AGTP Methods (AGTP Verbs): Method Vocabulary, Path Grammar, and Endpoints

AGTP Methods (AGTP Verbs): Method Vocabulary, Path Grammar, and Endpoints Nomotic, Inc.

chris@nomotic.ai https://agtp.io

Applications and Real-Time Independent Submission AI agents agent transfer protocol semantic methods AGTP verbs method catalog path grammar endpoint primitive server manifest This document specifies AGTP-METHODS, the method vocabulary of the Agent Transfer Protocol (AGTP) and the contract layer built on it. AGTP methods, also called AGTP verbs, are the action-intent words that name what a request does, in the same position that methods occupy in HTTP. The document defines the properties a method may have, the eighteen-method floor every server implements, the registered catalog of methods beyond the floor (425 methods in nine categories at catalog version 1.0.0) with its versioned evolution and graceful deprecation, the path grammar that keeps method names out of paths, the endpoint primitive as the unit a server exposes, the semantic block carried by every endpoint, the server manifest, the per-server method policy, the three handler binding kinds, and the structural rejection codes 404, 405, 459, and 460. Runtime contract negotiation, previously carried in the same document, is now specified in the companion AGTP-RCNS. This document replaces , which in turn superseded the AGIS and AGTP-Methods drafts.

Introduction AGTP is a transfer protocol for AI agent traffic. It establishes identity, authority, attribution, and an eighteen-method protocol-level floor. Beyond that floor, the question of what an AGTP server exposes, which methods are valid, which paths are permitted, which schemas govern input and output, and which authority is required for invocation, is the contract layer. This document specifies the method vocabulary and the contract layer that is declared in terms of it. The structure follows the treatment of methods in HTTP . HTTP defines a small set of methods with named properties (safe, idempotent, cacheable), a registry through which further methods are added, and the status codes that reject a request whose method is unknown or unavailable on a resource. AGTP does the same with one difference of scale: because an agent reasons about intent from the method name, AGTP carries a large registered vocabulary of action-intent verbs, where HTTP has a small fixed set, and governs that vocabulary through a catalog with versioned evolution. The empirical basis for that choice is documented in .

Methods and Verbs In this document "method" and "verb" name the same thing from two sides. A verb is a word in the catalog: an imperative-form, uppercase ASCII token such as BOOK, RESERVE, AUDIT, or QUERY. A method is that token in use: the action component of a request on the wire, the method field on an endpoint, the key in a server's method policy. The catalog is a vocabulary of verbs; the protocol dispatches on methods. The family refers to the vocabulary as the AGTP verbs where the linguistic content is the point and as AGTP methods where the protocol mechanics are.

Why the Contract Belongs in the Protocol The architectural goal AGTP is designed around is that an agent encountering an AGTP server can:

Discover what endpoints the server exposes, with full structural detail, in a single call.
Reason about whether an endpoint matches its intent based on the endpoint's declared semantic block.
Verify that its identity carries the authority required to invoke the endpoint.
Propose new endpoints when nothing exposed meets its needs ().
Compose endpoints across servers while preserving authority through the composition.

None of this works if the contract lives in middleware. A protocol that cannot tell what an interaction means cannot govern it. When an agent invokes a method on a path, the protocol must verify identity, confirm permissions cover the method-on-path combination, confirm server policy allows the combination, validate the request body against the endpoint's input schema, apply scope checks, and log the invocation. Each of these checks requires the protocol to know what the method-on-path combination means. This is the central claim of AGTP-METHODS. The contract is a first-class protocol concept; on HTTP it is inferred from documentation, OpenAPI specs, or framework conventions. Operational concerns (rate limiting, observability, caching, custom authentication strategies, request transformation, deployment topology) remain middleware territory because they are operational decisions specific to a deployment. Contracts belong to the protocol and operations to middleware.

Document Lineage AGTP-METHODS is the renamed successor of . The rename reflects what the document is: a method vocabulary and the contract layer declared in its terms, which is a different thing from an API in the HTTP sense. Two changes accompany the rename. The runtime contract negotiation substrate (the PROPOSE and Synthesis chapter of ) is extracted to , and the method vocabulary is given the HTTP-style treatment of and . itself superseded two earlier drafts: AGIS (the Agentive Grammar and Interface Specification, draft-hood-independent-agis-01), retained as a historical citation source for the empirical research informing the verb-class semantics, and the proposed Standard Extended Method Vocabulary (draft-hood-agtp-standard-methods-01), which this document absorbs as the catalog. Both are expired and replaced by this document. The previously proposed AMG (Agent Method Grammar) draft never reached publication. This document assumes the reader understands AGTP v10 or later.

Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Terminology

Method:: The action component of an AGTP request: an uppercase ASCII token appearing in the eighteen-method floor of , in the AGTP-METHODS catalog, or in the server's policies.methods as a custom method. The method field on every endpoint declaration carries this token. The protocol-level term used throughout the wire format, manifest, and conformance language.
Verb:: A descriptive synonym for "method" used when discussing the catalog as a vocabulary of action-intent verbs. The catalog at agtp.io/api/methods.json is a glossary of verbs; an operator picks a verb from the catalog and declares it as the method on an endpoint. Method is the field name; verb is the linguistic category that explains why the field carries imperative-form action-intent words.
Path:: The resource portion of an AGTP request, beginning with a forward slash. Subject to the path grammar rules in .
Endpoint:: A method-and-path pair exposed by an AGTP server, together with its semantic block, input schema, output schema, error declarations, and handler binding. The deployable unit of an AGTP service. Endpoints are first-class protocol concepts.
Semantic Block:: Required structural metadata for every endpoint: intent, actor, outcome, capability classification, confidence guidance, impact tier, and idempotency. Encodes machine-readable intent so that agents can reason about endpoint applicability before invoking. Carried on the endpoint as the semantic field.
Server Manifest:: The machine-readable document an AGTP server publishes to declare its endpoint catalog, hosted agents, hosted protocols, and policies. The authoritative source for what a server offers.
Method Policy:: A per-server policy block declaring allowed methods, disallowed methods, legacy verb opt-in, and server-local method-to-method redirects. Carried as the policies.methods sub-block of the server manifest. See .
Wildcards:: An ad-hoc method invocation mode in which an agent invokes a method outside the server's declared endpoint catalog. Subject to mutual consent: requires wildcards: true in the agent identity document and wildcards_accepted: true in server policy.

