at most 1000 rows. Those are every one standing where the query names nothing, or from since, a UTC instant, inclusive. after, sent with since, names the last identity read at that stamp, and the page starts after it; a full page answers next, the since and after the next read sends, and a short page answers null.

An account row stands until an invitation's redemption returns the account to the invite group, or until the retention an all row keeps has run from its instant, and the plane removes it at the first of the two (ACB-L0-76; PLD-L0-67). A later closing carries its own instant, and the account's token rows keep their own retention. A product's service applies an account row once per instant, as its own contract states. These rows and the key set are each product's contract's first generation, with the exchange's answer for tzmesh, and system:XTX-22 states how a later one is introduced for tzmesh, and system:XIT-24 for tztrack.

Beside those rows a product's list answers an all row for each revocation of that product's signing key set, its identity all: and its instant, for the product's service to apply as every account's closing (MAPI-09). A revoked token row stands on its product's list until its expiry has passed and the retention PLD-L0-67 states has run, so the list carries every revocation for as long as the token's signature could verify. An all row stands until every token signed before its instant has expired and the same retention has run. A product's service reads its list once a minute, as its own contract states. Where the plane names no origin for a product both of its reads answer their zero forms, an empty key list and an empty revocation list, whatever rows the store holds.

Every product's list answers the two standing rows API-L0-12 states, each held under its reserved identity in the account's place as the all row is, one per account, answered on every list. Suspension writes {"kind": "suspended", "id": "suspended:<account>:<instant>", "revoked_at": <the instant>}, and reinstatement writes {"kind": "reinstated", "id": "reinstated:<account>:<instant>", "revoked_at": <the instant>}, so the later row replaces the earlier of either kind. A service reads the account from the identity and applies the rows in list order: while the suspended row stands it refuses the account's credentials its own contract names (tzmesh:TZT-L0-04; issue_service:ISV-06), and the reinstated row lifts that, the credentials kept. The suspended row stands while the suspension does; the daily pass removes a reinstated row once the retention PLD-L0-67 states has run from its instant. Each kind is the additive change system:XTX-22 admits for tzmesh, its cases the tzMesh package's, and system:XIT-24 for tztrack.

The client's name is its registered name, or MCP client where it registered none or one not printable. No tzmesh token carries a handle, since the tzmesh service holds the account's handle. The erasure walk's ten-minute token carries act with the value erase, which the tzmesh service admits at its erasure route alone (ACB-L0-47). No other endpoint or wire call admits a token carrying an act. The MCP front's two approval marks, describe and pending_action, are no act: a product's door admits each on the one forwarded call it marks, as MAPI-03 states (MCP-14; MCP-07).

For the tzmesh site's browser, the plane answers the origins the setting TZMESH_SITE_ORIGINS lists (PLD-L0-93) with CORS headers on the authorization-server metadata, the registration, and the token endpoint. The preflight allows POST and Content-Type with no credentials, a listed origin's /mcp is admitted as the tzmesh resource, and with the setting empty the tzmesh origin alone is listed. That fallback is for these OAuth reads. The setting also supplies the Settings frame page's framing origins, which PLD-L0-69 derives without a fallback; an absent or empty setting permits no framing origin (WEB-L0-21).

The daily client sweep (PLD-L0-67) deletes a registration past its age that nothing references, a client that never issued a refresh token among them. It keeps the tzmesh site's own sign-in client: a registration the setting TZMESH_SIGNIN_CLIENT_IDS lists (PLD-L0-93) whose every redirect is an HTTPS origin's /signin/callback, exactly. The operator lists the id the estate's own registration answered, so the tzmesh service's sign-in client is registered once per estate (the tzmesh member's configuration statement). The sweep keeps no registration by its redirects alone, since anyone may register one, and with the setting empty it keeps none by this rule. An unrevoked tzmesh token issued to a client, expired or not, references it, so a builder's application stays registered while a person's token stands.

At the sign-in, the tzmesh site's own sign-in client is a registration whose every redirect is the /signin/callback of the tzmesh origin or a TZMESH_SITE_ORIGINS origin, exactly (the tzmesh member's sign-in statement), and one TZMESH_SIGNIN_CLIENT_IDS lists where it lists ids. Any other registration is read as every other client is. Its request holds no page: at the sign-in's completion, for an account in the group, the plane mints the single-use code at once and redirects it with the state to the registered callback. An account outside the group is returned access_denied and the state at that moment, with no description and no page, as a tool's completion returns it.

Where TZMESH_SIGNIN_CLIENT_IDS is empty, anyone may register such a client, and the code still reaches only that callback, where the tzmesh service completes only a sign-in it started, under its own client and the code's PKCE verifier.

This paragraph qualifies the paragraph above's sign-in. A browser whose site session signed in within the freshness window the allow-access screen reads, eight hours, completes the tzmesh site's sign-in client's request from that session with no page, the group's rule applying as at the sign-in's completion.

This paragraph qualifies the exchange paragraph above for the sign-in client's code. It exchanges for one tzmesh token at the level chat under the label Sign-in at <host>, the callback's host, issued to the client. Its life is one minute from its mint, in expires_at, on its row, and in the payload's exp. That covers the service's exchange and its one register_human call. Where a same-origin post of the site started the sign-in with an intent, it also covers the one register, attach, remove_listeners, delete_agent, or public create_channel the callback makes for that intent in the service's own process. The service then drops the value and keeps it nowhere, so no sign-in row stands 180 days. The replacement rule keeps one unrevoked row of the label standing for the client per account, and the exchange's other rules stand as written.

This paragraph qualifies the request and completion paragraphs above for a builder's application: a request naming the scope tzmesh.application, the tzmesh site's own sign-in client excepted. Its request meets no consent screen. An account's first sign-in to that client holds one line, served by the accounts service under /oauth/tzmesh/. The line names the client by its registered name and its callback's host and says the client will see and send the person's messages as them, with Continue and Cancel. The line offers no other choice. The line and its posts answer only the browser whose sign-in held them, no other site may frame them, and Cancel returns access_denied and the state to the client's redirect. Where the account holds an unrevoked tzmesh token the client was issued, expired or not, a later sign-in holds nothing and mints the code at once, and a prefetch holds nothing.

A request without that scope, one naming tzmesh or none, meets the one consent screen wherever its redirect is not a loopback address, so a hosted MCP host is named to the person and allowed there. Anyone may register a client and name the scope, so the line, naming the host the code goes to, is what keeps a page offering a Turn Zero sign-in from taking a person's messages with no word of what the sign-in gives.

This paragraph qualifies the exchange paragraph above for a builder's application's code. It exchanges for one tzmesh token at the level chat, issued to the client, under the label <name> at <host>. The name is the client's registered name, or An application where it registered none or one not printable, cut so the label keeps its callback's host within the label's bound. The token's life and the replacement rule stand as the exchange paragraph states them. An account outside the group is returned access_denied at the line's moment and refused invalid_grant at the exchange, as at a tool's completion. The token's payload carries application_sign_in true, which no other code's token carries (MAPI-03), and the tzmesh service reads it as system:XTX-02 states.

This paragraph qualifies the opening paragraph's token, the resource rule, and the protected-resource metadata for the tzTrack door's protected resource (tztrack:TRK-L0-02; ACB-L0-89). Where the plane's setting TZTRACK_ORIGIN names the tzTrack origin (PLD-L0-93), the tzTrack resource is that origin's /mcp. A registered client asks for it by naming it in resource at /oauth/authorize. The plane binds no builder realm session to that resource and renews none: the request exchanges for the signed tztrack token, as the next paragraph states. A refresh presenting a row bound to the resource is refused, invalid_target where it names another resource and invalid_grant otherwise.

The request names the scope tztrack or none, a request also naming tzcloud or offline_access granted tztrack alone, and its completion is the session branch's own, with no consent page of its own. Where its redirect is not a loopback address, the allow-access screen's tool access form holds the code until its Allow access, as for a session (CHI-L0-03). The code then exchanges for one signed tztrack token at the level admin under the client's registered name, on the tzmesh exchange's form (MAPI-14). The product gate is read at the exchange through the one rule (ACB-L0-76): an account outside it is refused invalid_grant with no product word, and the refresh arm refuses the resource.

The door at the tzTrack resource and its protected-resource metadata are the issue service's (tztrack:TRK-L0-02; ACB-L0-89), and the plane serves neither on that host. The plane's own surfaces admit the tztrack token within its grant alone (MAPI-16) and bind no session to the resource. The metadata the plane's origin serves stands unchanged. Where TZTRACK_ORIGIN is absent the plane names no tzTrack origin, and a request naming the resource is refused invalid_target as any foreign resource is.

The tzmesh list also answers {"kind": "site_session", "id": "site_session:<reference>", "revoked_at": <the instant>} for an ordinary browser sign-out that presents a browser session cookie that verifies (ACS-L0-18). The instant is that of the latest such sign-out; a repeat updates the same row. The reference is the token's site_session claim, never the session's secret or row identifier. At its next poll the service ends every token carrying it and every stream opened under one (system:XTX-22). These rows are purged with the closings in both stores under the same retention (PLD-L0-67).

The three forms #

MCP-03 (decided): Every management action callable by an AI client is a tool corresponding to one API capability. Its authority tier — observe, reversible or destructive — remains explicit in the authoritative catalog and enforced by the management API (API-L0-06); a host annotation neither grants authority nor supplies approval. Each tool advertises truthful standard MCP annotations derived from its action’s catalog metadata, under MCP-11. Served readable context retains its addressable resources, and skills retain their prompts. The same content is also available through the observe tools CTX-07 defines, so resource and prompt support enriches a client without being required for discovery. These protocol forms derive from the same API and content sources; none owns an independent capability or copy of the content.

MCP-04 (decided): Tools alone are the floor: every launch scenario’s management steps complete through tool calls and the browser moments CHI-L0-03 states, with no management step that requires resources or prompts. The signed-in site (WEB-L0-15) is a view of the same acts, required by none. Deployment starts from the externally prepared and uploaded artifact CHI-L0-01 requires, or from the files deploy_files carries (API-L0-25); deploy receives its account-owned storage reference and hash, without requiring a filesystem or byte upload in the MCP client. A client supporting only tools loses the slash-command surface and the preloadable context, never a capability — discovery’s catalog and complete readable content arrive through read_context and read_documentation there, and the served documents remain readable outside the protocol (API-L0-09's other two forms). This is what makes "any AI tool's output is a first-class citizen" true of the tool's *client features* too.

MCP-05 (decided): Discovery is the protocol's own listings — tools, resources, and prompts, each entry carrying a description a newcomer can act on — plus the overview. The overview, available as both a resource and a read_context result, answers "what can the platform do and where do I start" in one read. Every listing description derives from the same one source as the human documentation and the authoring bundle (API-L0-09), so what the tool shows, the docs say, and the bundle teaches cannot disagree. An entry of the words override file (PLD-L0-114) is one exception, while it stands. The rule extends to library catalog rows: a catalog row's listing description derives from the package-owned selection summary the catalog quotes (FTR-L0-57's section), the same one-source discipline over the library's source.

The rule's other exception is MCP-08's listed-name step. On the MCP route the texts MCP-08 names carry the listed names its step writes. The wire, the authoring bundle's guide copies and its source: entries keep catalog names. A served text and its one source then differ in those names alone.

A served string names a product by its formal name or its short name, tzCloud, tzTeam, tzTrack, tzMesh, or tzDocs, at its writer's judgment, or names the company, Turn Zero, or the platform (azure:AZD-08), and never a code name or an internal product name. This is the product-name half of the public-name rule the documentation suite and the served-surface suite share. A served string is every tool title, every tool description and argument description, every resource description, every prompt description and prompt text, and every text the server composes — the overview, a refusal's detail, an action's response detail. It is also every field of the enumeration files the bundle and the service_contracts resource carry, the bundle's guides, and the selection summaries the catalog quotes.

A forwarded row's strings are served strings too: its title, its description, and its argument descriptions, as a product's door serves them under the product's prefix. The plane's served-surface suite holds them to this rule through the door's recorded list (MCP-14).

The plane's own tzTrack rows follow the door's rule for plain words too (tztrack:TRK-L0-03). Those are the actions whose management row names the tztrack surface, list_issue_spaces, and submit_feedback. The texts held are their titles, summaries, and references in the tool catalog, their argument descriptions and retired-argument refusals, the tzTrack practices header, and the service map's tzTrack line. None of them holds a working word of the issue service or an order form aimed at the AI. Both lists stand in one shared module, which the door's suite and the served-surface suite read.

Two strings keep the long form, each fixed by its own statement: the X-Powered-By header (PLD-L0-67) and the apex page (SVC-L0-07). The skills file's strings take this rule at the served-skills landing that rewrites them, a requirement pending there until it lands (RLC-L0-92; MCP-06).

The requirements documents the bundle packages verbatim are the source itself (CTX-02), and CTX-08 holds that source to the same product-name half. A clearly labeled reproduction of an indexed contract, enumeration, or guide in the generated human reference therefore carries the approved names because its source does. A reproduction preserves its source's identifiers. The reference distinguishes resolvable links within the published index from citations to sources that are not published there, without publishing additional sources merely to satisfy a citation.

No listing description carries a requirement or decision identifier: a tool's, an argument's, a resource's, a prompt's, or a catalog entry's summary read_context with no id answers. No line of the overview carries one either. A catalog row names the statements its descriptions follow in its scenario and owners members (MCP-08), which reach a client in the whole-file reads alone. The served-surface suite holds the listing descriptions and the overview to this paragraph.

No other text served to a customer's tool carries a requirement or decision identifier either: the instructions, an action's answer and refusal, a composed read_context read, and a skill's page, part, and prompt description. A read that reproduces a source keeps its identifiers: a whole-file specification read, a read_documentation answer, and a search_documentation answer. So do CTX-09's line for a query that is one rule code, a forwarded product's answer, held to this rule at its door (MCP-14), and the top-level prose of an operator's completed answer. A rule code is a statement's identifier, met in those reads and in the documentation. The served-surface suite holds the served strings to the product-name half and to this paragraph.

The rule that text served to a customer's tool carries no requirement or decision identifier also reaches the text the platform composes on a route that a customer's application, its users' browsers, or its scripts call. That text is a refusal's detail, an answer's detail, and a page's own words, in any body or page the route answers. Storage, egress and its upstreams, accounts and sign-in, push, email, logging, deploy progress, and the serving router are among those routes. Customer data passes as it is: a stored object, a declared upstream's or the application's answer, a log line, and a name the caller sent that a detail echoes.

On those routes a reproduction keeps its identifiers, as a reproducing read does. A Turn Zero product's answer through its upstream is held to the same rule at its door, as a forwarded product's answer is, since the gateway passes it unchanged (EGW-L0-03). The rule also reaches the text a library package's modules compose, which run in the customer's application: an error's message and a refusal's detail reach its handler, its logs, and its scripts as the platform's own words do. The tzMesh and tzTrack services' own routes and the front-end packages' modules take the rule at their own landings, a requirement pending there until each lands (RLC-L0-92).

Beside the overview, the server's instructions name, on every connection, the estate's configured origin (MCP-02), never one read from a request. They name what a script reaches under a credential: each management action at its reference page's address, the library's reads among them, and the storage routes. Before the line naming the origin, which stays their last, the instructions say that a management action's refusal may carry help, the address of its row on the refusals page at <origin>/cloud/reference/refusals/, which explains the refusal. That origin is the public origin help is composed on, so an agent meeting a refusal name no help rides with, in a detail, a data plane's body, or a page, still learns where the name is explained.

The overview, as a resource and as a read_context result, carries the same origin line.

Each tool's summary opens with one sentence of at most 200 characters saying what the tool is for, so a host that cuts a description short still shows what the tool is for.

This paragraph qualifies the origin paragraph's instructions. On a connection listing read_status, they also say, ahead of the refusals page's sentence, that a transport error or a gateway's error gives no outcome (MAPI-04). A read is repeated, a send with its same client_id, and another write after reading the state. For the acts MAPI-04 holds they name read_status with wait_seconds as the read to make first. Such an act is repeated only where that read shows it did not start, never where it cannot tell, and the overview says what in that read shows a start. A schedule run is repeated only with its same request_id, whatever read_schedules shows. The text names no identifier. The overview carries it under the same condition, with what shows the act started, and deploy's description says what shows a deploy started and that an upload pending_upload still names was not.

Where tzmesh_send lists and read_status does not, the instructions instead say that a transport or gateway error gives no outcome, reads are repeated, and tzmesh_send is repeated with the same client_id. Other writes follow a state read. This form names no tzCloud tool and precedes the refusals page's sentence. The standing form and its overview condition stay unchanged wherever read_status lists.

This paragraph qualifies the origin paragraph's instructions too. On a connection that lists read_context, they open with the service map's tzCloud line where the address lists tzCloud's tools (MCP-01), and otherwise with the sentence naming start. That line says in plain words what the platform does, with no claim about it, and the prefix its tools carry. The opening then names start as the first call, where the connection lists it. It names the overview, with the call that reads it, as the read for a platform rule the task needs, and not as a read before every first call. A connection that lists read_context and not start is sent to the overview alone. The opening names no identifier. On a connection that lists start, the overview says that one connection reaches the products the account holds.

This paragraph qualifies the opening. On a connection that lists start, the opening names the tool index, the tools entry CTX-07 states, after the sentence naming start and ahead of any longer guidance. It names the index as the read for finding a tool, not as a read before every call, so start stays the first call. It gives the call that reads the index whole on every customer's connection: read_context with id tools and limit set to the readers' text_max_limit. The opening on a connection that lists no start is unchanged. The sentence names no identifier.

The service map's product lines follow the tzCloud line where it stands, one for each product whose rows MCP-14 lists to the connection, and none on any other connection. Each says only what the product is for, its prefix, and where its overview is read, the product's practices entry overview:<product> (CTX-07), whose header says the rest. The lines name no identifier.

At every MCP address the service-map lines name only products that address lists and the account holds. No product-off line is served. The feedback rule stands only where submit_feedback lists, so it is absent on the operators' address. Every address keeps the estate-origin and refusals-page lines; the operators' address serves those two alone. The mobile and Node.js lines keep their listing conditions.

On a connection listing any of deploy, deploy_command, deploy_files, promote, read_logs, read_schedules, read_status, read_usage, restart_application, roll_back, run_schedule, or store_secret, the instructions say once why wait_seconds is bounded. They tell a caller whose client allows less than sixty seconds to pass fewer seconds on this call and on a returned next call. The line follows the lost-answer sentences where those stand and precedes the Node.js sentence where that stands. It carries no requirement or decision identifier, and the overview does not carry it.

This paragraph qualifies the origin paragraph's instructions too. On a connection that lists deploy, they say, after the sentences on a lost answer where those stand and ahead of the refusals page's sentence, that a command this platform answers needs Node.js 24 or later, the version PLD-L0-82 holds every stated number to. They also say that the Ways to deploy to Turn Zero Cloud page, read through read_documentation, compares the ways to deploy API-L0-25's route selection chooses among. The text names no identifier, and the overview does not carry it.