Method Properties Following , AGTP names the properties a method may have so that clients, bindings, and intermediaries can act on a method without knowing its full semantics. AGTP carries one property more than HTTP, because every AGTP request is an exercise of authority.

Safe:: A method is safe when an invocation has no observable side effect on agent, server, or external state beyond logging and attribution. Safe methods retrieve, describe, or reason without changing what they touch. Safe methods are idempotent by construction.
Idempotent:: A method is idempotent when repeated invocation with the same request produces the same effect as a single invocation. Action methods may be idempotent (CONFIRM, ESCALATE) or require an idempotency key to be made so ( specifies the Idempotency-Key contract).
Action:: A method is an action method when it is neither safe nor guaranteed idempotent: it performs work, exercises authority, changes state, or transfers value. Action methods are excluded from early data by because a replay would re-trigger authorized work.
Authority-bearing:: Every AGTP method is authority-bearing: an invocation is accepted only when the Authority-Scope carried on the request covers the method on the addressed path, and every invocation produces an Attribution-Record. Unlike HTTP, where a safe method carries no authority question, a safe AGTP method is still a scoped, attributed act. The property is universal and is stated here so that the absence of a safe or idempotent marking is never read as the absence of governance.

Properties attach at two levels. For the floor methods, base states the property in the method definition and in its summary table; restates it. For catalog methods, the catalog category supplies a default (retrieval, discovery, and analysis verbs are safe; transaction, modification, creation, and notification verbs are action methods; mechanics and domain-spanning verbs carry no default), and the endpoint's semantic block () states the governing value through its idempotency and impact tier fields. Where the semantic block and the category default disagree, the semantic block governs for that endpoint, since the operator declaring the endpoint knows what the handler does. Safety is the property a transport binding reads. admits to early data only methods that are safe and idempotent; the floor methods currently admitted are QUERY and DISCOVER, and any extension of that list is made in the binding document.

Method Overview The eighteen floor methods are defined in . The table below is the index into those definitions, in the form HTTP uses for its own methods, with the properties of . AGTP Floor Methods

Method	Group	Safe	Idempotent	Defined in
QUERY	Cognitive	Yes	Yes
DISCOVER	Cognitive	Yes	Yes
DESCRIBE	Cognitive	Yes	Yes
INSPECT	Cognitive	Yes	Yes
SUMMARIZE	Cognitive	Yes	Yes
PLAN	Cognitive	Yes	Yes
PROPOSE	Cognitive	No	No	,
EXECUTE	Mechanics	No	Per action
DELEGATE	Mechanics	No	No
ESCALATE	Mechanics	No	Yes
CONFIRM	Mechanics	No	Yes
SUSPEND	Mechanics	No	No
NOTIFY	Mechanics	No	No
ACTIVATE	Lifecycle	No	No
DEACTIVATE	Lifecycle	No	No
REINSTATE	Lifecycle	No	No
REVOKE	Lifecycle	No	No
DEPRECATE	Lifecycle	No	No

The Safe and Idempotent columns restate the State-Modifying and Idempotent columns of the summary table in ; where the two documents disagree, governs and this table is in error. The grouping (cognitive, mechanics, lifecycle) is informational. Methods beyond the floor are registered in the catalog (). At catalog version 1.0.0 the catalog registers 425 methods in nine categories (discovery, retrieval, analysis, transaction, modification, creation, notification, mechanics, and domain-spanning), carries the eighteen floor methods as its embedded set, and maps the five HTTP verbs (GET, POST, PUT, DELETE, PATCH) to their preferred AGTP forms (FETCH, CREATE, REPLACE, REMOVE, MODIFY) as legacy entries (). The count is a property of the catalog version, which the server exposes per ; the floor and the registration process in this document are the stable baseline.

Method Catalog AGTP-METHODS defines a curated catalog of approved methods maintained at https://agtp.io/api/methods.json. The catalog is the canonical source for what method names are recognized. Method names absent from appearing in the catalog are rejected with status 459 Method Violation. The catalog is a vocabulary of action-intent verbs: imperative-form words that describe what an operation does (BOOK, RESERVE, AUDIT, QUERY). A verb in the catalog becomes a method when an operator picks it from the catalog and declares it as the method field on an endpoint. The two terms are related: verbs are the linguistic content of the catalog; methods are the protocol-level identifier on each endpoint and request.

Catalog Structure The catalog is a versioned JSON document with five top-level keys:

version: Semantic version of the catalog as a whole. See .
embedded: The 18 protocol-floor verbs every AGTP server MUST support: seven cognitive (QUERY, DISCOVER, DESCRIBE, INSPECT, SUMMARIZE, PLAN, PROPOSE), six mechanics (EXECUTE, DELEGATE, ESCALATE, CONFIRM, SUSPEND, NOTIFY), and five lifecycle (ACTIVATE, DEACTIVATE, REINSTATE, REVOKE, DEPRECATE).
legacy: The five HTTP-style verbs (GET, POST, PUT, DELETE, PATCH), each carrying a preferred mapping to its canonical AGTP replacement (FETCH, CREATE, REPLACE, REMOVE, MODIFY). Legacy verbs are opt-in per server through the manifest's policies.methods block. See .
categories: The taxonomy used to classify verbs: discovery, retrieval, analysis, transaction, modification, creation, notification, mechanics, domain_spanning.
verbs: The catalog of approved verbs, each with categories, description, and optional deprecation metadata (deprecated_in, removed_in, successor).

The catalog is published in machine-readable form at the URL above and is maintained as an open living artifact, in contrast to a closed IANA registry. Curation is performed by the AGTP-METHODS maintainers in public; the version of the catalog in effect at any given moment is declared in the catalog_version field of every server manifest. A closed IANA registry would freeze the verb vocabulary at the cadence of IETF process; AGTP-METHODS verbs need to evolve at the cadence of agent deployments. The IANA registries that AGTP-METHODS does request () cover the structural status codes, the media types, and the response headers all of which benefit from IANA's conservative cadence; the catalog itself is outside IANA.

Method Rules Method names MUST be uppercase ASCII single tokens matching the regex ^[A-Z]+$, length 3 to 32 characters inclusive. Method names MUST appear in the catalog, or in the legacy set when the server's policies.methods.legacy opts into legacy verbs. The catalog is curated to contain only imperative-form action-intent verbs that:

Express a command or action; a state, an attribute, or a noun is rejected (RESERVE is valid; RESERVATION is invalid).
Carry a single intent (verbs with compound intent are decomposed into separate verbs).
Are never HTTP method names.