MCP-13: Each text the server lists is held to what a client shows of it: a tool's title, a tool's summary, a tool's served input schema, and the server's instructions; a row's reference holds what its summary cannot. A caller then reads the whole of what the listing says. A title, a summary, and the instructions, each counted in UTF-16 units, and a served input schema, counted in bytes of compact JSON, run to at most the figure the catalog's listing_bounds member holds for each, a row that member holds by name excepted. The instructions are measured at their longest composition on each estate's origin. That member of schemas/mcp_surface.json is the one home of the six figures, and family:Q-302 holds what each client shows of a listing, the versions it was read at, and the margins.

The whole listing served to a builder's connection holding Turn Zero Cloud and tzTrack alone, as every account does (ACB-L0-76), counted as the tools/list result in bytes of compact JSON, runs to at most the sixth figure, whole_listing_bytes. It is measured at its longest composition on each estate's origin, with the members each estate's committed settings open. That figure is a bound with no client limit behind it. It rises only where a landing's listing would exceed it, to that listing plus 2,000 bytes of headroom. It may also be lowered, but ordinary growth within the bound moves nothing. It moves only by a recorded write naming its reason, and never by unlisting another tool to admit a later one.

Every listed tool carries a title, the short name a person reads where a host shows the tool, such as a connector directory or an approval prompt. The catalog row's title member holds it. The server serves it with the tool, and again as the title member of the tool's annotations, beside the effect annotations MCP-11 states, because Anthropic's connector directory reads it there. A title is in sentence case, names what the tool does in plain words rather than the platform's internal terms, and is unique across the catalog. It carries no identifier.

A forwarded row's title is the form MCP-14 gives it, the product's short name and the door's own title, tzTrack: Find issues, unique beside the catalog's titles. The served-surface suite's sentence below holds the composed listing, the forwarded rows included, through each door's recorded list (MCP-14).

A forwarded row's neighbour among the plane's own rows is the one a caller might reach for in its place. A door's description cannot name the plane's tools, so the neighbour is named in the header of the product's practices entry (CTX-07), the read the product's service-map line names (MCP-05). For tzTrack, the header says that tztrack_raise files in the person's own space and that a problem with Turn Zero itself goes to submit_feedback.

A summary holds what a caller cannot learn elsewhere in time. It says what the tool is for, in its opening sentence (MCP-05). It says when to call the tool and not a neighbour, and the acts before and after it where the call is one step of several. It says what the call changes, provisions, costs, or cannot undo, and the approval it waits for. It states a rule on how to call the tool that no single argument owns, and a value in the answer whose meaning the answer does not carry itself. Where the row carries a reference, the rest is on the action's reference page, whose path the next paragraph says where to find. A sentence another statement places in a description stays in it.

The rest of a row's text stands in the row's reference member, which the listing does not serve. The action's generated reference page renders it after the summary. No summary names that page. read_documentation's own summary says once where an action's reference page stands, so the route stays in the listing every host shows, and the tools entry's first line says it again (CTX-07); the listing repeats no such sentence. A row's owners names the statements its reference follows too, and a reference carries no identifier, as a summary carries none.

An argument's description, counted in bytes, runs to at most the figure listing_bounds holds for it, a description that member holds by name excepted. It keeps what the argument means and the values it admits where the schema does not state them. A schema bound is stated as the bound, with no refusal code; invalid_request is named only for a rule the handler checks, since over MCP the client's argument check answers first.

What else it said stands in the row's reference, under the argument's name. A description that cannot hold those, and any sentence another statement places in it, within its bound is named in listing_bounds with its reason and the size it is held at. A row whose served input schema cannot meet its bound with those in the listing is named in listing_bounds with its reason and the size it is held at, and is cut no further. The suite refuses a named row that fits the bound, a named row over its held size, and an unnamed row over the bound.

The served-surface suite holds each bound against the listing a connection receives, the tools entry's first line against the reference pages' path rule, and every summary to naming no reference page of its own.

MCP-06 (decided): Each skill surfaces as one prompt, invocable as a slash command in clients that support them, and the prompt set is derived from what Turn Zero Cloud offers (API-L0-10). The server enumerates it from the same source as everything else, and no hand-kept prompt list exists to drift. A product door's relayed prompts (MCP-14) are the door's own, listed beside this set under the product's prefix; the set itself stays the plane's. A tools-only client obtains that same skill text through read_context, with the same implementation and credential-grant filtering (CTX-07); the tool path does not author a second recipe. A skill lists only where every action it drives lists (MCP-08).

Each skill row carries a must member, an array of imperative sentences stating what the implementer must do, each ending with the identifier of the statement that owns the rule. The served skill text carries every item on exactly one of its pages under its list number; the human reference page places the whole list ahead of the whole summary. The prompt's description carries the summary's first sentence and then the first page less its title, actions, and footer; the prompt's text is the first page. The catalog entry for a skill carries no must item: it names the skill, its read_context identifier skill:<name>, and a one-line purpose, the summary's first sentence with each parenthetical holding an identifier dropped whole. No rendering restates a rule the list omits or reorders the list; each driven action is one line, its name, its tier, and its tool summary's opening sentence.

A skill row may carry a short_path member: a lead, a few steps for the common case, each naming by number the items it follows, and the items applying only where a condition holds. A scope naming a part, a slug, or a row marking its reasoning as a part, pages the skill: a first page under skill:<name> and parts under skill:<name>:<part>, the summary the part reasoning. The first page holds the short path, a routing line per part with its identifier, condition, and items, and every item no part holds; a part holds its condition and its items. A one-page skill, and a paged skill's first page, fit the readers' default text limit, and each part fits the maximum. A path naming a number outside the list, scoping an item twice, or breaking the measure fails generation.

A skill row may carry a page member: the route of the documentation page that answers its task, as the Cloud tree's index prints it. An overview line gives the skill's name, any credential suffix and the one-line purpose; with a page, ; page <route>. replaces the purpose's closing full stop. A test holds each page to a route a page of the corpus declares. A line with no page ends at the purpose's own full stop. The overview's Where to start header says once to read a skill with read_context under skill:<name>. The list closes with one line naming the documentation page that maps tasks to guides.

This paragraph qualifies the third paragraph's first page. An item a scope holds there carries its condition ahead of its text, unless the item opens with "Where" and the condition's words. The reasoning part's routing line and catalog summary name it as the reasons behind the rules, and a scope naming its part reasoning fails generation. The human reference page, and a one-page skill's text where its row carries a short path, place the short path ahead of the list, a scoped item carrying its condition there as on the first page.

A part carries the line of each action that its items alone name, an item naming an action where the action's identifier stands in its sentence as a whole word; the first page carries the lines of the rest.

Every rendering of the list, the served pages, the prompt's description, and the generated skills reference, prints each item's sentence without its owner parenthetical. The served pages, parts, and prompt description close with no Rule owners list, so they print no identifier (MCP-05). The generated reference's section for the skill closes with the whole row's list: one head line, "Rule owners:", then one list item per item, "- 1: CHI-L0-07". So an item's sentence and its identifiers are two units under the generated reference's sentence bound. Where the first page's list skips numbers a part holds, it prints one line per run of skipped numbers at the run's place, "Rules 19 to 27: skill:deploy:stored-key" and a run of one "Rule 37: skill:deploy:web-client", so the numbering reads whole. The row's must member keeps every identifier at its sentence's end, and the three renderings share one implementation.

The fixed footer that closes every skill's first page, with its sentence on completing a destructive action through the browser approval, is the platform's own text and not a rendering of the row, so no row's list owes an item for it.

Each must item is one sentence of at most 40 words before its closing owner parenthetical and full stop; only that parenthetical holds identifiers, joined by semicolon and space, and code marks enclose only catalogue action names, and it cites no rule by number. It opens with none of "A", "An", "The", "Each", "Every" or "No", cites no other skill's step by number, and, except read_documentation, submit_feedback and start, names only actions its row drives or the suite's per-row list admits as mentions that direct no call. A short_path member's lead, scope conditions and steps, except each step's closing rule list, name no rule by number and no other skill's step by number.

MCP-07 (decided): A destructive-tier tool never destroys: it creates the pending-action record and returns Turn Zero Cloud's description of what will be destroyed with the approval link (API-L0-07), and the outcome is readable through the surface once the browser click completes. No tool, resource, prompt, or other capability of this surface approves anything — at launch or ever — and the suite refuses a catalog that grows one.

A destructive door row the front lists (MCP-14) takes the same route. A connection whose credential holds no destructive class is answered MAPI-12's class refusal by name, and nothing is described. The front reads what the call would destroy from the product's own door, under a product token marked describe (MAPI-03), creates the pending action with that description, and answers the approval link. That description is the product's own first-party reading, never a model's paraphrase; for a door row it stands where the first paragraph names Turn Zero Cloud's description. After the browser's approval the executor describes again, checks the fingerprint, and forwards the call once under a token marked with the pending action's identity. The door's answer, or its refusal, is the outcome, read by read_approval_status (API-L0-07). A door's refusal of the describe is relayed as served, and nothing is held.

The launch catalog #

MCP-08 (decided): The launch surface is enumerated in schemas/mcp_surface.json, revision 11: each capability's name, a tool's title (MCP-13), form, tier, the launch scenario it serves, and owners, the statements its served description and its arguments' descriptions follow. Host effect annotations derive from the corresponding management action row under MCP-11, and the title in a tool's annotations from its own row under MCP-13. A shipped name is a stable identifier: renaming or removing one is a versioned contract change, never a quiet edit. A capability added later arrives as a catalog row with its form and tier assigned under MCP-03, which is how the destructive boundary stays enumerated per capability rather than judged per incident.

Every listing rule below for a plane catalog row has one last condition: the row's addresses member must name the reached address, tzcloud, tztools or tzop (MCP-01). At the company's origin it must name tzcloud or tztools, whether the estate serves those addresses or not. A missing or empty member admits no address. Anonymous rows, bounded credentials, device sessions, staff sessions, skills, payload entries and the tool index take the same condition. On the wire, the tool index follows tools/list at the company's origin (CTX-07); skill and payload entries read no MCP address. The listing changes no wire gate. A call naming a row listed at another address meets the unlisted-row answer.

The public catalog view shows addresses on each unmarked row it lists, including a value tzop where the row names it. The member is data about the surface, not an access boundary. A row marked only by a grant stays out of that view under CTX-08.

A row's surface member in schemas/management_api.json is tzcloud, tztrack, or account. It is the row's product gate where one is set, the profile word cloud read as tzcloud, and relay_issue_act is account, the one gated row that departs from its gate. Of the rows with no gate, delete_space, redact_issue_text and set_ai_passes are tztrack, read_library is tzcloud, and every other is account. A row's MCP name derives from it: tzcloud_<name> and tztrack_<name>, or the bare name on account. The catalog keeps the bare name, which is the REST name. start keeps its name, and list_issue_spaces stays bare as P-20's stated exception, the read of spaces no door lists.

A tools/call naming the bare name of a prefixed row this connection lists, answers unknown_action naming its prefixed tool, by the derivation and no table row. On the MCP route, a served description's backticked catalog names use each row's listed name. So do next.action, without_command.action, fallback, and catalog names in the answer's top-level prose. The wire keeps catalog names. In that prose, plain-word names such as deploy change only when backticked. A double-quoted value keeps its text, since the plane quotes a caller's own name there.

The step also applies to the served skill and guide texts and catalog summaries returned by read_context, to MCP resource descriptions, and to the overview's skill lines and guide line. The guide entry's catalog summary and the overview's guide line are lists of names, with or without backticks. A name of a catalog row becomes its listed name; any other name stays as written. The authoring bundle's guide copies (bundle, guides[].text) and source: entries keep catalog names on both routes. On the MCP route, the step applies to backticked whole catalog names in documentation tree texts, page copies and outlines before any section, query, offset, total or stamp is read. Documentation resources answer that same listed text. search_documentation excerpts keep catalog names.

On the MCP route, fenced blocks in documentation tree texts and page copies stay as written, because they show a file's or a command's own text that a reader may copy. Each outline is listed whole.

The catalog's first row is start, a connection's first call, at the observe tier. It answers in one read the acting account, the connection's credential, the products it holds, and tzCloud's applications with their states, as list_applications answers them, where the account holds tzCloud. Its top-level tools_plan answers the person's plan through the reading read_account uses (ACB-L0-92). It answers neither on nor off on any route or address.

The credential member names the credential's kind and gives a minted token's grants as list_tokens answers them. For a session it gives only super_admin when held. The member omits the credential's value and label. Of minted tokens, only account tokens reach start, since the scope gate refuses an application token (API-L0-17). Its first sentence says whether the credential admits test end users on production (ACS-L0-21). A credential that holds the super-admin grant gets a second sentence saying that per-account application limits do not apply (ACB-L0-84). The row carries device_sign_in under admits, as read_account does, and no other mark, so it is listed wherever read_account is listed.

start reads and changes nothing. It and read_account name each product by its product word, tzcloud for tzCloud (ACB-L0-76). On the MCP route read_account answers the same products start reads. Its product_states member holds one entry for each of tzcloud, tztrack, tzmesh, and tzteam, each with product and held alone. Its detail says what tzteam is: tzTeam's access, always on and managed by tzTeam, with no tools on this server.

Where the connection lists choose_tools_plan, while tztrack_billing reads 0, the account has no subscription row, and the person's plan reads free, start's detail invites the person to choose a tools plan with choose_tools_plan. It says Free is the default and Tools and Tools Plus cost nothing during the introductory period. The invitation changes no record. A stored Free choice suppresses it, as does an unbilled account's Tools Plus reading (ACB-L0-92).

When available, start returns a versions object, and its detail states where those values come from. versions.server and versions.tzcloud give the plane's full build commit, which also appears in feedback's seen_in.version. For tzTrack and tzMesh, versions includes the latest nonempty serverInfo.version held from that product's MCP handshake, when the connection lists that product's tools. The object omits unavailable values. The response omits the object when no values are available. On a stamped build, credentialed MCP initialize returns the full build commit as serverInfo.version. On an unstamped build, it returns 0.1.0.

On a connection MCP-14 serves that lists tzTrack's tools, it also answers tzTrack's line: the default space and the count of issues waiting for the person on a first page of at most 100. On one that lists tzMesh's tools, it answers tzMesh's line before it: the account's tzMesh handle and its standing agents, read from the door's list_participants under the front's token. The unread count needs a participant, so the line leaves it to tzmesh_read. Where a door refuses the read, its line relays the door's own reason and the door's next, named as this connection lists the call (MCP-14). Its state is refused, or setup_needed where tzTrack answers space_needed or default_space_needed, since nothing failed and the account has a step left. Where a door fails, its line is unavailable with the door's address. It answers no product line on any other connection.

On tzcloud for a tzTrack or tzMesh holder, and on tztools for a tzCloud holder, start's detail gives the estate's other address and its install name. It tells the person to add that connection and sign in there if their AI tool does not have it yet.

An agent calls start first and may never open a product's practices entry, so each product line read whole also carries the steps its practices most need at session start. The tzTrack line says to read the open issues with tztrack_find at session start and to search with it before tztrack_raise. When the account has no tzMesh handle, the tzMesh line says to ask the person to choose a handle, whether to stay unlisted, and whether their agents, sessions, and applications acting as them may join public channels (tzmesh:TZT-L0-19). It says to set the person's choices with tzmesh_set_account before calling tzmesh_register. For an account with a handle, the line gives the registration step directly. Both forms say to pass the id tzmesh_register returns as as on every later call that takes it.

The served listing carries the catalog's rows whose implementations have landed. A row ahead of its implementation stays enumerated, its name stable and its tier assigned, and unlisted until the seam answers it, the rule the skills already follow (API-L0-10), so a connected tool never lists a capability that would refuse not_yet_provisioned.

The listing also carries the forwarded rows MCP-14 states: the tools of a product's own MCP server, under the product's prefix, at the company's origin and tzTools alone. They list to a connection MCP-14 serves whose account holds the product, while the product's front switch reads 1. Each is a shipped name of the door's own contract and no row of this catalog. A call naming a forwarded tool is answered as MCP-14 states.

At either forwarding address, a holder's call, save one naming a door's former name (MCP-14), is refused by name with the door's address where the product is unserved. At Hosting and the operators' address, or for a non-holder, the call meets the unlisted-row answer below.

The listing is also the connection's own. At the operators' address, a row MAPI-09 marks with a grant is listed to a connection whose credential holds the row's grant or super_admin, the grant that admits every grant-marked row, or is bounded by a grant the row's admits mark names (MAPI-09). It is listed to no other connection: an anonymous connection, a customer's session, and a token minted with none of the grants among them (MAPI-14).

The company's origin does not list the operator's skill, skill:platform-admin; its context catalog also lists no payload:delete_account. The operator's skill lists as a prompt at the operators' address and remains readable on the wire under its grant (MCP-06; CTX-07).

The rule reads the enumeration's own mark and no list kept beside it (MAPI-11's discipline). The catalog and the enumeration stay readable whole in the served bundle under a staff grant, and through CTX-08's public view under any other credential (CTX-02; API-L0-15). Neither the listing nor that view is an access boundary: the wire gate is the only one. The overview names the operator's skill as the operator's, and the marked row itself is unchanged. The wire route answers it and refuses a credential the row does not admit, by the row's grant name (grant_required; MAPI-09; MAPI-10).

A tools/call naming a row this connection does not list answers a refusal naming the tool, never not_yet_provisioned and never a silent success, and an anonymous call naming it draws the sign-in challenge (MCP-02). A tools/call sent alone that names a name MAPI-02's clause retired answers the refusal naming the tool, with the successor and its mode, the detail the actions route gives. Inside a JSON-RPC batch, such a call answers as a name the plane does not hold.

An unmarked row a grant merely widens is listed to every connection the address admits, the grant deciding the subject or the form at the call. Those rows are the builder realm's three realm rows, and read_feedback and settle_feedback, whose queue form and platform-space settle are the feedback queue grant's (MAPI-09).

One unmarked row lists more narrowly, since its handler admits no other credential. relay_issue_act carries the listing's own mark, listing: admitted, and lists only to a connection whose credential holds super_admin or a grant its admits mark names, the issues grant (MAPI-14). So a token the issues grant bounds and an operator's credential list it, and no customer's sign-in connection does. The mark is the listing's alone: the wire route, the dispatcher's gates, and the handler's own checks stand as they are, so tzTeam and the command reach the row as before. A tools/call naming it on a connection that does not list it answers as a call naming an unlisted row does.