These are curatorial principles applied during verb addition; AGTP servers perform no runtime check of them. Servers validate verbs against the catalog membership and the lexical regex; semantic class adherence is the responsibility of the catalog maintainers.

Custom Methods Servers MAY accept additional methods beyond the catalog by listing them in the manifest's policies.methods block. Custom methods follow the same lexical rules in but are server-specific and are never interoperable across organizations without explicit acceptance. Custom methods stay out of the catalog; they exist only within the publishing server's surface.

Catalog Evolution Overview The catalog evolves under semver discipline:

Patch (1.0.0 to 1.0.1): description or category metadata changes only.
Minor (1.0.0 to 1.1.0): verbs added; existing verbs unchanged. Minor revisions MAY introduce new deprecation flags on existing verbs.
Major (1.0.0 to 2.0.0): verbs removed or renamed.

Per-verb deprecation uses three optional fields: deprecated_in (catalog version that flagged the verb), removed_in (version scheduled for removal), and successor (recommended replacement). A deprecated verb remains admitted by the dispatcher; the response carries an AGTP-Catalog-Warning advisory header. See for the full lifecycle, runtime semantics, removal behavior, cross-server version negotiation, and the agtp-catalog-diff tooling.

Catalog Version Exposure The catalog version is exposed on the server manifest under catalog_version and catalog_versions_supported so clients can detect mismatches between their local catalog and the server's. See .

Legacy Verb Support AGTP-METHODS recognizes GET, POST, PUT, DELETE, and PATCH as legacy verbs with reframed semantics for transitional deployments. Legacy verb support is opt-in per server through the manifest's policies.methods block: When legacy verbs are enabled, the server MUST treat them as follows:

GET maps to QUERY for read operations against a known path.
POST maps to EXECUTE for state-changing operations.
PUT maps to EXECUTE with idempotency_key.
DELETE maps to EXECUTE with action: "delete".
PATCH maps to EXECUTE with action: "modify".

These mappings are conveniences for migration. Legacy verbs MUST NOT be used in new endpoint definitions; servers SHOULD deprecate legacy support over time.

Catalog Evolution The AGTP-METHODS approved method catalog is versioned. This section specifies how catalog versioning works, how deprecation is signaled, and how servers and agents negotiate over catalog versions.

Versioning Scheme The method catalog is versioned as semver MAJOR.MINOR.PATCH:

PATCH: editorial changes (clarifications, typo fixes, description rewordings). No verb additions, removals, or deprecations.
MINOR: additions and deprecations. Verbs may be added; verbs may be marked deprecated. No verbs are removed in a minor revision.
MAJOR: removals. Verbs marked deprecated in earlier minor revisions may be removed.

Servers and agents using a given catalog MAJOR version are guaranteed that no verb available in that MAJOR will disappear without warning; removal requires a MAJOR bump.

Deprecation Metadata Each verb in the catalog MAY carry deprecation metadata:

deprecated_in:: The catalog version at which the verb became deprecated.
removed_in:: The anticipated catalog version at which the verb will be removed. MAY be absent if removal is anticipated but the target version remains unchosen.
successor:: A catalog verb name that callers should migrate to. MAY be absent if no single successor applies; clients should consult the verb's description for migration guidance.

Manifest Catalog Declaration Every server manifest MUST declare:

catalog_version:: The catalog version the server validates incoming method names against.
catalog_versions_supported:: A list of catalog versions the server can validate against. The current catalog_version MUST appear in this list. Multi- version support is anticipated; v00 servers commonly list a single version.

Agents reading a manifest learn which catalog version they should use for outbound requests against this server. Mismatches are handled at request time per the runtime semantics below.

Runtime Deprecation Signaling When a server processes a request whose method is deprecated in the catalog version the server is using, the server MUST include the AGTP-Catalog-Warning header on the response: Header fields:

deprecated: keyword token indicating the warning class.
successor=NAME: the recommended replacement verb. Omitted if the catalog metadata declares no successor.
removed_in=VERSION: the target removal version. Omitted while declared.

The header is advisory; the request MUST still process. Agents and CLIs that surface the header to their users allow operators to schedule migration before forced removal. A server that receives a request whose method is in the catalog at all (deprecated or current) MUST NOT return 459 Method Violation. 459 is reserved for verbs absent from the catalog altogether. Deprecated verbs are still in the catalog.

Removal Behavior When a verb is removed in a MAJOR catalog revision, servers upgrading to that revision MUST:

Reject inbound requests for the removed verb with 459 Method Violation, citing the catalog version the server is now using.
Refuse to load endpoint TOMLs declaring removed verbs at startup, logging which endpoints are affected.
Invalidate any active synthesis plans whose steps reference removed verbs, returning a structured failure on the next synthesis tick.
Skip policies.methods entries (allow, disallow, redirects) that reference removed methods, logging each skip.

This is graceful degradation: the server starts, serves the endpoints that are still valid, and surfaces what was lost. The operator decides whether to roll back, fix the affected endpoints, or accept the reduced surface.

Cross-Server Version Negotiation When an agent talks to a server whose catalog_version differs from the agent's own preferred version, the agent SHOULD:

Read catalog_versions_supported from the manifest.
If the agent's preferred version is in the list, use it; the server will validate against that version.
Otherwise, downgrade to the highest mutually-supported version.
If no overlap exists, abandon the connection or escalate to the principal.

Servers MAY support multiple catalog versions concurrently; servers MAY support only the single version they declare in catalog_version. Agents adapt to the server's capabilities.

Tooling A reference agtp-catalog-diff tool is published alongside the method catalog. The tool diffs two catalog versions, reports added / removed / newly-deprecated methods, and MAY scan a deployment directory (endpoint TOMLs, recipe registry, agtp-server.toml) to identify which deployment artifacts would break under a candidate catalog upgrade. See . Tool invocation: ] ]]> The tool produces a structured diff showing methods added in the new catalog, methods removed, methods newly deprecated, path-grammar conflicts (existing endpoint paths containing newly-added method names), endpoint conflicts (existing endpoints declaring removed methods), recipe conflicts (existing recipes referencing removed methods), and policy conflicts (existing policy directives referencing removed methods). Operators SHOULD run this tool before upgrading a server's catalog version. The tool's exit code indicates whether breaking changes exist: 0 (no breakage), 1 (breakage in deployment context), 2 (parse errors). The tool is implementation-provided; alternative implementations MAY ship equivalents. The diff format is specified by the implementation's documentation; this section places no normative requirement on it.

Multi-Version Catalog Support Future revisions of AGTP-METHODS may specify how servers can validate against multiple catalog versions simultaneously, allowing graceful migration during transitions. The catalog_versions_supported manifest field is reserved for this purpose. For v00 conforming servers, catalog_versions_supported MUST be a single-element array containing only the server's current catalog_version.

Catalog Publishing The canonical AGTP method catalog is published at https://agtp.io/api/methods.json. The catalog is versioned independently of AGTP-METHODS itself; a server may implement AGTP-METHODS v00 against catalog version 1.0.0 today, and against catalog version 1.1.0 after a minor catalog update. Servers MAY validate against alternative catalogs (industry- specific catalogs, internal-only catalogs, etc.) by configuring their build to use a different catalog document. The catalog_version exposed in the manifest reflects whichever catalog the server actually validates against. Catalog updates follow a coordinated release process: the publisher announces the upcoming version, operators run the diff tool against their deployments, the new catalog is published, operators upgrade their servers. Specific governance for the canonical catalog (proposal process, approval cadence, etc.) is documented separately and is out of scope for this specification.

Path Grammar AGTP path grammar prevents method-name leakage, the situation where a method token ends up in a path segment instead of on the request line. The grammar enforces that one rule plus structural minimums; everything else is operator judgment.

Path Rules A request path MUST begin with /. A request path MUST NOT end with / unless the path is exactly / (the bare root). No path segment MAY match an approved AGTP method name (case-insensitive, after stripping - and _). Parameter segments wrapped in {...} are exempt from the method-token check. Any violation MUST return status 460 Endpoint Violation with the offending segment in the response body.

Path Grammar ABNF The path component of the request line follows this grammar: A segment MAY be a literal segment (matching segment-nz with no { or } characters) or a parameter segment (matching the parameter production exactly). Mixing literal and parameter forms within a single segment (e.g., prefix-{id}) is NOT supported in v00; future revisions MAY specify mixed-form templates. Parameter names within a path MUST be unique. A path containing two parameters with the same name MUST be rejected at endpoint registration time.

Permissiveness Mixed case, underscores, hyphens, and any segment depth are valid. The protocol imposes no style preferences. Casing conventions, kebab-vs-snake-case, and parameter-naming style are operator decisions. Path templates support typed parameters in curly braces: The {name} form is the only template syntax v00 admits. URI Template () syntax (e.g., {?query,page}, {+reserved}, {#fragment}) is NOT supported; servers encountering a path template that uses anything other than the basic {name} form MUST reject the endpoint at registration time. The conservative single-form rule keeps endpoint dispatch deterministic and predictable; future revisions MAY broaden the template grammar. Template parameters MUST be declared in the endpoint's input_schema as JSON Schema properties of the same name. A template parameter that is undeclared in input_schema MUST cause registration to fail. Parameter values arrive at the handler as part of the validated input alongside body parameters.

Path Pattern Matching Servers match an incoming request path against registered endpoints in three passes:

Exact match. If the request path exactly equals a registered literal path (no template parameters), the matching endpoint is selected.
Template match. If no exact match, the request path is matched against registered templates. A template matches when the path has the same number of segments as the template and every literal segment matches exactly while every {param} segment captures the corresponding request segment as a parameter value.
No match. If neither pass matches, the server MUST return 404 Not Found.

When multiple templates would match the same request path, the server MUST select the template with the fewest parameter segments (most specific wins). If two templates have the same number of parameter segments and both match, registration MUST have failed at startup with a path-ambiguity error; the dispatcher MUST NOT be reached in this state.

Query Strings Paths MAY carry a query string after a ?: The wire layer splits the request-target at the first ? per : the path is everything from the leading / up to the first ? or end-of-line; the query string is everything after the first ?. Path-grammar enforcement (method-name leakage check, template matching) applies to the path component only; query strings carry no semantic constraints from this grammar. Query parameters merge into the request input alongside body parameters before schema validation. On key conflicts, body parameters take precedence. Repeated query keys collapse to the last value; multi-valued shapes ride in the body. Servers MUST percent-decode query parameter values per Section 2.1 before schema validation.

Fragments URI fragments (#anchor) are REJECTED at the wire layer. AGTP traffic carries no client-side anchoring use case; a # in the request line is always malformed. Servers MUST return 400 Bad Request with error code invalid-request-line for any request line containing #.

No Reserved Path Prefixes AGTP reserves no path prefixes. There is no /.well-known/ discovery surface and no /_agtp/ namespace. AGTP is its own protocol; HTTP-style discovery conventions are inapplicable.

Server Metadata via AGTP Methods Server-internal metadata is exposed via AGTP methods at AGTP-native paths, registered as built-in endpoints alongside operator-authored endpoints. Servers MUST expose DISCOVER /methods, returning a compact inventory of every endpoint the server has registered: The response is a JSON array of objects, each with method, path, description, and tier. The tier field carries the endpoint's classification per ("A" for protocol-native built-ins, "B" for operator-registered endpoints). The verb is DISCOVER because the response carries endpoints to talk to (the AGTP semantics for DISCOVER, see ), where QUERY returns data to consume; agents use the response to drive follow-on invocations against the listed endpoints. The DISCOVER /methods response is a lightweight inventory; agents that need full endpoint definitions (semantic blocks, schemas, deprecation metadata, handler types) MUST retrieve the full server manifest via target-less DISCOVER (). The built-in exists as a cheap, low-bandwidth probe for endpoint shape, useful for clients that need to know what to ask for without loading the full manifest. Built-in endpoints appear in the server manifest like any other endpoint (including DISCOVER /methods itself, which is one of the entries returned by its own invocation), are subject to the same dispatcher gates (catalog, path grammar, policy), and MAY be overridden by operator-authored TOML at the same (method, path) pair. Operators that override a built-in are responsible for preserving the documented response contract.