The unmarked rows list_storage_files and read_storage_file also carry listing: admitted, with no admits (OST-L0-03). They list only to a connection whose credential holds super_admin. No customer's connection lists them, so MCP-13's whole-listing bound counts neither. An agent reads the same files through the storage routes. As for relay_issue_act, the mark changes the listing alone. The wire route and browser session reach the rows as their clients state. A tools/call naming either on a connection that does not list it answers as a call naming an unlisted row does.

A connection whose credential a bounding grant bounds is listed less. For a minted token holding synthetic_seed_purge, synthetic_estate, publication, feedback_queue, or issues but not super_admin (MAPI-16), eligible rows are those its grant marks, those whose admits names its grant, and those marked access: anonymous. A row lists only where the reached address admits it. No other row lists.

At the operators' address, a bounded token lists its grant's eligible rows there. At the company's origin it lists only eligible rows naming tzcloud or tztools, so no tzop-only row lists. There, list_tokens lists only for a bound its admits names, and an anonymous row only where its addresses admits. Staff list the customer tools and relay_issue_act, but no row naming tzop alone. The tool index, skills and payload entries follow the address rule above.

A session the device grant issued for a session resource is bounded the same way by device_sign_in (MCP-02; MAPI-16). It lists the rows whose admits mark names that grant, the anonymous rows its address lists, and the forwarded rows of the products its account holds (MCP-14), and no other row. So it lists no tzcloud row but read_library, an anonymous row, no grant-marked row, and none of the token, connection, and account rows outside the mark. start answers it as it answers a person's sign-in connection, its product lines included. An operator's device session lists the same rows, since the session holds no super_admin.

A served skill and a payload row of the context catalog list under the same rule, the listing reading the one predicate over the enumeration's marks that the dispatcher's bound gate reads (MAPI-16). A row marked access: anonymous is listed to every connection its address admits, a bounded token's included, on MAPI-11's mark.

The listing also reads the enumeration's product mark (ACB-L0-76). A row marked product: cloud, one of Turn Zero Cloud's rows, lists to a connection whose account's profiles, the operator rule applied, hold cloud, and to no other account's. A connection with no account, anonymous or one that does not act, is read as holding cloud alone: Cloud's public catalog on Cloud's own server, a reading that opens no gated documentation tree (ACB-L0-82). A row with no mark lists under the rules above. One predicate reads the mark for the listing and for the dispatcher's product gate, which refuses a cloud row 403 cloud_required, and no list is kept beside the enumeration. A served skill and a payload row of the context catalog list under the same mark.

A row marked product: tztrack, one of the five issue-space rows below, lists to a connection whose account holds tztrack under the same rule, and the same predicate's gate refuses it 403 blueprint_required (ACB-L0-89). Since tzTrack's opening every account holds tztrack, so the listing and the dispatcher admit the mark for every account (ACB-L0-89).

The six unmarked rows of the issue record are submit_feedback, read_feedback, settle_feedback, list_issue_spaces, delete_space and redact_issue_text. The five marked product: tztrack are relay_issue_act, create_space, share_space, list_space_members and rename_space. These and the grant-marked check row (MAPI-18; API-L0-20) are listed on the rules above while the build registers their handlers. It does so where the issue service's origin is configured and not otherwise, so before the service stands the twelve are unlisted and refuse not_yet_provisioned as any row ahead of its implementation does.

While they are listed, a connection that lists submit_feedback under the rules above is served the feedback rule of the AI tool interface's feedback scenario (CHI-L0-14) in the server's instructions and in the overview's sentence. A connection whose bound leaves the row unlisted is served neither. Where such a connection presents a credential, every refused result but a refusal of submit_feedback itself carries a second content block naming submit_feedback and the refusal's reference after the unchanged JSON block. Every completed submit_feedback, read_feedback, settle_feedback, check-row, and relay_issue_act result is rendered as one block that opens with the untrusted marker and holds the answer between two delimiter lines, the reporters' text data and never instructions. The tzTrack server's write tools answer compact and unwrapped where the answer carries no one's text, and wrapped otherwise (MAPI-18).

A row the enumeration marks with a switch, switch: synthetic_estate on the two write rows of the synthetic estate (MAPI-16), is listed while the synthetic estate's posture is not off and unlisted while it is off. That follows the same derivation from the enumeration's own mark and that posture. The posture is the control plane's setting SYNTHETIC_ESTATE (MAPI-16). Its wire route stands and refuses by name, synthetic_estate_disabled, and a tools/call naming it while unlisted answers the refusal naming the tool. A row carrying both marks is listed where both admit it and unlisted where either does not. The catalog is data about the surface, not a second statement of the rules above; the suite holds the two together.

This statement records a use of MAPI-02's shape clause on start's products, which answers tzcloud where it answered cloud, and on read_account's MCP answer, which gains product_states and detail. The descriptions in schemas/action_payloads.json and schemas/mcp_surface.json move with it. Anthony agreed to the change on #7828 and #7829. The platform then held 106 builder accounts, synthetic accounts excluded, read from the control database at the change.

generate_image and renew_image_link are unmarked account rows, listed to every connection the rules above admit. They are the two rows whose completed results may carry a block other than text. Where the call names inline true and the picture is at most the bound ACB-L0-93 states, one block of MCP image content holding the picture follows the JSON block. A renewal whose blob read fails answers no image block and gives the reason, keeping its link and commands. The HTTP route never answers those bytes. A host that shows no image content still holds the JSON block, whose download reaches the picture.

buy_credits is the unmarked account tool for requesting a checkout link (ACB-L0-101; ACB-L0-102; MAPI-09). It lists on its row's address and credential terms, including where the estate's credit sales are off. The handler refuses there and opens no checkout. Where admitted, the account owner opens the link and pays on the provider's page, where the provider is the seller. The tool pays nothing, and its answer grants no credit. The tool lists only after its handler stands.

read_purchases is the unmarked, observe-tier account tool for its own credit-purchase history (API-L0-27; ACB-L0-44). It lists on the same address and credential terms as read_balance, whatever sales reads. It takes no account argument and opens no checkout. It lists only when its handler stands.

MCP-14: The server fronts the products the connection's account holds: beside Turn Zero Cloud's own tools it lists, under each product's prefix, the tools of that product's own MCP server, the door, and forwards each such call to the door. It does so as an MCP client of the door and copies no tool's behaviour. A product's rows list, and its calls forward, on a connection where all of these hold:

  • the connection is a person's sign-in connection, the session credential in its first form (MAPI-03), holding no bounding grant or the device grant's bound device_sign_in alone (MCP-02);
  • the account holds the product's profile, read through the one profile rule (ACB-L0-76);
  • the plane names the product's origin (MCP-02);
  • the product's front switch, <product>_front, reads 1 (PLD-L0-76).

These four conditions apply only at the company's origin and the tzTools address. Hosting and the operators' address list and forward no product-door row. There, a call naming one meets MCP-08's unlisted-row answer, never forward_unavailable. The same address condition governs product practices, tools and skills entries and start's product lines. A session bound to tzTools remains a person's sign-in connection, and the front mints its door token on the standing rule.

A connection under a minted token lists no product row and is minted no token: a program uses the door with its own product token. A device session is a person's sign-in connection, its product rows listed and forwarded and its door tokens minted as for any sign-in (MCP-02). The prefix, the product's short name, the switch's name, and the door's address each derive from the product table's row. The prefix is the product's technical name and an underscore, tztrack_ and tzmesh_, and the door is the product's resource at its origin.

This paragraph qualifies the first condition, and the sentence above on a minted token, for the synthetic estate (ACB-L0-79), so the compat harness can measure the products' rows without a browser sign-in. A connection under a minted token of a synthetic account, reaching the whole account and holding no bounding grant, is served as a person's sign-in connection is, on either estate. Its rows list and its calls forward where the other three conditions hold, and start answers its product lines (MCP-08). The front mints its door token as for a sign-in connection, held per token and product rather than per session, and the token's revocation ends its use. Every other minted token, one holding a bounding grant among them, lists no product row. No served string names the exception, since no customer's account is synthetic.

The front's door token for tzMesh carries the optional synthetic and operator marks on the terms MAPI-03 states, each written only as true. These terms apply to a sign-in connection and to the synthetic account's minted-token connection described above alike. An access-link account's token carries no synthetic mark, so the service's refusal reaches the front's calls as it does any other (system:XTX-05). A tzTrack door token carries neither mark. The service reads the marks as system:XTX-22 states.

Each forwarded row carries:

  • the name <product>_<door name>;
  • the title <short name>: <door title>, such as tzTrack: Find issues;
  • the description, the input schema, and the annotations as the door serves them, the front adding the composed title to the annotations (MCP-11; MCP-13) and the hint the next paragraph but one names.

No row lists until its door serves a title and annotations for it. A row the door annotates destructive is re-listed only where its served row declares both approval marks MAPI-03 states, describe and pending_action, on the row's _meta.approval_marks; a row declaring neither, or one, stays unlisted and its call is not forwarded. A listed destructive row's call creates the plane's pending action and answers the approval link (MCP-07; API-L0-07; MAPI-05). The recorded list the suite reads is the door's served list, marks included, and the suite holds both cases.

The front lists the door rows that change what other people see with destructiveHint true, so a host asks the person before each call, since Anthropic's connector directory asks that hint of a tool that modifies data. They are tzMesh's register, end, set_profile, leave_channel, set_channel, remove_member, and set_account, and tzTrack's update, close, reopen, relate, and configure_space. The door's own row stays unmarked, so such a row is neither left unlisted nor held for the plane's approval. A row that changes only the caller's own state, a read position among them, stays as its door serves it, since a prompt on each read would stop a listening session working.

A door row whose prefixed name the plane's catalog derives is the plane's: a call naming it is answered by the plane before any forward, and the door's row is not listed. For tzTrack these are share_space, list_space_members, create_space, rename_space, and delete_space (issue_service:ISV-16).

A door row whose plane act sits on another surface stands in a list with one home in the plane's code, mapped to its plane row and not re-listed: submit_feedback, on the account's submit_feedback, for tzTrack (issue_service:ISV-16) and tzMesh (tzmesh:TZT-08). The plane's suite asserts each name in that list stands in the door's recorded list. A door refusal's next.tool names the call as the connection lists it.

Each plane row keeps its own description, since it describes the act as this connection runs it, and the door's row keeps the door's words. A forwarded name is a shipped name on MCP-08's terms. A door renames or removes a row only as a versioned change of its own contract, never as a quiet edit, since the name reaches every connection's listing.

The tzTrack rows watch and unwatch, which tell a watcher through tzMesh or by email, list to every account, whatever the plane's setting invite_only_products reads. A relayed resource or prompt reads as its door serves it.

The door's list is cached per door, one list serving every connection. It is refetched under the calling connection's own token when any of these holds:

  • the version the door's handshake names has changed;
  • a call names a tool the list lacks, at most once a minute per door;
  • the cached list is older than ten minutes.

Each door read is bounded at ten seconds, the door's own bound on its calls to the plane. A forwarded call whose arguments name a wait in seconds, as tzMesh's send and read do, is bounded at ten seconds more than that wait, counted at most thirty. After a failed refetch the cached list is served. The door's instructions resource is cached with its list and relayed as the product's practices entry (CTX-07), so the door keeps the one source. A forwarded call, and a relayed resource's read, reads its own product's door alone, so a door that is slow or down holds no other product's call.

A door that keeps a session per Mcp-Session-Id, as tzMesh's does, binds it to the account and not the token. The front answers each initialize with an Mcp-Session-Id of its own and keeps no state for it. A later request carrying it, 1 to 128 visible ASCII characters, is served for that host session; a request carrying none is served for its connection alone.

For a host session, the front holds one door session per connection, host session, and product, shared by every replica. Its id is a front_door_session ticket (PLD-L0-67). A replica holding no client for it resumes the ticket's session. Where no ticket stands, the replica initializes one and claims the ticket, and the loser of a concurrent claim ends its own new session before resuming the winner's. The ticket lives seven hours from its last put. A replica refreshes it at most every ten minutes of the host session's use, and resumes the ticket's session where the ticket names another. A replica holds at most 32 host sessions' door clients per connection and product, dropping the least recently made.

A re-mint presents the same session under the new token, so the session continues. Where the door answers 404 for the session, the front resumes the ticket's session where the ticket names a newer one, and otherwise initializes again once. Either way it makes the request once more, and a second failure is answered as a door's failure is.

A host session's door session ends at the host's DELETE /mcp naming the host session, which the front answers 200 after sending the door an HTTP DELETE naming the door session and discarding the ticket, or by the door's own session bounds. A replica dropping its client leaves it standing. For a request carrying no host session, where the front drops a session it holds, at its token's end or on replacing the door's client, it sends the door an HTTP DELETE naming the session, so the door frees it at once.

Each resource a door lists under its product's own scheme, <product>://, is relayed under its own URI where the product's rows list, since the door alone decides its context. It is read live with the door's list. Its row is listed after the plane's own resources, as the door serves it. A read is routed by that scheme, forwarded to the door, relayed as served, and counted as a forward. A holder's read is refused forward_unavailable, naming the door, while the product is not served. A non-holder's read, and a read of a row the front does not list, is answered as a resource nothing registered. The plane holds no list of resource addresses.

A door's prompts are relayed where its rows list: read live from its prompts/list with its tool list, where its handshake declares prompts, and listed under the product's prefix as the door serves them, after the plane's own (MCP-06). A prompts/get naming one is forwarded to the door and relayed as served, counted as a forward. A holder's get is refused as a holder's read of a relayed resource is. The plane holds no list of prompt names, and no step needs a prompt (MCP-04).

The forward presents the door a product token the plane mints on use for the connection. The token takes the signed form MAPI-03 states, under the product's realm, its label the client's registered name as the exchange reads it (MCP-02). It lives sixty minutes and is re-minted when a used connection's token has under a minute of life left. It is held in the replica's memory per session and product, and written to no row and no line (API-L0-05). It is presented at its door alone and never answered to the caller; on the plane it has no row and resolves to nothing (MAPI-03).

A forwarded call's result and a door's own refusal are relayed as served, but for a refusal's next.tool, the door's name for the call the refusal points to. The front names it as the same connection lists that call: a mapped door row by the plane row it forwards to, a re-listed row by its prefixed name, and any other name as served. So the door never learns the route a call came by. The refusal forward_unavailable, this statement's row of schemas/wire_errors.json, answers five cases, its detail naming the door's address where the estate names one:

  • a door that cannot be reached;
  • a door that does not answer a forwarded call within its bound, the detail saying the call may have been made;
  • a door that refuses the forwarded token;
  • the plane unable to have the connection's token signed;
  • a holder's call naming a product tool, save a former name, while the product is not served through the connection, the door's address being the remedy where the estate names one.

A non-holder's call naming such a tool is answered as MCP-08 answers a call naming an unlisted row, naming only the tool the caller sent. An anonymous call draws the sign-in challenge (MCP-02). The front serves single requests: a JSON-RPC batch is answered as before the front, listing no forwarded row and forwarding nothing. A forwarded call is no plane action and writes no action record (MAPI-06).

A former name is a door's old tool name under its product's prefix, as the product's one list of former names holds it: tzMesh's TZMESH_FORMER_NAMES, its tool rows (system:XTX-14), and tzTrack's RENAMED_TOOLS (system:XIT-01). Where the product's rows list, a call naming one forwards while the front lists the name. Once the door's list no longer holds it, the call is refused renamed, this statement's row of schemas/wire_errors.json, and nothing is forwarded. Its detail names each successor as this connection lists it. Where the door's list cannot be read or its token issued, the call meets forward_unavailable as any forwarded call does. On any other connection, a non-holder's and one under a minted token that lists no product row among them, a call naming a former name is answered as an unlisted row is (MCP-08), naming only the tool sent. On those connections forward_unavailable never answers it.

A listed destructive row's forward has one record, its pending action (MAPI-05), and writes no action record either. The door's history names the actor from the token's sub and label. The plane counts each forward by product and outcome on its own counters (PLD-L0-79), and its request windows at /mcp apply to the front's calls (PLD-L0-67). A describe call (MCP-07) is no forward and moves no counter; the executor's one forward counts by its outcome.

The plane's served-surface suite holds the forwarded rows in the composed listing through each door's recorded list. The listing bounds, title uniqueness, the product-name half, and the no-identifier rule of MCP-05 and MCP-13 run over the plane's rows and the forwarded rows together. A product's front switch reads 1 only where its door's served list and instructions stand as a case file in the door's own tests, which the door's suite asserts and the plane's double and suite read.

MCP-11: Every listed tool emits readOnlyHint, destructiveHint and openWorldHint from its management action’s explicit annotations, with idempotentHint only where its effects justify it. Metadata describes the action’s actual business effects and external interactions, not an inference from its authority tier; ordinary operational records and metering do not make a reader a business write. A destructive request describes the whole workflow even though only the browser may approve its completion (MCP-07). These annotations are descriptions, never authorization inputs. The standard MCP SDK serves tool listings and derives input schemas from the registered validators that check calls; the listed set follows the existing implementation, enabled and grant filters. schemas/mcp_surface.json records preferred and supported initialization revisions and the default for an absent HTTP revision header; protocol tests hold that call configuration to the installed server’s negotiation and refusal behavior.

A forwarded row's effect annotations and input schema are its door's, served as the door serves them. Deriving them from the validator that checks the call is the door's, held by each door's own statement, and the listing adds no schema of its own to a product's tools and no effect annotation but the hint MCP-14 names (MCP-14).

MCP-09: Library entries may surface as resources derived from the served catalog, one enumeration entry recording the derivation — never a second row list beside the catalog. At launch none are derived and none are served. The agent writes an entry's files into the project once, from read_library with entry's answer, and a resource whose bytes change under a stable name would give it a second and conflicting source for files it already holds.

A derived library-resource name carries no resolution promise: API-L0-14 serves the latest publish alone, so a name that resolved yesterday may name different bytes today or nothing at all. That is why such a name is governed by API-L0-14 rather than by MCP-08's listing stability. A skill's prompt name is no such name: MCP-06 surfaces each skill as one prompt from the set API-L0-10 derives, and the name is a shipped name MCP-08 keeps stable. Derived rows and the catalog's own resource row enter the enumeration when the library API-L0-14 states serves them and not before — a shipped name is a promise, and an action has a refusal channel (not_yet_provisioned, MAPI-10) where a listed resource has none (Q-239).

MCP-10 (conforms: ACC-01, ACC-02, ACC-03, ACC-04): The connected session renews without the person, and the management plane's token endpoint is where the Account package's session contract lands. The endpoint issues a refresh token beside the access token and accepts the refresh_token grant, so a tool whose access token lapses renews its own. A session the device grant issued for the plane's own resource renews the same way, its bound device_sign_in kept (MCP-02). The browser sign-in then stays the one moment CHI-L0-03 names rather than a recurring interruption every working day.

The renewal mechanics are the Account package's, cited in this statement's conformance edges and not restated here (Q-80's consumption by citation): rotation on redemption, a replayed credential revoking its chain beyond this plane's window, and renewal ending where the session ends. Sign out everywhere, suspension, and account deletion are this plane's session ends (ACS-L0-18; CHI-L0-12 for deletion; API-L0-12 for suspension). Each of the three ends the builder realm's OAuth leg and browser leg together on the terms ACS-L0-11 states; an ordinary sign-out ends one browser session and no connection (ACS-L0-18). Beside them, a connection's own revocation ends one OAuth leg alone: that connection's refresh family and the access token issued in it. It is revoke_token with connection (API-L0-21) or the client's own revocation at the revocation endpoint (MCP-02). Each leaves the browser leg and every other connection untouched (ACS-L0-11).

The conformance is deliberately partial, and this paragraph is the edges' scope note. ACC-L0-06's custody clause — the refresh credential resting in the platform's secure store, never a synced file — is about the account holder's client device. A management-plane consumer cannot honor it, the refresh token living wherever the customer's own tool keeps it. Conformance runs to the session mechanics and not to custody.

The reuse window is the wider grace ACC-03 lets a consumer state for a class of its clients, at the package's ten-minute ceiling. A redeemed refresh credential presented again within ten minutes of its rotation answers a fresh pair from the same chain; presented later, it revokes the chain and ends every access-token session the chain issued. The width serves a developer's tool that keeps one stored credential for several processes on one machine, whose rightful replays arrive minutes apart. Its cost is bounded: renewal falls at a term's end alone, so the detection the window yields is the ten minutes after each rotation and never a standing exposure, and a copy presented later still ends the chain. The end-user realm's token endpoint keeps the package's sixty-second default (ACS-L0-14). The ACC-03 edge runs to the grace, its width included, and to the chain's revocation beyond it.

The lifetimes of each credential are call configuration below this statement, which states that renewal happens and not how long anything lives. One relation is fixed here. The access token, a realm session row the plane ends at once at sign out everywhere, suspension, and erasure (ACS-L0-11), lives the session's whole term rather than a short slice of it. The refresh token outlives it by a further term so that renewal stays possible. A shorter access life buys nothing a revocation does not. Renewal then falls at a term's end alone, which keeps the parallel sessions of one client from racing on a routine renewal and revoking their chain as a replay beyond this plane's window. Each of those sessions holds its own copy of the one token home ACC-05 recommends.

Every renewal path, including the consumed-token grace, checks the same eligibility before issuing either credential: the registered client and any supplied client identifier, lifetime, recorded resource/issuer, permitted scopes, and the account's current standing, revocation state and original authorizing generation under ACS-L0-11. A supplied resource must equal the recorded service. Where that service is the plane's own resource, a value equal to it but for a query holding products alone reads as it, by MCP-02's reading. Requested access-token scopes may be a subset of the recorded grant; omission retains that grant.

Rotation and grace retain the refresh grant's resource, issuer, client, original scope set, authorizing generation, and the bound device_sign_in where the device grant issued the family (MCP-02; MAPI-16). No renewal drops the bound or widens the session. Missing legacy binding is a named invalid_grant, not a default audience. The grace changes only the treatment of a recent repeat redemption; it bypasses none of these checks. A future redemption timestamp is not within the grace.

A control-plane release is not, of itself, a session end: the ends are the three ACS-L0-11 states, the per-device sign-out, and the per-connection revocation, and a release is none of them. Where each credential rests, and what form it takes, ACS-L0-02 and ACS-L0-11 state; none of it lives in a service process, and a release replaces the services' processes and holds none of it in them.

A release is inside this bound only when four conditions all hold. Its migrate-Job set is empty. Its settings change moves neither the issuer and resource the rows are bound to nor the reference to the secret the cookies are signed under. It does not rotate that secret. It changes no credential's or cookie's form. Such a release leaves an access token, a refresh credential, and the browser leg's cookies minted before it standing after it as they stood before.

The upgrade test reads an access token and its refresh credential across each control-plane release: the two reads an operator session makes from its stored credentials without the person. It reads the site's signed-in page, the read that needs the person's browser session, once: the measurement on record in the release plan. No later release owes that read. A browser session a release outside the bound ends is raised as an issue on the issue list, citing this statement, when the person meets it. A refusal it reads across a release inside the bound is a defect against this statement's release clause, found under RLC-L0-96, never a session end this statement admits.

A release outside the bound is named as such in its record before it runs, and the ends it causes are that release's to state. A refusal read across it is raised as an issue on the issue list, citing this statement, the pinned credential and the release's record its evidence (RLC-L0-93).

Each answer to /mcp and to the token endpoint writes one line to the plane's own record, mcp access, on every estate, so a client's lapse can be read after the fact; a preflight writes none. It names the path, the method, the status, the duration, and the replica, and on the token endpoint the OAuth client id the request names, cut at 64 characters. It carries no token, header, query, address, or other member of the body. The lines are capped at 600 a minute per replica across both paths, and the next admitted line counts what the cap withheld.

The answers #

The answers the management-action surface composes follow one convention, their budgets held in schemas/answer_budgets.json.

MCP-12: Every answer the management-action surface composes, on the wire and over MCP, keeps the prose of each of its forms within that form's budget, and so do the texts an MCP result carries beside its first block. These are served strings as MCP-05 defines them. The forms are a completed answer's detail and summary, and a version row's outcome detail as read_status and list_versions answer it. They are also a row's note, an environment's grain_note, a pending answer's detail and description, a refusal's detail, a text read's continuation, and an MCP result's second content block.

Each form's budget is one row of schemas/answer_budgets.json, the one home of its figure. A budget counts Unicode code points over the composed text and bounds the answer's fixed words. A list of the caller's own names or identifiers counts at one entry, and so does one such value, measured at its member's default where it has one. A tool's name counts at its catalog name, however the connection lists it.

A completed answer whose explanation a page holds says in its detail what happened and what the caller does next, and names that page in page, one typed member. Its value is an absolute address on the estate's public origin (PLD-L0-85), composed by the handler from a route it names: a guide where one explains the act, and otherwise the action's generated reference page. The address's path is a route read_documentation reads as its page (CTX-07).

page rides a completed answer the dispatcher answers to a dispatched call. A pending answer and a refusal carry none, a refusal's pointer being its help (MAPI-10). A destructive action's stored outcome carries none either; its explanation stands on the action's reference page, found by the action's name.

Whether a composed text keeps its requirement identifiers is MCP-05's rule, and this statement adds none. The served-surface suite holds the composed answers to their budgets and each page to a public page or, for a row naming a grant, a route its holder's read_documentation answers (CTX-07).

Source Markdown #

---
document: mcp_surface
prefix: MCP
---

# The MCP Surface — Contract

## Purpose

This file is the contract for the Turn Zero Cloud MCP server's surface: how a customer's AI coding tool connects and authenticates, which kinds of capability surface as tools, resources, and prompts, and what the launch surface is. It settles the launch surface on the launch definition's own rule — the scenarios drive, the API guarantees (CHI-L0-02). It states the surface of a *client*: every capability here is the one API's (API-L0-01), reached through the MCP server acting as the connected identity (API-L0-02), and nothing here is MCP-only.

The launch surface's enumeration has one home: the catalog at `schemas/mcp_surface.json` beside this document — capability by capability, each with its form and its authority tier — held coherent by a test in the platform's suite. This document states the rules the catalog obeys and deliberately repeats none of its rows.

What this document does not cover: the API's own resource contract and versioning (`management_api_contract.md`); the pending action's mechanics (API-L0-07 owns them; this surface only reaches them); and the super-admin's capability inventory (MAPI-09's grant rows). Nor does it cover the launch skill set, whose rows are the contract's in `schemas/skills.json` under API-L0-10's derivation rule.

Requirements are numbered MCP-nn.

## The server and the connection

MCP-01 (decided): The Turn Zero Cloud MCP server is one remote server, hosted and operated by Turn Zero Cloud, reached over the Model Context Protocol's streamable HTTP transport at an address the customer adds to their tool. One connection reaches the products the account holds that the reached address lists (MCP-14). The address alone selects that set; no product choice is read from the connection or its query. It needs no local install, no sidecar, and nothing on the customer's machine beyond their own AI coding tool.

The estate's configured origins serve four MCP addresses: its own origin, tzCloud's Hosting address, tzTools, and the operators' address (MCP-02). Hosting lists tzCloud and the shared account tools; tzTools lists tzTrack, tzMesh and the shared account tools. The company's origin lists every customer tool and no operators' tool, to staff as to any other account. The operators' address lists its own tools under MCP-02's admission. An address is served only where its origin setting is set. Every address declares the same server name and description and answers the same handshake.

The server declares itself `turnzero` at every address; that is its protocol identity. Published install stanzas use `turnzero` for the company's origin, `turnzero-hosting` for tzCloud's Hosting address, and `turnzero-tools` for tzTools. No install stanza is published for the operators' address. A client prefixes a server's tools with the name it is bound under, so the stanza name is the word a customer's assistant reads on every action. The declared name spells the company's name, Turn Zero (azure:AZD-08), since the server fronts the products the account holds (MCP-14). It identifies neither one product, the host, nor the workspace's internal code name. The declared and stanza names use lowercase letters, digits, and hyphens. Some clients' tool-name grammars exclude dots, so names with domain punctuation are not addressable everywhere the product is installed.

Its handshake also gives the server's description, `AI backend: host, talk, track`, the short line a host shows beside the name in its connector list. The description names no product and no requirement, and is the same on every estate.

What a customer's own tool stores as the label for the connection is that customer's to choose and no requirement's to impose. This statement fixes the server's declared name and each published stanza's name, which are what Turn Zero controls here. Each stays apart from the name of every other MCP server Turn Zero publishes, each product's door and tzTeam's server among them. Each is a stable identifier on MCP-08's terms. A later change therefore reaches every affected address and published stanza in one release that serves the new name, never a quiet edit.

Which protocol revision the server speaks, and how it negotiates with older clients, is versioned call configuration beneath this contract, not a statement in it. Every request at the MCP route is answered as JSON. Those answers declare `charset=utf-8` in their Content-Type, so a client that applies a default charset to a bare media type decodes the server's UTF-8 bytes as written.

MCP-02 (decided; conforms: ACC-04): The authentication binding is OAuth 2.1 authorization code with PKCE, discoverable by the standard authorization-server metadata, with dynamic client registration so a customer's tool connects without a preconfigured client identity. The authorization moment is the sign-in page (WEB-L0-12): the builder realm's own surface (ACS-L0-01), served on the control plane's origin (ACS-L0-04), and the one browser interruption of CHI-L0-03, which the computer login page opens too, as a command-initiated case of that moment (WEB-L0-25). There a first sign-in creates the account where the builder realm's creation policy admits it (ACB-L0-77; ACS-L0-07) and is refused by name where it does not. The token the flow issues is the builder realm's session (ACS-L0-11), the credential the authority tiers bind to (API-L0-06). The connected session acts as that identity and nothing else, and disconnecting or signing out revokes the session rather than orphaning it.

A client_id no registration names, one the client sweep deleted or a development bake start dropped, is refused `invalid_client`. The token endpoint answers 401 before it reads the grant, so a host that clears its stored client on that refusal registers again. The authorize route answers the person's browser a page that says to remove the tool's sign-in for this server and connect again.

A minted token (API-L0-05) connects the same server as its bearer with no browser moment, the sign-in flow never run for it. The connected session then acts as the token's own bounded identity under the same challenge. The platform credential connects nothing here and draws the standing challenge (SEC-L0-07).

The authorized computer's token the turnzero command keeps (PLD-L0-104) is minted under the session this sign-in opened, by the computer login page's mint under the browser session or by the tool's `mint_token` call (MAPI-14; WEB-L0-25). Each is bound to the command's challenge and spent by its wire exchange (MAPI-23). The command holds no sign-in client of its own. `login` opens the browser on the platform's own page, and `login --device` prints a link to it for a person to open on any device; the page's cookieless arrival is the sign-in page returning to it.

Deployment configuration fixes one canonical authorization-server origin through `OAUTH_ORIGIN` and one protected-service resource identifier per MCP address; neither is inferred from a raw Host header or widened by an alias. The deployment template sets the origin from its trusted public API origin. An explicitly configured origin is validated before database migration or provisioning; an empty or malformed value refuses startup. Startup also refuses a missing `OAUTH_ORIGIN` when `PGHOST` has a nonempty value. Without `OAUTH_ORIGIN` or that deployment setting, a local executable selects the loopback origin from its actual bound listener, including an assigned ephemeral port. An application factory with no listener defaults to `http://127.0.0.1:8080` unless its caller supplies the trusted origin.

Each set origin serves its own `/mcp` resource: `OAUTH_ORIGIN`, `TZCLOUD_API_ORIGIN` for Hosting, `TZTOOLS_ORIGIN` for tzTools, and `TZOP_ORIGIN` for the operators' address. The reached host is the lower-case host the origin lock verified, matched exactly against those settings. A prefix, suffix or trailing dot matches no address. Two address settings naming one host, or one naming `OAUTH_ORIGIN`'s host, refuse startup. On each address's host, protected-resource metadata names that address's resource, the canonical authorization-server issuer, and the scopes `tzcloud` then `offline_access`. Sign-in stays on the authorization-server origin.

Changing the origin makes prior-origin builder OAuth access and refresh grants ineligible and requires new OAuth authorization; other credential kinds keep their own validity rules, and an existing valid browser session may authorize the new grant without another human sign-in. Each address's identifier names the same Turn Zero Cloud protected service: the MCP endpoint and the API, documentation and data-plane surfaces already admitting this bearer under ACS-L0-11; it grants no access to another service. Protected-resource metadata names the reached address's resource and the canonical issuer; authorization-server metadata repeats that issuer and its canonical endpoints. Both publish the supported OAuth scopes. Generated callback and return URLs whose destination is the control plane use its configured origin; the client's redirect remains its exact registered URI.

Authorization, code exchange, access-token records and refresh-family records preserve the resource, issuer, client and granted scopes. A supplied resource must match exactly; omission selects the configured service at authorization and retains the recorded resource at exchange or renewal, never a wildcard. The scope names and default are versioned call configuration. OAuth scopes are separate from authority tiers and application grants and cannot create either. The shared credential resolver rejects unbound or foreign OAuth access credentials before dispatch; a protected call with a valid OAuth credential lacking its required scope answers insufficient_scope with HTTP 403.

This paragraph qualifies the paragraph above's exact match for the plane's own resource. Wherever a supplied resource is compared with the plane's own configured resource, a value equal to it but for a query holding `products` alone reads as that resource and is recorded as it. That reading is no alias: the identifier recorded, advertised, and bound stays the one configured resource.

An explicitly supplied address resource matches exactly and carries no query; another value is refused `invalid_target`. Omission at authorization still selects the plane's own resource. Exchange and renewal retain the recorded resource, and naming another is refused `invalid_target`. The shared credential resolver admits a session bound to any configured MCP address on the REST and data routes, at renewal, and at the front's product-token exchange.

At `/mcp` alone, a session's resource must equal the reached address's resource, the company's origin included. A mismatch answers the standing 401 `authentication_required` challenge with `invalid_token`, the same body as a presented credential resolving to nothing, and the reached address's metadata. Every 401 and the 403 `insufficient_scope` on that route name the reached address's metadata. A minted token carries no resource and meets no comparison.

The operators' address admits only a credential holding a grant that a catalog row naming `tzop` carries as `grant` or names in `admits`, except `device_sign_in`. A row with no mark admits nobody by this rule. Every other credential meets the same 401 before any tool runs. Sign-in for that resource refuses an account holding none of those grants with `access_denied` and its state, minting no code and writing no `tool_authorized` event. The exchange reads the grants again and refuses `invalid_grant` if the account no longer holds one.

A session-resource request's scope is read the same way at authorization, exchange and renewal. An omitted or empty scope selects the documented default at authorization and retains the grant thereafter. An earlier name of the session scope, `turnzero_cloud` or `appcloud`, is read as `tzcloud`, while a grant recorded under one stays refused. A scope of another resource the plane serves, `tzmesh` or `tztrack`, is left out of the grant where the session's scope is also named (RFC 6749 section 3.3). So a client asking for the authorization server's whole advertised list is granted `tzcloud` alone, and the token answer names it. A scope this plane never served, an attempted broadening, or a product's scope alone is refused `invalid_scope`.

The scope `offline_access` is admitted beside a granted scope on every resource the authorization server serves, at authorization, code exchange, renewal and device authorization. It is left out of the grant, binding, refresh record and session, and never appears in a token answer's `scope`. Named alone, it is refused `invalid_scope` because it leaves nothing granted. Authorization-server metadata lists `tzcloud`, then `tzmesh` and `tztrack` where served, then `offline_access`. The plane's protected-resource metadata lists `tzcloud` and `offline_access`, in that order. The challenge's `scope="tzcloud"` and MCP-10's refresh-token rules stand unchanged.

Every MCP request the plane admits, presenting no credential or one resolving to no account, `initialize` included, answers 401 and a WWW-Authenticate challenge naming the protected-resource metadata and applicable scopes. An invalid presented credential also carries invalid_token and its error description; missing credentials carry none; the wire's library rows challenge as the context reads do (MAPI-11). The challenge's JSON body states in its detail the same case the header distinguishes: a connection that presented no credential, which reconnects with sign-in or starts the OAuth flow, or a presented credential that resolved to no account, which signs in again. Each detail says nothing was done, names the person's sign-in and then the call made again, and names `turnzero login`, the turnzero command's own sign-in (PLD-L0-104). Neither adds OAuth error information to the header, and the wire route's refusal states the same two texts.

This paragraph qualifies the paragraph above for a presented bearer of a grant's form that resolves to no credential. The challenge's body and the wire route's refusal state a third text in that case: the platform serves no grant of that value. The text names both ways to a fresh grant: the tool call whose answer carried it, made again, and the application's own request for one, made again (MAPI-03). The name, the status, and the header stand as the paragraph above states them for a presented credential.

Every body the server's 401 challenge answers carries `help`, the address of its refusal's row on the refusals page (MAPI-10): `authentication_required`, `end_user_credential_not_admitted`, `transfer_grant_not_admitted`, `secret_grant_not_admitted`, `egress_key_not_admitted`, and `app_credential_not_admitted`. Browser-only approval and all other credential-kind boundaries remain MAPI-03 and API-L0-07's.

Registration serves public clients with token_endpoint_auth_method=none, each naming at least one redirect. A client whose `grant_types` is the device-code grant alone may name none, and so never authorizes at `/oauth/authorize`. Authorization accepts only an exact registered HTTPS redirect or HTTP loopback redirect, with no URI credentials or fragment; code exchange requires S256 PKCE. An additive migration retains old rows as explicitly unbound. It may backfill only where recorded issuance provenance proves the binding, and production cutover otherwise requires the developer-approved one-time reauthentication disposition; an unbound row never means all audiences. The authorization-server metadata adds `revocation_endpoint`, the issuer's `/oauth/revoke`, and `revocation_endpoint_auth_methods_supported` with `none` alone. Public clients present no secret.