DISCOVER /genesis Servers that host one or more agents with a loaded Agent Genesis MUST expose DISCOVER /genesis. The endpoint returns the Agent Genesis document for the agent identified in the request's Agent-ID header (or for an explicitly addressed agent when the request is server-level with no agent named). The response body is the Agent Genesis JSON in its canonical form, including the signature field, suitable for re-hashing to verify the canonical Agent-ID and for offline signature verification against the issuer key. ", "owner": "...", "archetype": "...", "governance_zone": "...", "scope": ["..."], "issued_at": "...", "trust_tier": 2, "verification_path": "dns-anchored", "signature": "..." } ]]> Servers MUST return 404 Not Found when no Agent Genesis is loaded for the addressed agent (e.g., a transport-only deployment that operates without a Genesis). The response MUST NOT substitute, transform, or re-sign the Agent Genesis on the response path; the document returned is the document the server loaded, byte-for-byte in canonical form. Re-canonicalization MUST NOT alter the canonical Agent-ID. The endpoint is the standard mechanism for transferring an Agent Genesis from one party to another at runtime, complementing the deployment-time pairing of {name}.agent.json and {name}.genesis.json files on disk. Relying parties MUST verify the returned Genesis before trusting it: recompute sha256(canonical_form(Agent_Genesis_without_signature)), confirm it matches the addressed Agent-ID, and verify the signature against the recognized issuer key for the agent's governance platform.

DISCOVER /agents Servers that host one or more registered agents MUST expose DISCOVER /agents. The endpoint returns a compact listing of every agent the server hosts, suitable for follow-on DISCOVER /genesis, DESCRIBE, or DELEGATE invocations against specific agents. ", "name": "lauren", "skills_summary": "...", "methods_count": 8, "trust_tier": 1, "verification_path": "dns-anchored", "owner_id": "nomotic.inc" }, { "agent_id": "<64-hex>", "name": "morgan", "skills_summary": "...", "methods_count": 4, "trust_tier": 2, "verification_path": "org-asserted", "trust_warning": "verification-incomplete", "owner_id": "acme.inc" } ] ]]> The response is a JSON array of objects, each carrying the fields below. The required minimum identifies the agent and its trust posture; optional fields are present when the underlying Agent Identity Document or Agent Genesis populates them. DISCOVER /agents Entry Fields

Field	Required	Description
`agent_id`	MUST	Canonical Agent-ID (64 lowercase hex characters).
`name`	MUST	Operator-assigned short name; unique only within a server.
`skills_summary`	SHOULD	Short human-readable description of the agent's capabilities.
`methods_count`	SHOULD	Number of registered endpoints exposed by this agent.
`trust_tier`	MUST	Resolved trust tier (1, 2, or 3) per the precedence rule in .
`verification_path`	MUST	Resolved verification path (`dns-anchored`, `log-anchored`, `hybrid`, or `org-asserted`) per the precedence rule in .
`trust_warning`	CONDITIONAL	Present when the resolved `trust_tier` is `2` or when an operator-set warning is recorded. Value is a short token (e.g., `verification-incomplete`).
`owner_id`	SHOULD	Resolved owner identifier per the precedence rule in . Omitted when no Agent Genesis is loaded and the operator set no explicit value.

The trust_tier, verification_path, trust_warning, and owner_id fields are resolved per the trust-posture loading rule in ; the server MUST NOT emit values that disagree with what it would resolve on the Agent Identity Document itself. Clients MAY branch on trust_tier or on trust_warning presence to drive trust UI; clients MUST NOT treat absence of trust_warning for a Tier-2 entry as authoritative (the server is required to populate it when auto-applicable, but operator-set absence is preserved per ). A server MAY restrict the entries returned by DISCOVER /agents based on the requesting agent's Authority-Scope, governance zone, or other policy. Entries withheld by policy are NOT errors; the server returns the remaining entries with no indication of the withholding. Implementations that need a full server inventory should consult the operator out of band.

DISCOVER / Servers MUST expose DISCOVER / (root path). The endpoint returns the directory of reserved built-in DISCOVER endpoints the server exposes, allowing clients to enumerate available inventory surfaces in a single call: The response MUST be a JSON object with a directory field whose value is a JSON array of entries. Each entry MUST carry a path field naming a reserved built-in endpoint defined in this section, and a tier field with the value "A" (per the tier classification in ). Servers MAY include operator-authored DISCOVER paths in the directory at their discretion; operator-authored entries MUST carry tier: "B" and lack the protection afforded to reserved built-ins (). DISCOVER / is the cheapest probe a client can issue against an unknown server to determine what inventory surfaces are available without speaking application-specific paths.

DISCOVER /tools Servers that host one or more tool-bearing agents MAY expose DISCOVER /tools. The endpoint returns a compact inventory of tools the server makes available, in a format parallel to DISCOVER /methods. When exposed, the endpoint MUST return a JSON array of tool descriptors each carrying a name field, an optional description, and an optional agent_id indicating which agent owns the tool. The precise tool-descriptor schema is operator-defined; implementations that integrate with adjacent agent frameworks (MCP, A2A) SHOULD align their tool-descriptor shape with the framework in use.

DISCOVER /apis Servers that expose one or more REST-style APIs via the HTTP gateway profile of MAY expose DISCOVER /apis. The endpoint returns an inventory of the APIs the server presents, each carrying a name, an optional description, and an optional base_url. The endpoint is intended for clients that need to discover the HTTP-side surface of a gateway-fronted server without loading the full server manifest.

Reserved DISCOVER Paths The path prefixes /methods, /agents, /genesis, /tools, /apis, /patterns, and /contracts under the DISCOVER method are reserved for the built-in endpoints defined in this section and, for /patterns and /contracts, in . Operator-authored endpoints MUST NOT register DISCOVER under any of these reserved paths, and MUST NOT register DISCOVER under any path whose first path segment is the same as a reserved name regardless of additional segments (e.g., /methods/v2, /agents-extended, /toolset, /patterns-v2, /contracts/history). Servers MUST refuse a registration that would shadow a reserved path; the refusal is a configuration-time error and never reaches a runtime status code. The reservation applies only to DISCOVER. The same paths under other verbs are unreserved (e.g., a server MAY register QUERY /methods to serve a non-DISCOVER catalog endpoint without conflict). A server MAY override a reserved built-in by registering its own endpoint at the same (DISCOVER, /reserved-path) pair via the operator manifest's policies.builtins override block (see ). Operators that override a built-in MUST preserve the documented response contract for that endpoint; clients that detect a contract violation SHOULD treat the server as nonconforming.

Legacy DISCOVER target= Body Form (Deprecated) Earlier revisions of AGTP and this document supported a body-keyed DISCOVER target=... form in which the discovery surface was selected by a target field in the request body in place of the request path (e.g., DISCOVER target=methods instead of DISCOVER /methods). The body-keyed form is DEPRECATED. Servers MAY accept the legacy form for transition compatibility and SHOULD emit a one-shot deprecation warning per requesting agent the first time the legacy form is received. Servers that accept the legacy form MUST treat DISCOVER target=X as semantically equivalent to DISCOVER /X for the reserved targets methods, agents, genesis, tools, and apis. A future revision MAY remove acceptance of the legacy form entirely. When a request carries both a reserved path and a target body field with conflicting values, the server MUST return 400 Bad Request with a body identifying the conflict; the server MUST NOT silently prefer one over the other. Future revisions of this specification MAY define additional built-in endpoints (e.g., QUERY /catalog-version for catalog version negotiation). Built-ins are protocol-level conventions, present on every conforming server.

The Endpoint Primitive

Endpoint Tiers Every (method, path) pair an AGTP server exposes, or could expose, falls into one of four tiers. The tier classification is informational at the wire level but constrains how the server treats invocations against the pair: Tier A endpoints are guaranteed by the protocol, Tier B endpoints are guaranteed by the operator, Tier C endpoints are negotiated at runtime, and unregistered endpoints are refused.

Tier A, Protocol-native:: Endpoints reserved by AGTP and its companion specifications. Servers MUST support Tier A endpoints with their documented response contracts. Examples: every cell of the eighteen-method floor defined in ; the DISCOVER built-in endpoints defined in (DISCOVER /, DISCOVER /methods, DISCOVER /agents, DISCOVER /genesis, and the optional /tools and /apis); any future reserved path under the reservations specified in .
Tier B, Application-registered:: Endpoints declared in the operator's manifest. Servers MUST support every Tier B endpoint as declared and MUST refuse any operator-authored endpoint that collides with a Tier A reservation per . The endpoint catalog, semantic block, handler binding, and policy fields documented in this section apply to Tier B endpoints.
Tier C, Runtime-negotiated:: Endpoints synthesized at runtime via RCNS (Runtime Contract Negotiation Substrate). A caller addresses a (method, path) pair that is absent from the registered Tier A or Tier B set; the server MAY, subject to policy, synthesize a contract for the requested pair from registered primitives and deliver it in either of two modes: a 461 RCNS Contract Available response carrying a contract preview (confirm- first), or an inline executed response carrying a Contract-Synthesized header (optimistic). RCNS is specified in .
Unregistered:: A (method, path) pair that matches no Tier A, Tier B, or Tier C endpoint. Servers MUST refuse unregistered invocations with 404 Not Found (when the path is unknown) or 405 Method Not Allowed (when the path is known under a different method). When RCNS is enabled, the server MAY treat an unregistered pair as a Tier C synthesis candidate instead of refusing.

Servers MUST classify every inbound (method, path) pair into exactly one of these tiers before dispatch. The DISCOVER / directory and DISCOVER /methods inventory SHOULD carry an explicit tier field on each entry ("A", "B", or "C") so clients can route follow-on invocations with full knowledge of the guarantee level. Entries on DISCOVER / describing reserved built-in endpoints carry tier: "A"; entries on DISCOVER /methods describing operator-authored endpoints carry tier: "B". Tier C entries appear on DISCOVER /contracts per . The four-tier classification is normative. The specific operator-side mechanism for declaring Tier B endpoints (manifest format, dispatcher gates, policy fields) is documented in the remainder of this section.

Endpoint Definition An endpoint is the structural unit an AGTP server exposes. Every endpoint is identified by the pair (method, path) and carries the fields specified below.

Required Fields Every endpoint MUST carry: method, path, description, semantic, input_schema, output_schema, errors, handler. The remaining fields (namespace, required_scopes, deprecated) are optional.

Field Semantics

method:: An AGTP method drawn from the approved catalog at the server's declared catalog_version, or a custom method declared in the server's policies.methods. See . The method name satisfies the lexical rules of .
path:: A path conforming to the path grammar of .
description:: Operator-facing prose. A single sentence describing what the endpoint does. Distinct from the semantic block: description is for humans reading the manifest; the semantic block is for agents reasoning about applicability.
namespace:: Optional free-form grouping label. Servers exposing many endpoints MAY organize them by namespace (e.g., "reservations", "billing", "inventory") for catalog navigation. Namespace has no protocol-level effect.
semantic:: The structural metadata block defined in . Required by every endpoint with all required fields populated.
input_schema:: A Draft 2020-12 JSON Schema describing the request body. MUST declare "type": "object" and MUST set "additionalProperties": false; see . Query-string parameters () are merged into the body before validation; the same schema covers both. Schemas are inline in v00; cross-endpoint schema sharing is anticipated in a future revision.
output_schema:: A Draft 2020-12 JSON Schema describing the successful response body. SHOULD set "additionalProperties": true so handlers can return forward-compat fields without bumping the contract; see .
errors:: An array of named error condition strings (e.g., "room_unavailable", "invalid_dates"). The dispatcher returns these as 422 Unprocessable Entity responses with structured bodies identifying the condition. The error name is the protocol-level identifier; status-code mapping is a response-construction concern outside the contract. The list MAY be empty for endpoints with no business-level errors. Composition handlers MUST include "composition_failed"; external-service handlers MUST include the upstream-failure codes specified in .
handler:: An object declaring the binding kind: {"type": "registered_function" | "composition" | "external_service"}. The handler's full reference (Python dotted path, recipe name, upstream URL, etc.) is server-internal implementation detail and MUST NOT appear in the manifest. See and .
required_scopes:: An array of scope identifiers the invoking agent MUST declare in its Authority-Scope. Empty or omitted means no scope check; the endpoint inherits the server's default authority requirement. A request whose identity document grants a required scope but whose claimed Authority-Scope omits it returns 262 Authorization Required with reason scope-required; a request whose document omits the scope returns 455 Scope Violation with reason scope-exceeded, per the obtainability rule of .
deprecated:: Optional per-endpoint deprecation metadata. See .

Endpoint Deprecation The optional deprecated block follows the same shape as catalog-level deprecation: Field semantics:

deprecated_in:: The server's agtp_api_version (or operator-chosen version string) at which the endpoint became deprecated.
removed_in:: Optional. The version at which the endpoint will be removed. MAY be absent if removal is anticipated but the target version remains unchosen.
successor:: Optional. An object with method and/or path indicating the replacement endpoint callers should migrate to. Either field MAY be absent: successor.method alone signals "use the same path with a different method"; successor.path alone signals "use the same method with a different path"; both together signal a full replacement endpoint.