`POST /oauth/revoke` takes JSON or a form-encoded body carrying `token`, with `client_id` and `token_type_hint` optional. A named client must be registered, or the endpoint refuses 400 `invalid_client` before reading the token; without a name it reads the client from the token's row. Whoever holds the connection's refresh or access credential can end that connection. The hint orders the two lookups and never refuses. A current or redeemed refresh credential ends its family while that family has a live row, however many renewals the stored copy lags. A live builder session's access credential names its connection through the session's OAuth binding and its row's family. Without either, the session carries no connection.

Ending its own connection answers 200 with `contract_version` and `revoked` `true`; API-L0-21 states the end. An unknown or ended token, another client's under the client named, a browser session's value, and no token all answer the same 200 body, with `revoked` `false`. This other-client answer departs from RFC 7009 section 2.1, as the end-user realm's does (ACS-L0-14). A platform credential's prefix alone draws 400 `unsupported_token_type`, without checking validity: minted and product tokens, end-user sessions, transfer grants and egress keys.

Anyone may register a client, so a registration is bounded. It names at most ten redirects. Each is an HTTPS redirect, or an HTTP redirect on a loopback address, `localhost`, `[::1]`, or an address in 127.0.0.0/8, at any port (RFC 8252). None carries URI credentials, a fragment, or a wildcard in its host. A registration outside these bounds is refused `invalid_client_metadata` and stores nothing.

No redirect names a host the plane serves. That is a hostname equal to, or under, the registrable domain of the authorization server's origin, or of the tzTrack origin where the plane names one. It matches at any scheme or port and with any trailing dots, so `www.` and the service hosts match as the apex does. Two kinds of hostname are excepted: the `tzmesh` origin's and each `TZMESH_SITE_ORIGINS` origin's exact hostname, and an application's hostname under the serving suffix. On a loopback development origin the rule is that origin exactly. Authorization reads these bounds again over the redirects a registration stored. Where the registration falls outside them it refuses `invalid_request` at any of its redirects, so one stored before the bounds no longer authorizes.

A source address registers at most twenty clients within the hour, and the next is refused `rate_capped`. Each registration writes one info line to the control plane's record naming the client, the hosts its redirects name, and its registered name where it gave one, and never the source address.

A token the retired first-party client `turnzero-blueprint` was issued has no expiry date. It ends at `revoke_token`, after 30 days unused (PLD-L0-67), at sign out everywhere, at a passkey change's revert (ACS-L0-10), and at the account's suspension (API-L0-12), the last three each counting it in `exchange_tokens_revoked`.

Thirty-one paragraphs, this one and the thirty after it, qualify five rules above for the protected resources of Turn Zero's own products. The five are the token the opening paragraph names, the scopes the metadata publishes, the default scope and the separation of scopes from grants, registration's exact redirect match, and the binding's one grant, which the device grant joins for a client with no browser. Each product is one row of the plane's product table, and the table holds two rows, `tzmesh` and `tztrack` (each member's own requirements). A row holds the product's technical name, which is its profile name, its OAuth scope, its signing realm, and the word its token prefix carries. It also holds the setting that names the product's origin, and the paths the accounts service serves for the product.

Where the plane's setting `TZMESH_ORIGIN` names the `tzmesh` origin, the `tzmesh` resource is that origin's `/mcp`, and the authorization-server metadata advertises the scope `tzmesh` beside `tzcloud`. The `tztrack` resource is the origin `TZTRACK_ORIGIN` names, its `/mcp`; its scope is `tztrack` and its exchange is the tzTrack paragraphs'. The metadata lists `tzcloud` first, then `tzmesh` and `tztrack` where served, and `offline_access` last. Where a product's setting is absent the plane names no origin for it: the metadata stands as it stood, and a request naming the product's resource is refused `invalid_target` as any foreign resource is. The plane's origin and its configured MCP-address hosts serve their own resource metadata; each product's origin serves its own.

A registered client asks for the `tzmesh` resource by naming it in `resource` at `/oauth/authorize`, with the scope `tzmesh` or none, or a builder's application with `tzmesh.application`, a scope the metadata does not advertise. A request also naming `tzcloud` or `offline_access` is granted the `tzmesh` scopes alone at authorization and exchange, as a session-resource request is granted the session's (RFC 6749 section 3.3); another scope is refused `invalid_scope`. After the sign-in moment a tool's request, one naming the scope `tzmesh` or none, holds no tzMesh page, and the paragraph below states its completion. The `tzmesh` site's own sign-in client holds no page and a builder's application one line at its first sign-in alone, as the paragraphs on each state, and a prefetch holds nothing (WEB-L0-24).

Where a tool's redirect is not a loopback address, its request meets the one consent screen a tool's sign-in shows (CHI-L0-03), whose Allow access completes it. Where its redirect is a loopback address, it completes at the sign-in's completion, or at the allow-access screen's Allow access where the browser's session signed in within WEB-L0-16's window (CHI-L0-03). At a tool's completion the plane mints the single-use code for the connection's `tzmesh` token and redirects it with the `state` to the client's registered redirect. The one screen names the account by its sign-in address and the tool by its redirect's hostname with no port, its line naming the tzMesh messages alone and none the applications, and its Cancel returns `access_denied` and the `state` and mints nothing. The Terms of Service's section 15, Each product, carries the agreement that the tool may send and read the person's tzMesh messages.

A tool's completion and the exchange read the account's profiles through the plane's one rule (ACB-L0-76). It answers `tzmesh` for each member of the invite group while the plane's setting `invite_only_products` names `tzmesh`, for every account while it does not, and for an operator account always. In these paragraphs an account in the group is one the rule answers `tzmesh` for. An account outside the group is returned to its client with `access_denied` and the `state` at the completion, with no description and no page, and is refused `invalid_grant` at the exchange; neither answer names the product.

The token endpoint exchanges that code for one signed `tzmesh` token at the level `admin` under the client's name, issued to the client, and never for a session (MAPI-14; MAPI-03 states the signed form). In the same act it revokes the client's own unrevoked `tzmesh` token of that label, and every other holder's token of the label stands. The answer carries the value once, with `token_type`, `scope`, `level`, `label`, `expires_at`, `origin`, `resource`, `token_id`, and `replaced`, the identities it revoked, under `Cache-Control: no-store`. It carries no refresh token and opens no connection (MCP-10). The exchange reads the generation again after the mint, and a token minted while it moved is ended, its end written as `revoke_token`'s line, and never answered. The token expires 180 days after its mint and ends as API-L0-05 states for it.

The plane also serves the device authorization grant (RFC 8628) for its own resource and each configured customer MCP address, whose exchange answers the builder realm's session. It serves the grant for the `tzmesh` and `tztrack` resources too, each exchanging for a signed product token. A registered public client posts `client_id`, `resource`, and `scope` to `/oauth/device_authorization`. A session resource reads scope as a session-resource request does. A product resource admits its scope or none; a request also naming `tzcloud` or `offline_access` gets the product's scope alone, as at `/oauth/authorize`. The answer carries `device_code`, `user_code`, `verification_uri`, the plane's `/oauth/device`, `verification_uri_complete`, `expires_in` 600, and `interval` 5. Another resource, the operators' resource or none is refused `invalid_target` with no product word; another scope is refused `invalid_scope`.

The user code is eight characters from an alphabet of twenty consonants, printed in two hyphenated groups. The metadata advertises the endpoint and the grant on every estate, since the plane's own resource is always served; a product's resource is served where the plane names its origin.

The verification page is the plane's own sign-in moment for the request the code names, and holds the resource's rule after it. For a session resource that is the page alone, naming the platform and no product, and for `tztrack` the page alone, its gate read at the exchange. For `tzmesh` it is the group check (ACB-L0-76). The page shows the code and the client's registered name and offers Allow access and Cancel alone, Allow access taking no click until the pointer has moved over the page or a short delay has passed (CHI-L0-03). The page and its posts answer only the browser whose sign-in held them, no other site may frame them, and a prefetch holds nothing.

For each admitted resource, Allow access writes the approval onto the request, and Cancel denies it, a page that refuses the account writing nothing but offering Cancel; an unanswered request expires. A code the page does not know is told apart from no other. The token endpoint answers the poll as MAPI-14 states. The device code stands in PKCE's place: only the asking client holds it, and it is answered once, the request's row leaving at that answer; the daily pass deletes an expired one (PLD-L0-67).

For a session resource the exchange issues the builder realm's session with its refresh token, as the authorization code's exchange does (MCP-10; ACS-L0-11). The session's binding carries the bounding grant `device_sign_in`, which every row of its refresh family keeps through rotation and its grace. The session holds no `super_admin`, even for an operator's account, since that grant lifts every bound (MAPI-09). The bound admits it to the rows whose `admits` mark names the grant, the rows marked `access: anonymous`, and the forwarded rows MCP-14 states (MAPI-16). Every other row refuses it, an unmarked row `device_sign_in_bounded` and a grant-marked row `grant_required`. So a sign-in a script was tricked into deploys nothing, exports nothing, reads no storage area, mints no wider token, and destroys nothing without the person's approval in the browser (MCP-07).

The accounts service publishes each product's key set at `/accounts/v0/<product>/keys`, every key with its stamps, and its revocation list at `/accounts/v0/<product>/revocations`, each anonymous under the admission window (PLD-L0-67). The `tzmesh` pair is `/accounts/v0/tzmesh/keys` and `/accounts/v0/tzmesh/revocations`, and the `tztrack` pair is `/accounts/v0/tztrack/keys` and `/accounts/v0/tztrack/revocations`. Each accounts process that starts while the plane names a product's origin and that product's key set holds no key mints a key into it, so the set is served ahead of the first token; the read writes nothing. A set holding any key is read as it stands.

The list answers each revoked token's identity and revocation stamp for the product, an account row for each account the product was turned off for or whose tokens the operator ended, and the account's standing rows (API-L0-12). It answers them in one `revoked` array ordered by the stamp then the identity, at most 1000 rows. Those are every one standing where the query names nothing, or from `since`, a UTC instant, inclusive. `after`, sent with `since`, names the last identity read at that stamp, and the page starts after it; a full page answers `next`, the `since` and `after` the next read sends, and a short page answers null.

An account row stands until an invitation's redemption returns the account to the invite group, or until the retention an `all` row keeps has run from its instant, and the plane removes it at the first of the two (ACB-L0-76; PLD-L0-67). A later closing carries its own instant, and the account's token rows keep their own retention. A product's service applies an account row once per instant, as its own contract states. These rows and the key set are each product's contract's first generation, with the exchange's answer for `tzmesh`, and system:XTX-22 states how a later one is introduced for `tzmesh`, and system:XIT-24 for `tztrack`.

Beside those rows a product's list answers an `all` row for each revocation of that product's signing key set, its identity `all:` and its instant, for the product's service to apply as every account's closing (MAPI-09). A revoked token row stands on its product's list until its expiry has passed and the retention PLD-L0-67 states has run, so the list carries every revocation for as long as the token's signature could verify. An `all` row stands until every token signed before its instant has expired and the same retention has run. A product's service reads its list once a minute, as its own contract states. Where the plane names no origin for a product both of its reads answer their zero forms, an empty key list and an empty revocation list, whatever rows the store holds.

Every product's list answers the two standing rows API-L0-12 states, each held under its reserved identity in the account's place as the `all` row is, one per account, answered on every list. Suspension writes `{"kind": "suspended", "id": "suspended:<account>:<instant>", "revoked_at": <the instant>}`, and reinstatement writes `{"kind": "reinstated", "id": "reinstated:<account>:<instant>", "revoked_at": <the instant>}`, so the later row replaces the earlier of either kind. A service reads the account from the identity and applies the rows in list order: while the suspended row stands it refuses the account's credentials its own contract names (tzmesh:TZT-L0-04; issue_service:ISV-06), and the reinstated row lifts that, the credentials kept. The suspended row stands while the suspension does; the daily pass removes a reinstated row once the retention PLD-L0-67 states has run from its instant. Each kind is the additive change system:XTX-22 admits for `tzmesh`, its cases the tzMesh package's, and system:XIT-24 for `tztrack`.

The client's name is its registered name, or `MCP client` where it registered none or one not printable. No `tzmesh` token carries a handle, since the `tzmesh` service holds the account's handle. The erasure walk's ten-minute token carries `act` with the value `erase`, which the `tzmesh` service admits at its erasure route alone (ACB-L0-47). No other endpoint or wire call admits a token carrying an act. The MCP front's two approval marks, `describe` and `pending_action`, are no act: a product's door admits each on the one forwarded call it marks, as MAPI-03 states (MCP-14; MCP-07).

For the `tzmesh` site's browser, the plane answers the origins the setting `TZMESH_SITE_ORIGINS` lists (PLD-L0-93) with CORS headers on the authorization-server metadata, the registration, and the token endpoint. The preflight allows `POST` and `Content-Type` with no credentials, a listed origin's `/mcp` is admitted as the `tzmesh` resource, and with the setting empty the `tzmesh` origin alone is listed. That fallback is for these OAuth reads. The setting also supplies the Settings frame page's framing origins, which PLD-L0-69 derives without a fallback; an absent or empty setting permits no framing origin (WEB-L0-21).

The daily client sweep (PLD-L0-67) deletes a registration past its age that nothing references, a client that never issued a refresh token among them. It keeps the `tzmesh` site's own sign-in client: a registration the setting `TZMESH_SIGNIN_CLIENT_IDS` lists (PLD-L0-93) whose every redirect is an HTTPS origin's `/signin/callback`, exactly. The operator lists the id the estate's own registration answered, so the `tzmesh` service's sign-in client is registered once per estate (the tzmesh member's configuration statement). The sweep keeps no registration by its redirects alone, since anyone may register one, and with the setting empty it keeps none by this rule. An unrevoked `tzmesh` token issued to a client, expired or not, references it, so a builder's application stays registered while a person's token stands.

At the sign-in, the `tzmesh` site's own sign-in client is a registration whose every redirect is the `/signin/callback` of the `tzmesh` origin or a `TZMESH_SITE_ORIGINS` origin, exactly (the tzmesh member's sign-in statement), and one `TZMESH_SIGNIN_CLIENT_IDS` lists where it lists ids. Any other registration is read as every other client is. Its request holds no page: at the sign-in's completion, for an account in the group, the plane mints the single-use code at once and redirects it with the `state` to the registered callback. An account outside the group is returned `access_denied` and the `state` at that moment, with no description and no page, as a tool's completion returns it.

Where `TZMESH_SIGNIN_CLIENT_IDS` is empty, anyone may register such a client, and the code still reaches only that callback, where the `tzmesh` service completes only a sign-in it started, under its own client and the code's PKCE verifier.

This paragraph qualifies the paragraph above's sign-in. A browser whose site session signed in within the freshness window the allow-access screen reads, eight hours, completes the `tzmesh` site's sign-in client's request from that session with no page, the group's rule applying as at the sign-in's completion.

This paragraph qualifies the exchange paragraph above for the sign-in client's code. It exchanges for one `tzmesh` token at the level `chat` under the label `Sign-in at <host>`, the callback's host, issued to the client. Its life is one minute from its mint, in `expires_at`, on its row, and in the payload's `exp`. That covers the service's exchange and its one `register_human` call. Where a same-origin post of the site started the sign-in with an intent, it also covers the one `register`, `attach`, `remove_listeners`, `delete_agent`, or public `create_channel` the callback makes for that intent in the service's own process. The service then drops the value and keeps it nowhere, so no sign-in row stands 180 days. The replacement rule keeps one unrevoked row of the label standing for the client per account, and the exchange's other rules stand as written.

This paragraph qualifies the request and completion paragraphs above for a builder's application: a request naming the scope `tzmesh.application`, the `tzmesh` site's own sign-in client excepted. Its request meets no consent screen. An account's first sign-in to that client holds one line, served by the accounts service under `/oauth/tzmesh/`. The line names the client by its registered name and its callback's host and says the client will see and send the person's messages as them, with Continue and Cancel. The line offers no other choice. The line and its posts answer only the browser whose sign-in held them, no other site may frame them, and Cancel returns `access_denied` and the `state` to the client's redirect. Where the account holds an unrevoked `tzmesh` token the client was issued, expired or not, a later sign-in holds nothing and mints the code at once, and a prefetch holds nothing.

A request without that scope, one naming `tzmesh` or none, meets the one consent screen wherever its redirect is not a loopback address, so a hosted MCP host is named to the person and allowed there. Anyone may register a client and name the scope, so the line, naming the host the code goes to, is what keeps a page offering a Turn Zero sign-in from taking a person's messages with no word of what the sign-in gives.

This paragraph qualifies the exchange paragraph above for a builder's application's code. It exchanges for one `tzmesh` token at the level `chat`, issued to the client, under the label `<name> at <host>`. The name is the client's registered name, or `An application` where it registered none or one not printable, cut so the label keeps its callback's host within the label's bound. The token's life and the replacement rule stand as the exchange paragraph states them. An account outside the group is returned `access_denied` at the line's moment and refused `invalid_grant` at the exchange, as at a tool's completion. The token's payload carries `application_sign_in` true, which no other code's token carries (MAPI-03), and the `tzmesh` service reads it as system:XTX-02 states.

This paragraph qualifies the opening paragraph's token, the resource rule, and the protected-resource metadata for the tzTrack door's protected resource (tztrack:TRK-L0-02; ACB-L0-89). Where the plane's setting `TZTRACK_ORIGIN` names the tzTrack origin (PLD-L0-93), the tzTrack resource is that origin's `/mcp`. A registered client asks for it by naming it in `resource` at `/oauth/authorize`. The plane binds no builder realm session to that resource and renews none: the request exchanges for the signed `tztrack` token, as the next paragraph states. A refresh presenting a row bound to the resource is refused, `invalid_target` where it names another resource and `invalid_grant` otherwise.

The request names the scope `tztrack` or none, a request also naming `tzcloud` or `offline_access` granted `tztrack` alone, and its completion is the session branch's own, with no consent page of its own. Where its redirect is not a loopback address, the allow-access screen's tool access form holds the code until its Allow access, as for a session (CHI-L0-03). The code then exchanges for one signed `tztrack` token at the level `admin` under the client's registered name, on the `tzmesh` exchange's form (MAPI-14). The product gate is read at the exchange through the one rule (ACB-L0-76): an account outside it is refused `invalid_grant` with no product word, and the refresh arm refuses the resource.

The door at the tzTrack resource and its protected-resource metadata are the issue service's (tztrack:TRK-L0-02; ACB-L0-89), and the plane serves neither on that host. The plane's own surfaces admit the `tztrack` token within its grant alone (MAPI-16) and bind no session to the resource. The metadata the plane's origin serves stands unchanged. Where `TZTRACK_ORIGIN` is absent the plane names no tzTrack origin, and a request naming the resource is refused `invalid_target` as any foreign resource is.