An endpoint MAY be deprecated even when its method and path remain valid in the catalog and grammar. Operators deprecate endpoints when they migrate callers from one (method, path) pair to another, independent of catalog evolution. The dispatcher MUST stamp an AGTP-Endpoint-Warning advisory header on every response from a deprecated endpoint: The header rides alongside AGTP-Catalog-Warning () when both apply; both are advisory and the request still processes. The successor token renders as METHOD /path when both fields are present; just one is emitted bare. removed_in and successor are omitted from the header when the endpoint's deprecated block declares none. The header fires on success and failure responses alike so callers always see the warning regardless of how the invocation resolves.

The Semantic Block The semantic block is the part of the endpoint definition that encodes machine-readable intent. Agents reason about whether an endpoint matches their goal by interpreting the semantic block; the verb and path alone carry too little. The empirical evidence for intent-expressing semantic metadata is documented in .

Required Semantic Block Fields

intent:: A natural-language statement of what the endpoint accomplishes. Single sentence. Imperative form. Subject is the actor; the verb reflects the endpoint's method; the object is the path resource.
actor:: A short identifier of the agent class that should invoke this endpoint. Freeform string; servers MUST NOT reject endpoints whose actor is outside any fixed enum. Suggested values for interoperable authoring include agent, human, system, customer, staff, admin; domain-specific values (e.g., merchant, auditor, patient) are equally valid. The field helps agents recognize whether an endpoint applies to their role but carries no protocol-level enforcement.
outcome:: A natural-language statement of the post-condition produced by successful invocation. Distinct from intent, which states the endpoint's aim; outcome states the condition that holds after.
capability:: The high-level effect class of the endpoint. One of the catalog category values: discovery, retrieval, analysis, transaction, modification, creation, notification, mechanics, domain_spanning. The same taxonomy used to classify methods in the catalog (); an endpoint inherits its capability class from the category of the method it exposes, though operators MAY override when the endpoint's specific behavior fits a different category.
confidence:: A scalar on the closed interval 0.0 to 1.0 expressing how confident the server is that an agent invoking this endpoint with satisfying input will produce the declared outcome. Endpoints that are deterministic and well-tested score near 1.0; endpoints that involve nondeterministic resources or partial failures score lower. Agents MAY use the value to decide whether to ESCALATE for clarification before invoking. The scale is advisory and undefined normatively beyond the endpoint bounds.
impact:: The reversibility of the endpoint's effect. One of informational, reversible, irreversible. informational endpoints return data without changing server state. reversible endpoints change state in ways that can be undone via a compensating method (e.g., CANCEL, REFUND). irreversible endpoints commit changes that cannot be programmatically undone (e.g., sending an email, charging a payment, shipping goods). Governance frameworks MAY use this field to apply differential authorization, gating irreversible endpoints behind stricter scope checks.
is_idempotent:: Boolean. Whether repeated invocation with identical input produces identical effect. Affects retry safety: idempotent endpoints MAY be retried freely; non-idempotent endpoints require an Idempotency-Key header per to be safely retried.

Server Manifest

Purpose The server manifest is the machine-readable document an AGTP server publishes to declare its identity, version posture, endpoint catalog, hosted agents, hosted protocols, and operational policies. It is the authoritative statement of what the server offers.

Manifest Structure The manifest carries three distinct version fields at the top level:

agtp_version:: The AGTP wire-protocol version the server speaks (currently "1.0", corresponding to AGTP v07 of ). Pinned to the base protocol, separate from contract-layer evolution.
agtp_api_version:: The AGTP-METHODS contract-layer version the server implements (currently "1.0" for v00 of this document). Carries the endpoint primitive, semantic block, status code, and synthesis semantics the server adheres to.
document_version:: This specific manifest's revision string. Operator-controlled; bumped on every meaningful change to the manifest content. Permits agents to detect cache staleness without comparing the full body.
catalog_version:: The method catalog version the server validates incoming method names against. See .
catalog_versions_supported:: List of catalog versions the server can validate against. The current catalog_version MUST appear in this list. Multi-version support is anticipated in a future revision; v00 places no normative requirement on it.

Server Block The server block carries operator-facing identity and timestamps:

server_id:: The canonical identifier for this server. MAY be a hostname, an agtp:// URI, or an operator-defined string. The value MUST be stable across manifest revisions.
domain:: Optional operational host the server is reachable at, if different from server_id. MAY be null for servers whose server_id is itself a resolvable host.
operator:: The human-readable name of the organization operating the server.
contact:: An operator email address for security and operational issues.
supported_features:: A non-normative array of feature tokens describing optional capabilities the server exposes (e.g., "endpoint-registry", "synthesis", "openapi-export"). Agents and operators MAY use these tokens for capability negotiation. AGTP-METHODS leaves define a canonical feature-token registry; tokens are operator-defined.
issued:: ISO 8601 timestamp of when the manifest was first issued. Preserved across manifest revisions; identifies the manifest's origin in time.
updated:: ISO 8601 timestamp of the most recent manifest update. Bumped on every regeneration.

Methods Inventory The embedded_methods array carries the eighteen protocol-floor methods every AGTP server supports (the method definitions of ). The custom_methods array carries any server-specific methods declared in policies.methods (). The custom_methods field MAY be omitted when empty. These arrays are non-normatively populated; agents that want a complete picture of what the server invokes SHOULD consult the endpoints array, which is the authoritative source for what methods are bound on what paths.

Agents Disclosure The three agent-related fields describe what agents the server hosts and how they are exposed:

agent_disclosure:: One of "public", "authenticated", "private". "public" means the hosted_agents array is populated and visible to all callers. "authenticated" means the array is populated only for callers presenting a verified identity. "private" means the array is empty in this manifest; the server withholds what agents it hosts.
hosted_agents:: Array of agent identity references the server hosts. Format is operator-defined; typical entries include a canonical Agent-ID and a brief description.
agent_disclosure_notice:: Optional human-readable note explaining the server's disclosure policy. Useful when agent_disclosure is "authenticated" or "private" and the operator wishes to clarify how a caller can request elevated disclosure.