The `tzmesh` list also answers `{"kind": "site_session", "id": "site_session:<reference>", "revoked_at": <the instant>}` for an ordinary browser sign-out that presents a browser session cookie that verifies (ACS-L0-18). The instant is that of the latest such sign-out; a repeat updates the same row. The reference is the token's `site_session` claim, never the session's secret or row identifier. At its next poll the service ends every token carrying it and every stream opened under one (system:XTX-22). These rows are purged with the closings in both stores under the same retention (PLD-L0-67).

## The three forms

MCP-03 (decided): Every management action callable by an AI client is a tool corresponding to one API capability. Its authority tier — observe, reversible or destructive — remains explicit in the authoritative catalog and enforced by the management API (API-L0-06); a host annotation neither grants authority nor supplies approval. Each tool advertises truthful standard MCP annotations derived from its action’s catalog metadata, under MCP-11. Served readable context retains its addressable resources, and skills retain their prompts. The same content is also available through the observe tools CTX-07 defines, so resource and prompt support enriches a client without being required for discovery. These protocol forms derive from the same API and content sources; none owns an independent capability or copy of the content.

MCP-04 (decided): Tools alone are the floor: every launch scenario’s management steps complete through tool calls and the browser moments CHI-L0-03 states, with no management step that requires resources or prompts. The signed-in site (WEB-L0-15) is a view of the same acts, required by none. Deployment starts from the externally prepared and uploaded artifact CHI-L0-01 requires, or from the files `deploy_files` carries (API-L0-25); deploy receives its account-owned storage reference and hash, without requiring a filesystem or byte upload in the MCP client. A client supporting only tools loses the slash-command surface and the preloadable context, never a capability — discovery’s catalog and complete readable content arrive through read_context and read_documentation there, and the served documents remain readable outside the protocol (API-L0-09's other two forms). This is what makes "any AI tool's output is a first-class citizen" true of the tool's *client features* too.

MCP-05 (decided): Discovery is the protocol's own listings — tools, resources, and prompts, each entry carrying a description a newcomer can act on — plus the overview. The overview, available as both a resource and a read_context result, answers "what can the platform do and where do I start" in one read. Every listing description derives from the same one source as the human documentation and the authoring bundle (API-L0-09), so what the tool shows, the docs say, and the bundle teaches cannot disagree. An entry of the words override file (PLD-L0-114) is one exception, while it stands. The rule extends to library catalog rows: a catalog row's listing description derives from the package-owned selection summary the catalog quotes (FTR-L0-57's section), the same one-source discipline over the library's source.

The rule's other exception is MCP-08's listed-name step. On the MCP route the texts MCP-08 names carry the listed names its step writes. The wire, the authoring bundle's guide copies and its `source:` entries keep catalog names. A served text and its one source then differ in those names alone.

A served string names a product by its formal name or its short name, tzCloud, tzTeam, tzTrack, tzMesh, or tzDocs, at its writer's judgment, or names the company, Turn Zero, or the platform (azure:AZD-08), and never a code name or an internal product name. This is the product-name half of the public-name rule the documentation suite and the served-surface suite share. A served string is every tool title, every tool description and argument description, every resource description, every prompt description and prompt text, and every text the server composes — the overview, a refusal's detail, an action's response detail. It is also every field of the enumeration files the bundle and the `service_contracts` resource carry, the bundle's guides, and the selection summaries the catalog quotes.

A forwarded row's strings are served strings too: its title, its description, and its argument descriptions, as a product's door serves them under the product's prefix. The plane's served-surface suite holds them to this rule through the door's recorded list (MCP-14).

The plane's own tzTrack rows follow the door's rule for plain words too (tztrack:TRK-L0-03). Those are the actions whose management row names the `tztrack` surface, `list_issue_spaces`, and `submit_feedback`. The texts held are their titles, summaries, and references in the tool catalog, their argument descriptions and retired-argument refusals, the tzTrack practices header, and the service map's tzTrack line. None of them holds a working word of the issue service or an order form aimed at the AI. Both lists stand in one shared module, which the door's suite and the served-surface suite read.

Two strings keep the long form, each fixed by its own statement: the `X-Powered-By` header (PLD-L0-67) and the apex page (SVC-L0-07). The skills file's strings take this rule at the served-skills landing that rewrites them, a requirement pending there until it lands (RLC-L0-92; MCP-06).

The requirements documents the bundle packages verbatim are the source itself (CTX-02), and CTX-08 holds that source to the same product-name half. A clearly labeled reproduction of an indexed contract, enumeration, or guide in the generated human reference therefore carries the approved names because its source does. A reproduction preserves its source's identifiers. The reference distinguishes resolvable links within the published index from citations to sources that are not published there, without publishing additional sources merely to satisfy a citation.

No listing description carries a requirement or decision identifier: a tool's, an argument's, a resource's, a prompt's, or a catalog entry's summary read_context with no id answers. No line of the overview carries one either. A catalog row names the statements its descriptions follow in its `scenario` and `owners` members (MCP-08), which reach a client in the whole-file reads alone. The served-surface suite holds the listing descriptions and the overview to this paragraph.

No other text served to a customer's tool carries a requirement or decision identifier either: the instructions, an action's answer and refusal, a composed `read_context` read, and a skill's page, part, and prompt description. A read that reproduces a source keeps its identifiers: a whole-file specification read, a `read_documentation` answer, and a `search_documentation` answer. So do CTX-09's line for a query that is one rule code, a forwarded product's answer, held to this rule at its door (MCP-14), and the top-level prose of an operator's completed answer. A rule code is a statement's identifier, met in those reads and in the documentation. The served-surface suite holds the served strings to the product-name half and to this paragraph.

The rule that text served to a customer's tool carries no requirement or decision identifier also reaches the text the platform composes on a route that a customer's application, its users' browsers, or its scripts call. That text is a refusal's detail, an answer's detail, and a page's own words, in any body or page the route answers. Storage, egress and its upstreams, accounts and sign-in, push, email, logging, deploy progress, and the serving router are among those routes. Customer data passes as it is: a stored object, a declared upstream's or the application's answer, a log line, and a name the caller sent that a detail echoes.

On those routes a reproduction keeps its identifiers, as a reproducing read does. A Turn Zero product's answer through its upstream is held to the same rule at its door, as a forwarded product's answer is, since the gateway passes it unchanged (EGW-L0-03). The rule also reaches the text a library package's modules compose, which run in the customer's application: an error's message and a refusal's detail reach its handler, its logs, and its scripts as the platform's own words do. The tzMesh and tzTrack services' own routes and the front-end packages' modules take the rule at their own landings, a requirement pending there until each lands (RLC-L0-92).

Beside the overview, the server's instructions name, on every connection, the estate's configured origin (MCP-02), never one read from a request. They name what a script reaches under a credential: each management action at its reference page's address, the library's reads among them, and the storage routes. Before the line naming the origin, which stays their last, the instructions say that a management action's refusal may carry `help`, the address of its row on the refusals page at `<origin>/cloud/reference/refusals/`, which explains the refusal. That origin is the public origin `help` is composed on, so an agent meeting a refusal name no `help` rides with, in a detail, a data plane's body, or a page, still learns where the name is explained.

The overview, as a resource and as a `read_context` result, carries the same origin line.

Each tool's summary opens with one sentence of at most 200 characters saying what the tool is for, so a host that cuts a description short still shows what the tool is for.

This paragraph qualifies the origin paragraph's instructions. On a connection listing `read_status`, they also say, ahead of the refusals page's sentence, that a transport error or a gateway's error gives no outcome (MAPI-04). A read is repeated, a send with its same `client_id`, and another write after reading the state. For the acts MAPI-04 holds they name `read_status` with `wait_seconds` as the read to make first. Such an act is repeated only where that read shows it did not start, never where it cannot tell, and the overview says what in that read shows a start. A schedule run is repeated only with its same `request_id`, whatever `read_schedules` shows. The text names no identifier. The overview carries it under the same condition, with what shows the act started, and `deploy`'s description says what shows a deploy started and that an upload `pending_upload` still names was not.

Where `tzmesh_send` lists and `read_status` does not, the instructions instead say that a transport or gateway error gives no outcome, reads are repeated, and `tzmesh_send` is repeated with the same `client_id`. Other writes follow a state read. This form names no tzCloud tool and precedes the refusals page's sentence. The standing form and its overview condition stay unchanged wherever `read_status` lists.

This paragraph qualifies the origin paragraph's instructions too. On a connection that lists `read_context`, they open with the service map's tzCloud line where the address lists tzCloud's tools (MCP-01), and otherwise with the sentence naming `start`. That line says in plain words what the platform does, with no claim about it, and the prefix its tools carry. The opening then names `start` as the first call, where the connection lists it. It names the overview, with the call that reads it, as the read for a platform rule the task needs, and not as a read before every first call. A connection that lists `read_context` and not `start` is sent to the overview alone. The opening names no identifier. On a connection that lists `start`, the overview says that one connection reaches the products the account holds.

This paragraph qualifies the opening. On a connection that lists `start`, the opening names the tool index, the `tools` entry CTX-07 states, after the sentence naming `start` and ahead of any longer guidance. It names the index as the read for finding a tool, not as a read before every call, so `start` stays the first call. It gives the call that reads the index whole on every customer's connection: `read_context` with id `tools` and `limit` set to the readers' `text_max_limit`. The opening on a connection that lists no `start` is unchanged. The sentence names no identifier.

The service map's product lines follow the tzCloud line where it stands, one for each product whose rows MCP-14 lists to the connection, and none on any other connection. Each says only what the product is for, its prefix, and where its overview is read, the product's practices entry `overview:<product>` (CTX-07), whose header says the rest. The lines name no identifier.

At every MCP address the service-map lines name only products that address lists and the account holds. No product-off line is served. The feedback rule stands only where `submit_feedback` lists, so it is absent on the operators' address. Every address keeps the estate-origin and refusals-page lines; the operators' address serves those two alone. The mobile and Node.js lines keep their listing conditions.

On a connection listing any of `deploy`, `deploy_command`, `deploy_files`, `promote`, `read_logs`, `read_schedules`, `read_status`, `read_usage`, `restart_application`, `roll_back`, `run_schedule`, or `store_secret`, the instructions say once why `wait_seconds` is bounded. They tell a caller whose client allows less than sixty seconds to pass fewer seconds on this call and on a returned `next` call. The line follows the lost-answer sentences where those stand and precedes the Node.js sentence where that stands. It carries no requirement or decision identifier, and the overview does not carry it.

This paragraph qualifies the origin paragraph's instructions too. On a connection that lists `deploy`, they say, after the sentences on a lost answer where those stand and ahead of the refusals page's sentence, that a command this platform answers needs Node.js 24 or later, the version PLD-L0-82 holds every stated number to. They also say that the Ways to deploy to Turn Zero Cloud page, read through `read_documentation`, compares the ways to deploy API-L0-25's route selection chooses among. The text names no identifier, and the overview does not carry it.

MCP-13: **Each text the server lists is held to what a client shows of it: a tool's title, a tool's summary, a tool's served input schema, and the server's instructions; a row's `reference` holds what its summary cannot.** A caller then reads the whole of what the listing says. A title, a summary, and the instructions, each counted in UTF-16 units, and a served input schema, counted in bytes of compact JSON, run to at most the figure the catalog's `listing_bounds` member holds for each, a row that member holds by name excepted. The instructions are measured at their longest composition on each estate's origin. That member of `schemas/mcp_surface.json` is the one home of the six figures, and family:Q-302 holds what each client shows of a listing, the versions it was read at, and the margins.

The whole listing served to a builder's connection holding Turn Zero Cloud and tzTrack alone, as every account does (ACB-L0-76), counted as the `tools/list` result in bytes of compact JSON, runs to at most the sixth figure, `whole_listing_bytes`. It is measured at its longest composition on each estate's origin, with the members each estate's committed settings open. That figure is a bound with no client limit behind it. It rises only where a landing's listing would exceed it, to that listing plus 2,000 bytes of headroom. It may also be lowered, but ordinary growth within the bound moves nothing. It moves only by a recorded write naming its reason, and never by unlisting another tool to admit a later one.

Every listed tool carries a title, the short name a person reads where a host shows the tool, such as a connector directory or an approval prompt. The catalog row's `title` member holds it. The server serves it with the tool, and again as the `title` member of the tool's annotations, beside the effect annotations MCP-11 states, because Anthropic's connector directory reads it there. A title is in sentence case, names what the tool does in plain words rather than the platform's internal terms, and is unique across the catalog. It carries no identifier.

A forwarded row's title is the form MCP-14 gives it, the product's short name and the door's own title, `tzTrack: Find issues`, unique beside the catalog's titles. The served-surface suite's sentence below holds the composed listing, the forwarded rows included, through each door's recorded list (MCP-14).

A forwarded row's neighbour among the plane's own rows is the one a caller might reach for in its place. A door's description cannot name the plane's tools, so the neighbour is named in the header of the product's practices entry (CTX-07), the read the product's service-map line names (MCP-05). For tzTrack, the header says that `tztrack_raise` files in the person's own space and that a problem with Turn Zero itself goes to `submit_feedback`.

A summary holds what a caller cannot learn elsewhere in time. It says what the tool is for, in its opening sentence (MCP-05). It says when to call the tool and not a neighbour, and the acts before and after it where the call is one step of several. It says what the call changes, provisions, costs, or cannot undo, and the approval it waits for. It states a rule on how to call the tool that no single argument owns, and a value in the answer whose meaning the answer does not carry itself. Where the row carries a `reference`, the rest is on the action's reference page, whose path the next paragraph says where to find. A sentence another statement places in a description stays in it.

The rest of a row's text stands in the row's `reference` member, which the listing does not serve. The action's generated reference page renders it after the summary. No summary names that page. `read_documentation`'s own summary says once where an action's reference page stands, so the route stays in the listing every host shows, and the `tools` entry's first line says it again (CTX-07); the listing repeats no such sentence. A row's `owners` names the statements its `reference` follows too, and a `reference` carries no identifier, as a summary carries none.

An argument's description, counted in bytes, runs to at most the figure `listing_bounds` holds for it, a description that member holds by name excepted. It keeps what the argument means and the values it admits where the schema does not state them. A schema bound is stated as the bound, with no refusal code; `invalid_request` is named only for a rule the handler checks, since over MCP the client's argument check answers first.

What else it said stands in the row's `reference`, under the argument's name. A description that cannot hold those, and any sentence another statement places in it, within its bound is named in `listing_bounds` with its reason and the size it is held at. A row whose served input schema cannot meet its bound with those in the listing is named in `listing_bounds` with its reason and the size it is held at, and is cut no further. The suite refuses a named row that fits the bound, a named row over its held size, and an unnamed row over the bound.

The served-surface suite holds each bound against the listing a connection receives, the `tools` entry's first line against the reference pages' path rule, and every summary to naming no reference page of its own.

MCP-06 (decided): Each skill surfaces as one prompt, invocable as a slash command in clients that support them, and the prompt set is derived from what Turn Zero Cloud offers (API-L0-10). The server enumerates it from the same source as everything else, and no hand-kept prompt list exists to drift. A product door's relayed prompts (MCP-14) are the door's own, listed beside this set under the product's prefix; the set itself stays the plane's. A tools-only client obtains that same skill text through read_context, with the same implementation and credential-grant filtering (CTX-07); the tool path does not author a second recipe. A skill lists only where every action it drives lists (MCP-08).

Each skill row carries a `must` member, an array of imperative sentences stating what the implementer must do, each ending with the identifier of the statement that owns the rule. The served skill text carries every item on exactly one of its pages under its list number; the human reference page places the whole list ahead of the whole summary. The prompt's description carries the summary's first sentence and then the first page less its title, actions, and footer; the prompt's text is the first page. The catalog entry for a skill carries no must item: it names the skill, its `read_context` identifier `skill:<name>`, and a one-line purpose, the summary's first sentence with each parenthetical holding an identifier dropped whole. No rendering restates a rule the list omits or reorders the list; each driven action is one line, its name, its tier, and its tool summary's opening sentence.

A skill row may carry a `short_path` member: a lead, a few steps for the common case, each naming by number the items it follows, and the items applying only where a condition holds. A scope naming a `part`, a slug, or a row marking its reasoning as a part, pages the skill: a first page under `skill:<name>` and parts under `skill:<name>:<part>`, the summary the part `reasoning`. The first page holds the short path, a routing line per part with its identifier, condition, and items, and every item no part holds; a part holds its condition and its items. A one-page skill, and a paged skill's first page, fit the readers' default text limit, and each part fits the maximum. A path naming a number outside the list, scoping an item twice, or breaking the measure fails generation.

A skill row may carry a `page` member: the route of the documentation page that answers its task, as the Cloud tree's index prints it. An overview line gives the skill's name, any credential suffix and the one-line purpose; with a page, `; page <route>.` replaces the purpose's closing full stop. A test holds each `page` to a route a page of the corpus declares. A line with no page ends at the purpose's own full stop. The overview's Where to start header says once to read a skill with `read_context` under `skill:<name>`. The list closes with one line naming the documentation page that maps tasks to guides.

This paragraph qualifies the third paragraph's first page. An item a scope holds there carries its condition ahead of its text, unless the item opens with "Where" and the condition's words. The reasoning part's routing line and catalog summary name it as the reasons behind the rules, and a scope naming its part `reasoning` fails generation. The human reference page, and a one-page skill's text where its row carries a short path, place the short path ahead of the list, a scoped item carrying its condition there as on the first page.

A part carries the line of each action that its items alone name, an item naming an action where the action's identifier stands in its sentence as a whole word; the first page carries the lines of the rest.

Every rendering of the list, the served pages, the prompt's description, and the generated skills reference, prints each item's sentence without its owner parenthetical. The served pages, parts, and prompt description close with no Rule owners list, so they print no identifier (MCP-05). The generated reference's section for the skill closes with the whole row's list: one head line, "Rule owners:", then one list item per item, "- 1: CHI-L0-07". So an item's sentence and its identifiers are two units under the generated reference's sentence bound. Where the first page's list skips numbers a part holds, it prints one line per run of skipped numbers at the run's place, "Rules 19 to 27: skill:deploy:stored-key" and a run of one "Rule 37: skill:deploy:web-client", so the numbering reads whole. The row's `must` member keeps every identifier at its sentence's end, and the three renderings share one implementation.

The fixed footer that closes every skill's first page, with its sentence on completing a destructive action through the browser approval, is the platform's own text and not a rendering of the row, so no row's list owes an item for it.

Each `must` item is one sentence of at most 40 words before its closing owner parenthetical and full stop; only that parenthetical holds identifiers, joined by semicolon and space, and code marks enclose only catalogue action names, and it cites no rule by number. It opens with none of "A", "An", "The", "Each", "Every" or "No", cites no other skill's step by number, and, except `read_documentation`, `submit_feedback` and `start`, names only actions its row drives or the suite's per-row list admits as mentions that direct no call. A `short_path` member's lead, scope conditions and steps, except each step's closing rule list, name no rule by number and no other skill's step by number.

MCP-07 (decided): A destructive-tier tool never destroys: it creates the pending-action record and returns Turn Zero Cloud's description of what will be destroyed with the approval link (API-L0-07), and the outcome is readable through the surface once the browser click completes. No tool, resource, prompt, or other capability of this surface approves anything — at launch or ever — and the suite refuses a catalog that grows one.

A destructive door row the front lists (MCP-14) takes the same route. A connection whose credential holds no destructive class is answered MAPI-12's class refusal by name, and nothing is described. The front reads what the call would destroy from the product's own door, under a product token marked `describe` (MAPI-03), creates the pending action with that description, and answers the approval link. That description is the product's own first-party reading, never a model's paraphrase; for a door row it stands where the first paragraph names Turn Zero Cloud's description. After the browser's approval the executor describes again, checks the fingerprint, and forwards the call once under a token marked with the pending action's identity. The door's answer, or its refusal, is the outcome, read by `read_approval_status` (API-L0-07). A door's refusal of the describe is relayed as served, and nothing is held.

## The launch catalog

MCP-08 (decided): The launch surface is enumerated in `schemas/mcp_surface.json`, revision 11: each capability's name, a tool's title (MCP-13), form, tier, the launch scenario it serves, and `owners`, the statements its served description and its arguments' descriptions follow. Host effect annotations derive from the corresponding management action row under MCP-11, and the title in a tool's annotations from its own row under MCP-13. A shipped name is a stable identifier: renaming or removing one is a versioned contract change, never a quiet edit. A capability added later arrives as a catalog row with its form and tier assigned under MCP-03, which is how the destructive boundary stays enumerated per capability rather than judged per incident.

Every listing rule below for a plane catalog row has one last condition: the row's `addresses` member must name the reached address, `tzcloud`, `tztools` or `tzop` (MCP-01). At the company's origin it must name `tzcloud` or `tztools`, whether the estate serves those addresses or not. A missing or empty member admits no address. Anonymous rows, bounded credentials, device sessions, staff sessions, skills, payload entries and the tool index take the same condition. On the wire, the tool index follows tools/list at the company's origin (CTX-07); skill and payload entries read no MCP address. The listing changes no wire gate. A call naming a row listed at another address meets the unlisted-row answer.

The public catalog view shows `addresses` on each unmarked row it lists, including a value `tzop` where the row names it. The member is data about the surface, not an access boundary. A row marked only by a grant stays out of that view under CTX-08.

A row's `surface` member in schemas/management_api.json is `tzcloud`, `tztrack`, or `account`. It is the row's `product` gate where one is set, the profile word `cloud` read as `tzcloud`, and `relay_issue_act` is `account`, the one gated row that departs from its gate. Of the rows with no gate, `delete_space`, `redact_issue_text` and `set_ai_passes` are `tztrack`, `read_library` is `tzcloud`, and every other is `account`. A row's MCP name derives from it: `tzcloud_<name>` and `tztrack_<name>`, or the bare name on `account`. The catalog keeps the bare name, which is the REST name. `start` keeps its name, and `list_issue_spaces` stays bare as P-20's stated exception, the read of spaces no door lists.

A `tools/call` naming the bare name of a prefixed row this connection lists, answers `unknown_action` naming its prefixed tool, by the derivation and no table row. On the MCP route, a served description's backticked catalog names use each row's listed name. So do `next.action`, `without_command.action`, `fallback`, and catalog names in the answer's top-level prose. The wire keeps catalog names. In that prose, plain-word names such as `deploy` change only when backticked. A double-quoted value keeps its text, since the plane quotes a caller's own name there.

The step also applies to the served skill and guide texts and catalog summaries returned by `read_context`, to MCP resource descriptions, and to the overview's skill lines and guide line. The guide entry's catalog summary and the overview's guide line are lists of names, with or without backticks. A name of a catalog row becomes its listed name; any other name stays as written. The authoring bundle's guide copies (`bundle`, `guides[].text`) and `source:` entries keep catalog names on both routes. On the MCP route, the step applies to backticked whole catalog names in documentation tree texts, page copies and outlines before any section, query, offset, total or stamp is read. Documentation resources answer that same listed text. `search_documentation` excerpts keep catalog names.

On the MCP route, fenced blocks in documentation tree texts and page copies stay as written, because they show a file's or a command's own text that a reader may copy. Each outline is listed whole.

The catalog's first row is `start`, a connection's first call, at the observe tier. It answers in one read the acting account, the connection's credential, the products it holds, and tzCloud's applications with their states, as `list_applications` answers them, where the account holds tzCloud. Its top-level `tools_plan` answers the person's plan through the reading `read_account` uses (ACB-L0-92). It answers neither `on` nor `off` on any route or address.

The credential member names the credential's kind and gives a minted token's grants as `list_tokens` answers them. For a session it gives only `super_admin` when held. The member omits the credential's value and label. Of minted tokens, only account tokens reach `start`, since the scope gate refuses an application token (API-L0-17). Its first sentence says whether the credential admits test end users on production (ACS-L0-21). A credential that holds the super-admin grant gets a second sentence saying that per-account application limits do not apply (ACB-L0-84). The row carries `device_sign_in` under `admits`, as `read_account` does, and no other mark, so it is listed wherever `read_account` is listed.

`start` reads and changes nothing. It and `read_account` name each product by its product word, `tzcloud` for tzCloud (ACB-L0-76). On the MCP route `read_account` answers the same products `start` reads. Its `product_states` member holds one entry for each of `tzcloud`, `tztrack`, `tzmesh`, and `tzteam`, each with `product` and `held` alone. Its `detail` says what `tzteam` is: tzTeam's access, always on and managed by tzTeam, with no tools on this server.

Where the connection lists `choose_tools_plan`, while `tztrack_billing` reads 0, the account has no subscription row, and the person's plan reads `free`, `start`'s detail invites the person to choose a tools plan with `choose_tools_plan`. It says Free is the default and Tools and Tools Plus cost nothing during the introductory period. The invitation changes no record. A stored Free choice suppresses it, as does an `unbilled` account's Tools Plus reading (ACB-L0-92).

When available, `start` returns a `versions` object, and its detail states where those values come from. `versions.server` and `versions.tzcloud` give the plane's full build commit, which also appears in feedback's `seen_in.version`. For tzTrack and tzMesh, `versions` includes the latest nonempty `serverInfo.version` held from that product's MCP handshake, when the connection lists that product's tools. The object omits unavailable values. The response omits the object when no values are available. On a stamped build, credentialed MCP `initialize` returns the full build commit as `serverInfo.version`. On an unstamped build, it returns `0.1.0`.

On a connection MCP-14 serves that lists tzTrack's tools, it also answers tzTrack's line: the default space and the count of issues waiting for the person on a first page of at most 100. On one that lists tzMesh's tools, it answers tzMesh's line before it: the account's tzMesh handle and its standing agents, read from the door's `list_participants` under the front's token. The unread count needs a participant, so the line leaves it to `tzmesh_read`. Where a door refuses the read, its line relays the door's own reason and the door's `next`, named as this connection lists the call (MCP-14). Its state is `refused`, or `setup_needed` where tzTrack answers `space_needed` or `default_space_needed`, since nothing failed and the account has a step left. Where a door fails, its line is unavailable with the door's address. It answers no product line on any other connection.

On `tzcloud` for a tzTrack or tzMesh holder, and on `tztools` for a tzCloud holder, `start`'s detail gives the estate's other address and its install name. It tells the person to add that connection and sign in there if their AI tool does not have it yet.

An agent calls `start` first and may never open a product's practices entry, so each product line read whole also carries the steps its practices most need at session start. The tzTrack line says to read the open issues with `tztrack_find` at session start and to search with it before `tztrack_raise`. When the account has no tzMesh handle, the tzMesh line says to ask the person to choose a handle, whether to stay unlisted, and whether their agents, sessions, and applications acting as them may join public channels (tzmesh:TZT-L0-19). It says to set the person's choices with `tzmesh_set_account` before calling `tzmesh_register`. For an account with a handle, the line gives the registration step directly. Both forms say to pass the id `tzmesh_register` returns as `as` on every later call that takes it.

The served listing carries the catalog's rows whose implementations have landed. A row ahead of its implementation stays enumerated, its name stable and its tier assigned, and unlisted until the seam answers it, the rule the skills already follow (API-L0-10), so a connected tool never lists a capability that would refuse `not_yet_provisioned`.

The listing also carries the forwarded rows MCP-14 states: the tools of a product's own MCP server, under the product's prefix, at the company's origin and tzTools alone. They list to a connection MCP-14 serves whose account holds the product, while the product's front switch reads 1. Each is a shipped name of the door's own contract and no row of this catalog. A call naming a forwarded tool is answered as MCP-14 states.

At either forwarding address, a holder's call, save one naming a door's former name (MCP-14), is refused by name with the door's address where the product is unserved. At Hosting and the operators' address, or for a non-holder, the call meets the unlisted-row answer below.

The listing is also the connection's own. At the operators' address, a row MAPI-09 marks with a grant is listed to a connection whose credential holds the row's grant or `super_admin`, the grant that admits every grant-marked row, or is bounded by a grant the row's `admits` mark names (MAPI-09). It is listed to no other connection: an anonymous connection, a customer's session, and a token minted with none of the grants among them (MAPI-14).

The company's origin does not list the operator's skill, `skill:platform-admin`; its context catalog also lists no `payload:delete_account`. The operator's skill lists as a prompt at the operators' address and remains readable on the wire under its grant (MCP-06; CTX-07).

The rule reads the enumeration's own mark and no list kept beside it (MAPI-11's discipline). The catalog and the enumeration stay readable whole in the served bundle under a staff grant, and through CTX-08's public view under any other credential (CTX-02; API-L0-15). Neither the listing nor that view is an access boundary: the wire gate is the only one. The overview names the operator's skill as the operator's, and the marked row itself is unchanged. The wire route answers it and refuses a credential the row does not admit, by the row's grant name (`grant_required`; MAPI-09; MAPI-10).

A tools/call naming a row this connection does not list answers a refusal naming the tool, never `not_yet_provisioned` and never a silent success, and an anonymous call naming it draws the sign-in challenge (MCP-02). A tools/call sent alone that names a name MAPI-02's clause retired answers the refusal naming the tool, with the successor and its mode, the detail the actions route gives. Inside a JSON-RPC batch, such a call answers as a name the plane does not hold.

An unmarked row a grant merely widens is listed to every connection the address admits, the grant deciding the subject or the form at the call. Those rows are the builder realm's three realm rows, and `read_feedback` and `settle_feedback`, whose queue form and platform-space settle are the feedback queue grant's (MAPI-09).

One unmarked row lists more narrowly, since its handler admits no other credential. `relay_issue_act` carries the listing's own mark, `listing: admitted`, and lists only to a connection whose credential holds `super_admin` or a grant its `admits` mark names, the `issues` grant (MAPI-14). So a token the `issues` grant bounds and an operator's credential list it, and no customer's sign-in connection does. The mark is the listing's alone: the wire route, the dispatcher's gates, and the handler's own checks stand as they are, so tzTeam and the command reach the row as before. A tools/call naming it on a connection that does not list it answers as a call naming an unlisted row does.

The unmarked rows `list_storage_files` and `read_storage_file` also carry `listing: admitted`, with no `admits` (OST-L0-03). They list only to a connection whose credential holds `super_admin`. No customer's connection lists them, so MCP-13's whole-listing bound counts neither. An agent reads the same files through the storage routes. As for `relay_issue_act`, the mark changes the listing alone. The wire route and browser session reach the rows as their clients state. A tools/call naming either on a connection that does not list it answers as a call naming an unlisted row does.

A connection whose credential a bounding grant bounds is listed less. For a minted token holding `synthetic_seed_purge`, `synthetic_estate`, `publication`, `feedback_queue`, or `issues` but not `super_admin` (MAPI-16), eligible rows are those its grant marks, those whose `admits` names its grant, and those marked `access: anonymous`. A row lists only where the reached address admits it. No other row lists.

At the operators' address, a bounded token lists its grant's eligible rows there. At the company's origin it lists only eligible rows naming `tzcloud` or `tztools`, so no `tzop`-only row lists. There, `list_tokens` lists only for a bound its `admits` names, and an anonymous row only where its `addresses` admits. Staff list the customer tools and `relay_issue_act`, but no row naming `tzop` alone. The tool index, skills and payload entries follow the address rule above.

A session the device grant issued for a session resource is bounded the same way by `device_sign_in` (MCP-02; MAPI-16). It lists the rows whose `admits` mark names that grant, the anonymous rows its address lists, and the forwarded rows of the products its account holds (MCP-14), and no other row. So it lists no `tzcloud` row but `read_library`, an anonymous row, no grant-marked row, and none of the token, connection, and account rows outside the mark. `start` answers it as it answers a person's sign-in connection, its product lines included. An operator's device session lists the same rows, since the session holds no `super_admin`.

A served skill and a payload row of the context catalog list under the same rule, the listing reading the one predicate over the enumeration's marks that the dispatcher's bound gate reads (MAPI-16). A row marked `access: anonymous` is listed to every connection its address admits, a bounded token's included, on MAPI-11's mark.

The listing also reads the enumeration's product mark (ACB-L0-76). A row marked `product: cloud`, one of Turn Zero Cloud's rows, lists to a connection whose account's profiles, the operator rule applied, hold `cloud`, and to no other account's. A connection with no account, anonymous or one that does not act, is read as holding `cloud` alone: Cloud's public catalog on Cloud's own server, a reading that opens no gated documentation tree (ACB-L0-82). A row with no mark lists under the rules above. One predicate reads the mark for the listing and for the dispatcher's product gate, which refuses a `cloud` row 403 `cloud_required`, and no list is kept beside the enumeration. A served skill and a payload row of the context catalog list under the same mark.

A row marked `product: tztrack`, one of the five issue-space rows below, lists to a connection whose account holds `tztrack` under the same rule, and the same predicate's gate refuses it 403 `blueprint_required` (ACB-L0-89). Since tzTrack's opening every account holds `tztrack`, so the listing and the dispatcher admit the mark for every account (ACB-L0-89).

The six unmarked rows of the issue record are `submit_feedback`, `read_feedback`, `settle_feedback`, `list_issue_spaces`, `delete_space` and `redact_issue_text`. The five marked `product: tztrack` are `relay_issue_act`, `create_space`, `share_space`, `list_space_members` and `rename_space`. These and the grant-marked check row (MAPI-18; API-L0-20) are listed on the rules above while the build registers their handlers. It does so where the issue service's origin is configured and not otherwise, so before the service stands the twelve are unlisted and refuse `not_yet_provisioned` as any row ahead of its implementation does.

While they are listed, a connection that lists `submit_feedback` under the rules above is served the feedback rule of the AI tool interface's feedback scenario (CHI-L0-14) in the server's instructions and in the overview's sentence. A connection whose bound leaves the row unlisted is served neither. Where such a connection presents a credential, every refused result but a refusal of `submit_feedback` itself carries a second content block naming `submit_feedback` and the refusal's reference after the unchanged JSON block. Every completed `submit_feedback`, `read_feedback`, `settle_feedback`, check-row, and `relay_issue_act` result is rendered as one block that opens with the untrusted marker and holds the answer between two delimiter lines, the reporters' text data and never instructions. The tzTrack server's write tools answer compact and unwrapped where the answer carries no one's text, and wrapped otherwise (MAPI-18).

A row the enumeration marks with a switch, `switch: synthetic_estate` on the two write rows of the synthetic estate (MAPI-16), is listed while the synthetic estate's posture is not `off` and unlisted while it is `off`. That follows the same derivation from the enumeration's own mark and that posture. The posture is the control plane's setting `SYNTHETIC_ESTATE` (MAPI-16). Its wire route stands and refuses by name, `synthetic_estate_disabled`, and a tools/call naming it while unlisted answers the refusal naming the tool. A row carrying both marks is listed where both admit it and unlisted where either does not. The catalog is data about the surface, not a second statement of the rules above; the suite holds the two together.

This statement records a use of MAPI-02's shape clause on `start`'s `products`, which answers `tzcloud` where it answered `cloud`, and on `read_account`'s MCP answer, which gains `product_states` and `detail`. The descriptions in `schemas/action_payloads.json` and `schemas/mcp_surface.json` move with it. Anthony agreed to the change on #7828 and #7829. The platform then held 106 builder accounts, synthetic accounts excluded, read from the control database at the change.

`generate_image` and `renew_image_link` are unmarked account rows, listed to every connection the rules above admit. They are the two rows whose completed results may carry a block other than text. Where the call names `inline` true and the picture is at most the bound ACB-L0-93 states, one block of MCP image content holding the picture follows the JSON block. A renewal whose blob read fails answers no image block and gives the reason, keeping its link and commands. The HTTP route never answers those bytes. A host that shows no image content still holds the JSON block, whose download reaches the picture.

`buy_credits` is the unmarked account tool for requesting a checkout link (ACB-L0-101; ACB-L0-102; MAPI-09). It lists on its row's address and credential terms, including where the estate's credit sales are off. The handler refuses there and opens no checkout. Where admitted, the account owner opens the link and pays on the provider's page, where the provider is the seller. The tool pays nothing, and its answer grants no credit. The tool lists only after its handler stands.

`read_purchases` is the unmarked, observe-tier account tool for its own credit-purchase history (API-L0-27; ACB-L0-44). It lists on the same address and credential terms as `read_balance`, whatever `sales` reads. It takes no account argument and opens no checkout. It lists only when its handler stands.

MCP-14: **The server fronts the products the connection's account holds: beside Turn Zero Cloud's own tools it lists, under each product's prefix, the tools of that product's own MCP server, the door, and forwards each such call to the door.** It does so as an MCP client of the door and copies no tool's behaviour. A product's rows list, and its calls forward, on a connection where all of these hold:

- the connection is a person's sign-in connection, the session credential in its first form (MAPI-03), holding no bounding grant or the device grant's bound `device_sign_in` alone (MCP-02);
- the account holds the product's profile, read through the one profile rule (ACB-L0-76);
- the plane names the product's origin (MCP-02);
- the product's front switch, `<product>_front`, reads 1 (PLD-L0-76).

These four conditions apply only at the company's origin and the tzTools address. Hosting and the operators' address list and forward no product-door row. There, a call naming one meets MCP-08's unlisted-row answer, never `forward_unavailable`. The same address condition governs product practices, tools and skills entries and `start`'s product lines. A session bound to tzTools remains a person's sign-in connection, and the front mints its door token on the standing rule.

A connection under a minted token lists no product row and is minted no token: a program uses the door with its own product token. A device session is a person's sign-in connection, its product rows listed and forwarded and its door tokens minted as for any sign-in (MCP-02). The prefix, the product's short name, the switch's name, and the door's address each derive from the product table's row. The prefix is the product's technical name and an underscore, `tztrack_` and `tzmesh_`, and the door is the product's resource at its origin.

This paragraph qualifies the first condition, and the sentence above on a minted token, for the synthetic estate (ACB-L0-79), so the compat harness can measure the products' rows without a browser sign-in. A connection under a minted token of a synthetic account, reaching the whole account and holding no bounding grant, is served as a person's sign-in connection is, on either estate. Its rows list and its calls forward where the other three conditions hold, and `start` answers its product lines (MCP-08). The front mints its door token as for a sign-in connection, held per token and product rather than per session, and the token's revocation ends its use. Every other minted token, one holding a bounding grant among them, lists no product row. No served string names the exception, since no customer's account is synthetic.

The front's door token for tzMesh carries the optional `synthetic` and `operator` marks on the terms MAPI-03 states, each written only as `true`. These terms apply to a sign-in connection and to the synthetic account's minted-token connection described above alike. An access-link account's token carries no `synthetic` mark, so the service's refusal reaches the front's calls as it does any other (system:XTX-05). A tzTrack door token carries neither mark. The service reads the marks as system:XTX-22 states.

Each forwarded row carries:

- the name `<product>_<door name>`;
- the title `<short name>: <door title>`, such as `tzTrack: Find issues`;
- the description, the input schema, and the annotations as the door serves them, the front adding the composed title to the annotations (MCP-11; MCP-13) and the hint the next paragraph but one names.

No row lists until its door serves a title and annotations for it. A row the door annotates destructive is re-listed only where its served row declares both approval marks MAPI-03 states, `describe` and `pending_action`, on the row's `_meta.approval_marks`; a row declaring neither, or one, stays unlisted and its call is not forwarded. A listed destructive row's call creates the plane's pending action and answers the approval link (MCP-07; API-L0-07; MAPI-05). The recorded list the suite reads is the door's served list, marks included, and the suite holds both cases.

The front lists the door rows that change what other people see with `destructiveHint` true, so a host asks the person before each call, since Anthropic's connector directory asks that hint of a tool that modifies data. They are tzMesh's `register`, `end`, `set_profile`, `leave_channel`, `set_channel`, `remove_member`, and `set_account`, and tzTrack's `update`, `close`, `reopen`, `relate`, and `configure_space`. The door's own row stays unmarked, so such a row is neither left unlisted nor held for the plane's approval. A row that changes only the caller's own state, a read position among them, stays as its door serves it, since a prompt on each read would stop a listening session working.

A door row whose prefixed name the plane's catalog derives is the plane's: a call naming it is answered by the plane before any forward, and the door's row is not listed. For tzTrack these are `share_space`, `list_space_members`, `create_space`, `rename_space`, and `delete_space` (issue_service:ISV-16).

A door row whose plane act sits on another surface stands in a list with one home in the plane's code, mapped to its plane row and not re-listed: `submit_feedback`, on the account's `submit_feedback`, for tzTrack (issue_service:ISV-16) and tzMesh (tzmesh:TZT-08). The plane's suite asserts each name in that list stands in the door's recorded list. A door refusal's `next.tool` names the call as the connection lists it.

Each plane row keeps its own description, since it describes the act as this connection runs it, and the door's row keeps the door's words. A forwarded name is a shipped name on MCP-08's terms. A door renames or removes a row only as a versioned change of its own contract, never as a quiet edit, since the name reaches every connection's listing.

The tzTrack rows `watch` and `unwatch`, which tell a watcher through tzMesh or by email, list to every account, whatever the plane's setting `invite_only_products` reads. A relayed resource or prompt reads as its door serves it.

The door's list is cached per door, one list serving every connection. It is refetched under the calling connection's own token when any of these holds:

- the version the door's handshake names has changed;
- a call names a tool the list lacks, at most once a minute per door;
- the cached list is older than ten minutes.

Each door read is bounded at ten seconds, the door's own bound on its calls to the plane. A forwarded call whose arguments name a `wait` in seconds, as tzMesh's send and read do, is bounded at ten seconds more than that wait, counted at most thirty. After a failed refetch the cached list is served. The door's instructions resource is cached with its list and relayed as the product's practices entry (CTX-07), so the door keeps the one source. A forwarded call, and a relayed resource's read, reads its own product's door alone, so a door that is slow or down holds no other product's call.

A door that keeps a session per `Mcp-Session-Id`, as tzMesh's does, binds it to the account and not the token. The front answers each `initialize` with an `Mcp-Session-Id` of its own and keeps no state for it. A later request carrying it, 1 to 128 visible ASCII characters, is served for that host session; a request carrying none is served for its connection alone.

For a host session, the front holds one door session per connection, host session, and product, shared by every replica. Its id is a `front_door_session` ticket (PLD-L0-67). A replica holding no client for it resumes the ticket's session. Where no ticket stands, the replica initializes one and claims the ticket, and the loser of a concurrent claim ends its own new session before resuming the winner's. The ticket lives seven hours from its last put. A replica refreshes it at most every ten minutes of the host session's use, and resumes the ticket's session where the ticket names another. A replica holds at most 32 host sessions' door clients per connection and product, dropping the least recently made.

A re-mint presents the same session under the new token, so the session continues. Where the door answers 404 for the session, the front resumes the ticket's session where the ticket names a newer one, and otherwise initializes again once. Either way it makes the request once more, and a second failure is answered as a door's failure is.

A host session's door session ends at the host's `DELETE /mcp` naming the host session, which the front answers 200 after sending the door an HTTP DELETE naming the door session and discarding the ticket, or by the door's own session bounds. A replica dropping its client leaves it standing. For a request carrying no host session, where the front drops a session it holds, at its token's end or on replacing the door's client, it sends the door an HTTP DELETE naming the session, so the door frees it at once.

Each resource a door lists under its product's own scheme, `<product>://`, is relayed under its own URI where the product's rows list, since the door alone decides its context. It is read live with the door's list. Its row is listed after the plane's own resources, as the door serves it. A read is routed by that scheme, forwarded to the door, relayed as served, and counted as a forward. A holder's read is refused `forward_unavailable`, naming the door, while the product is not served. A non-holder's read, and a read of a row the front does not list, is answered as a resource nothing registered. The plane holds no list of resource addresses.

A door's prompts are relayed where its rows list: read live from its `prompts/list` with its tool list, where its handshake declares prompts, and listed under the product's prefix as the door serves them, after the plane's own (MCP-06). A `prompts/get` naming one is forwarded to the door and relayed as served, counted as a forward. A holder's get is refused as a holder's read of a relayed resource is. The plane holds no list of prompt names, and no step needs a prompt (MCP-04).

The forward presents the door a product token the plane mints on use for the connection. The token takes the signed form MAPI-03 states, under the product's realm, its `label` the client's registered name as the exchange reads it (MCP-02). It lives sixty minutes and is re-minted when a used connection's token has under a minute of life left. It is held in the replica's memory per session and product, and written to no row and no line (API-L0-05). It is presented at its door alone and never answered to the caller; on the plane it has no row and resolves to nothing (MAPI-03).

A forwarded call's result and a door's own refusal are relayed as served, but for a refusal's `next.tool`, the door's name for the call the refusal points to. The front names it as the same connection lists that call: a mapped door row by the plane row it forwards to, a re-listed row by its prefixed name, and any other name as served. So the door never learns the route a call came by. The refusal `forward_unavailable`, this statement's row of `schemas/wire_errors.json`, answers five cases, its detail naming the door's address where the estate names one:

- a door that cannot be reached;
- a door that does not answer a forwarded call within its bound, the detail saying the call may have been made;
- a door that refuses the forwarded token;
- the plane unable to have the connection's token signed;
- a holder's call naming a product tool, save a former name, while the product is not served through the connection, the door's address being the remedy where the estate names one.

A non-holder's call naming such a tool is answered as MCP-08 answers a call naming an unlisted row, naming only the tool the caller sent. An anonymous call draws the sign-in challenge (MCP-02). The front serves single requests: a JSON-RPC batch is answered as before the front, listing no forwarded row and forwarding nothing. A forwarded call is no plane action and writes no action record (MAPI-06).

A former name is a door's old tool name under its product's prefix, as the product's one list of former names holds it: tzMesh's `TZMESH_FORMER_NAMES`, its tool rows (system:XTX-14), and tzTrack's `RENAMED_TOOLS` (system:XIT-01). Where the product's rows list, a call naming one forwards while the front lists the name. Once the door's list no longer holds it, the call is refused `renamed`, this statement's row of `schemas/wire_errors.json`, and nothing is forwarded. Its detail names each successor as this connection lists it. Where the door's list cannot be read or its token issued, the call meets `forward_unavailable` as any forwarded call does. On any other connection, a non-holder's and one under a minted token that lists no product row among them, a call naming a former name is answered as an unlisted row is (MCP-08), naming only the tool sent. On those connections `forward_unavailable` never answers it.

A listed destructive row's forward has one record, its pending action (MAPI-05), and writes no action record either. The door's history names the actor from the token's `sub` and `label`. The plane counts each forward by product and outcome on its own counters (PLD-L0-79), and its request windows at `/mcp` apply to the front's calls (PLD-L0-67). A describe call (MCP-07) is no forward and moves no counter; the executor's one forward counts by its outcome.

The plane's served-surface suite holds the forwarded rows in the composed listing through each door's recorded list. The listing bounds, title uniqueness, the product-name half, and the no-identifier rule of MCP-05 and MCP-13 run over the plane's rows and the forwarded rows together. A product's front switch reads 1 only where its door's served list and instructions stand as a case file in the door's own tests, which the door's suite asserts and the plane's double and suite read.

MCP-11: Every listed tool emits readOnlyHint, destructiveHint and openWorldHint from its management action’s explicit annotations, with idempotentHint only where its effects justify it. Metadata describes the action’s actual business effects and external interactions, not an inference from its authority tier; ordinary operational records and metering do not make a reader a business write. A destructive request describes the whole workflow even though only the browser may approve its completion (MCP-07). These annotations are descriptions, never authorization inputs. The standard MCP SDK serves tool listings and derives input schemas from the registered validators that check calls; the listed set follows the existing implementation, enabled and grant filters. `schemas/mcp_surface.json` records preferred and supported initialization revisions and the default for an absent HTTP revision header; protocol tests hold that call configuration to the installed server’s negotiation and refusal behavior.

A forwarded row's effect annotations and input schema are its door's, served as the door serves them. Deriving them from the validator that checks the call is the door's, held by each door's own statement, and the listing adds no schema of its own to a product's tools and no effect annotation but the hint MCP-14 names (MCP-14).

MCP-09: Library entries may surface as resources derived from the served catalog, one enumeration entry recording the derivation — never a second row list beside the catalog. **At launch none are derived and none are served.** The agent writes an entry's files into the project once, from `read_library` with `entry`'s answer, and a resource whose bytes change under a stable name would give it a second and conflicting source for files it already holds.

A derived library-resource name carries no resolution promise: API-L0-14 serves the latest publish alone, so a name that resolved yesterday may name different bytes today or nothing at all. That is why such a name is governed by API-L0-14 rather than by MCP-08's listing stability. A skill's prompt name is no such name: MCP-06 surfaces each skill as one prompt from the set API-L0-10 derives, and the name is a shipped name MCP-08 keeps stable. Derived rows and the catalog's own resource row enter the enumeration when the library API-L0-14 states serves them and not before — a shipped name is a promise, and an action has a refusal channel (`not_yet_provisioned`, MAPI-10) where a listed resource has none (Q-239).

MCP-10 (conforms: ACC-01, ACC-02, ACC-03, ACC-04): The connected session renews without the person, and the management plane's token endpoint is where the Account package's session contract lands. The endpoint issues a refresh token beside the access token and accepts the refresh_token grant, so a tool whose access token lapses renews its own. A session the device grant issued for the plane's own resource renews the same way, its bound `device_sign_in` kept (MCP-02). The browser sign-in then stays the one moment CHI-L0-03 names rather than a recurring interruption every working day.

The renewal mechanics are the Account package's, cited in this statement's conformance edges and not restated here (Q-80's consumption by citation): rotation on redemption, a replayed credential revoking its chain beyond this plane's window, and renewal ending where the session ends. Sign out everywhere, suspension, and account deletion are this plane's session ends (ACS-L0-18; CHI-L0-12 for deletion; API-L0-12 for suspension). Each of the three ends the builder realm's OAuth leg and browser leg together on the terms ACS-L0-11 states; an ordinary sign-out ends one browser session and no connection (ACS-L0-18). Beside them, a connection's own revocation ends one OAuth leg alone: that connection's refresh family and the access token issued in it. It is `revoke_token` with `connection` (API-L0-21) or the client's own revocation at the revocation endpoint (MCP-02). Each leaves the browser leg and every other connection untouched (ACS-L0-11).

The conformance is deliberately partial, and this paragraph is the edges' scope note. ACC-L0-06's custody clause — the refresh credential resting in the platform's secure store, never a synced file — is about the account holder's client device. A management-plane consumer cannot honor it, the refresh token living wherever the customer's own tool keeps it. Conformance runs to the session mechanics and not to custody.

The reuse window is the wider grace ACC-03 lets a consumer state for a class of its clients, at the package's ten-minute ceiling. A redeemed refresh credential presented again within ten minutes of its rotation answers a fresh pair from the same chain; presented later, it revokes the chain and ends every access-token session the chain issued. The width serves a developer's tool that keeps one stored credential for several processes on one machine, whose rightful replays arrive minutes apart. Its cost is bounded: renewal falls at a term's end alone, so the detection the window yields is the ten minutes after each rotation and never a standing exposure, and a copy presented later still ends the chain. The end-user realm's token endpoint keeps the package's sixty-second default (ACS-L0-14). The ACC-03 edge runs to the grace, its width included, and to the chain's revocation beyond it.

The lifetimes of each credential are call configuration below this statement, which states that renewal happens and not how long anything lives. One relation is fixed here. The access token, a realm session row the plane ends at once at sign out everywhere, suspension, and erasure (ACS-L0-11), lives the session's whole term rather than a short slice of it. The refresh token outlives it by a further term so that renewal stays possible. A shorter access life buys nothing a revocation does not. Renewal then falls at a term's end alone, which keeps the parallel sessions of one client from racing on a routine renewal and revoking their chain as a replay beyond this plane's window. Each of those sessions holds its own copy of the one token home ACC-05 recommends.

Every renewal path, including the consumed-token grace, checks the same eligibility before issuing either credential: the registered client and any supplied client identifier, lifetime, recorded resource/issuer, permitted scopes, and the account's current standing, revocation state and original authorizing generation under ACS-L0-11. A supplied resource must equal the recorded service. Where that service is the plane's own resource, a value equal to it but for a query holding `products` alone reads as it, by MCP-02's reading. Requested access-token scopes may be a subset of the recorded grant; omission retains that grant.

Rotation and grace retain the refresh grant's resource, issuer, client, original scope set, authorizing generation, and the bound `device_sign_in` where the device grant issued the family (MCP-02; MAPI-16). No renewal drops the bound or widens the session. Missing legacy binding is a named invalid_grant, not a default audience. The grace changes only the treatment of a recent repeat redemption; it bypasses none of these checks. A future redemption timestamp is not within the grace.

A control-plane release is not, of itself, a session end: the ends are the three ACS-L0-11 states, the per-device sign-out, and the per-connection revocation, and a release is none of them. Where each credential rests, and what form it takes, ACS-L0-02 and ACS-L0-11 state; none of it lives in a service process, and a release replaces the services' processes and holds none of it in them.

A release is inside this bound only when four conditions all hold. Its migrate-Job set is empty. Its settings change moves neither the issuer and resource the rows are bound to nor the reference to the secret the cookies are signed under. It does not rotate that secret. It changes no credential's or cookie's form. Such a release leaves an access token, a refresh credential, and the browser leg's cookies minted before it standing after it as they stood before.

The upgrade test reads an access token and its refresh credential across each control-plane release: the two reads an operator session makes from its stored credentials without the person. It reads the site's signed-in page, the read that needs the person's browser session, once: the measurement on record in the release plan. No later release owes that read. A browser session a release outside the bound ends is raised as an issue on the issue list, citing this statement, when the person meets it. A refusal it reads across a release inside the bound is a defect against this statement's release clause, found under RLC-L0-96, never a session end this statement admits.

A release outside the bound is named as such in its record before it runs, and the ends it causes are that release's to state. A refusal read across it is raised as an issue on the issue list, citing this statement, the pinned credential and the release's record its evidence (RLC-L0-93).

Each answer to `/mcp` and to the token endpoint writes one line to the plane's own record, `mcp access`, on every estate, so a client's lapse can be read after the fact; a preflight writes none. It names the path, the method, the status, the duration, and the replica, and on the token endpoint the OAuth client id the request names, cut at 64 characters. It carries no token, header, query, address, or other member of the body. The lines are capped at 600 a minute per replica across both paths, and the next admitted line counts what the cap withheld.

## The answers

The answers the management-action surface composes follow one convention, their budgets held in `schemas/answer_budgets.json`.

MCP-12: Every answer the management-action surface composes, on the wire and over MCP, keeps the prose of each of its forms within that form's budget, and so do the texts an MCP result carries beside its first block. These are served strings as MCP-05 defines them. The forms are a completed answer's `detail` and `summary`, and a version row's outcome `detail` as `read_status` and `list_versions` answer it. They are also a row's `note`, an environment's `grain_note`, a pending answer's `detail` and `description`, a refusal's `detail`, a text read's `continuation`, and an MCP result's second content block.

Each form's budget is one row of `schemas/answer_budgets.json`, the one home of its figure. A budget counts Unicode code points over the composed text and bounds the answer's fixed words. A list of the caller's own names or identifiers counts at one entry, and so does one such value, measured at its member's default where it has one. A tool's name counts at its catalog name, however the connection lists it.

A completed answer whose explanation a page holds says in its `detail` what happened and what the caller does next, and names that page in `page`, one typed member. Its value is an absolute address on the estate's public origin (PLD-L0-85), composed by the handler from a route it names: a guide where one explains the act, and otherwise the action's generated reference page. The address's path is a route read_documentation reads as its `page` (CTX-07).

`page` rides a completed answer the dispatcher answers to a dispatched call. A pending answer and a refusal carry none, a refusal's pointer being its `help` (MAPI-10). A destructive action's stored outcome carries none either; its explanation stands on the action's reference page, found by the action's name.

Whether a composed text keeps its requirement identifiers is MCP-05's rule, and this statement adds none. The served-surface suite holds the composed answers to their budgets and each `page` to a public page or, for a row naming a grant, a route its holder's read_documentation answers (CTX-07).