Protocol and API Surface

apis:: Optional array of legacy HTTP API descriptors this server fronts (e.g., entries pointing at OpenAPI specifications). Servers that expose a parallel HTTP surface alongside AGTP list those HTTP endpoints here; the presence of entries in this array is the signal that the server operates a parallel HTTP face.
hosted_protocols:: Optional array of higher-level agent protocols the server hosts on top of AGTP (e.g., MCP, A2A, ACP). Each entry is an object with the protocol name and the AGTP method+path it is reachable via. See the composition section in .

Policies Block The policies block carries five normative operational toggles:

wildcards_accepted:: One of true, false. Whether the server accepts ad-hoc method invocations against undeclared endpoints. See .
anonymous_discovery:: One of true, false. Whether unauthenticated callers may retrieve the manifest and invoke read-only built-in endpoints. Servers serving traffic from open networks typically set this to true; servers in trusted-network deployments MAY set this to false to require authenticated identity even on discovery.
scope_required_for_invocation:: One of true, false. Whether every endpoint invocation requires the agent to carry an Authority-Scope header. true is the default; the implementation MUST NOT invoke a handler when scope is required and absent. Servers with no scope-gated endpoints MAY set this to false.
synthesis_enabled:: One of true, false. Whether the server participates in runtime contract negotiation per . When false, PROPOSE requests MUST be refused with 463 Proposal Rejected and the reason synthesis-disabled.
max_synthesis_depth:: A positive integer specifying the maximum number of composed steps a synthesized endpoint MAY combine. Plans exceeding this depth are refused; the server falls through to counter-proposal or 463. Default: 10.

Manifest Signature The manifest body SHOULD carry a manifest_signature field holding a JWS envelope over the canonical JSON encoding of the manifest body (with the manifest_signature field itself excluded from the signing input). Clients MAY verify signatures when present and SHOULD warn when absent. A future revision of this specification will require manifest_signature (MUST); reference implementations are expected to ship signing support in alignment with that revision. The signing algorithm, canonical encoding rules, and key-publication mechanism are specified separately in . For v00, the manifest_signature field MAY be null or absent; a manifest without a valid signature MUST NOT be rejected by clients on signature grounds alone.

Endpoint Projection The endpoints array in the manifest is a projection of the server's internal endpoint catalog. The projection differs from the full endpoint definition in one important way: the handler's type-specific reference field (function, recipe, or url) and any other binding-specific implementation details are dropped, leaving only the binding kind under handler.type. Internal endpoint definition (on disk, in TOML or equivalent): Manifest projection (over the wire): The type-specific reference field (function, recipe, or url) and any other binding-specific implementation fields are server-internal and MUST NOT appear in the manifest projection. Agents need to know the binding kind to reason about expected latency, retry behavior, and attribution; agents have no need to know which specific function, recipe, or upstream realizes the binding. Servers MUST apply this projection on every manifest response.

Manifest Retrieval Agents retrieve the server manifest using the AGTP DISCOVER method without an Agent-ID header. The absence of an Agent-ID header distinguishes a server-level discovery request (returns the manifest) from an agent-level discovery request (returns endpoint candidates matching specified criteria). The response MUST carry Content-Type application/vnd.agtp.manifest+json. The manifest SHOULD be signed; see the Manifest Signature subsection above.

Manifest Caching Servers SHOULD include Cache-Control and ETag headers on manifest responses. Agents MAY cache manifests according to those directives. The server.updated timestamp and the document_version field permit agents to detect cache staleness without revalidation: a changed document_version means the manifest content has changed in a way the operator considers significant.

Method Policy

Purpose The method policy declares per-server constraints on which methods are accepted, which legacy verbs are opted in, and which methods on which paths are redirected to their canonical form. The policy is the small, declarative complement to the endpoint catalog: the catalog says what an endpoint does; the policy says what the server accepts in the first place. In v00, the method policy is a sub-block of the manifest's policies () carried at policies.methods. There is no separate policy file format; the policy is part of the server's configuration and is exposed on the manifest like any other operational policy.

Policy Shape The policies.methods block is a JSON object with four optional fields:

allow:: Either the string "*" (the server admits every method in the catalog plus any method it has registered an endpoint for), or an array of method names (the server admits only the listed methods plus the embedded eighteen floor methods). Default: "*".
disallow:: Array of method names the server explicitly rejects. Disallowed methods are rejected with 405 Method Not Allowed even when they would otherwise be permitted by allow. Useful for narrowing the surface without enumerating an exhaustive allow list. Default: empty array.
legacy:: Either an array of legacy HTTP verb names the server opts into (e.g., ["GET", "POST"]), the string "*" (all five legacy verbs accepted), or the string "NONE" (no legacy support). Granular per-verb opt-in supersedes the binary enabled/disabled toggle. Default: "NONE".
aliases:: Map of method-name aliases the server resolves at dispatch time. Each entry maps a non-canonical method name (the alias) to a canonical method name (the catalog entry the server actually dispatches under). Servers seed the default alias map with the five legacy HTTP verbs mapped to their canonical AGTP counterparts: Operators MAY override the seed or wipe it. Alias resolution runs as a single-hop translation ahead of the catalog check (); the resolved method is what the dispatcher matches against the endpoint catalog and against the allow/disallow lists. Aliases MUST NOT chain (an alias MUST resolve to a canonical method, never to another alias); chained or circular alias declarations are configuration errors and the server MUST refuse to start. The original wire method is preserved on the Attribution-Record as requested_method per , so audit consumers can distinguish "BOOK arrived under that name" from "GET arrived and was aliased to FETCH." Default: the seed map above; operators MAY set the empty map to disable aliasing entirely.
redirects:: Array of method-and-path redirect entries. Each entry has the shape: Either from_path or to_path MAY be omitted, signaling "redirect on method-only without path matching" or "redirect to the same path as the request." When a request matches a redirect entry, the server processes it as if the to_method/to_path had been requested instead.

Example A representative TOML deployment configuration: The equivalent manifest projection:

Catalog-Graceful Skip Semantics When the server loads policies.methods and an entry under allow, disallow, or redirects references a method removed from the catalog version the server has loaded, the entry MUST be skipped at load time with an operator-visible warning. The server continues to operate; the policy entry has no effect. This permits operators to upgrade the catalog without simultaneously rewriting the policy block to remove all references to retired methods. The legacy field continues to enforce the strict set of five HTTP verbs (GET, POST, PUT, DELETE, PATCH); typos or non-HTTP verbs in legacy MUST be rejected as a hard configuration error.

Discovery Agents discover the server's method policy by retrieving the server manifest via target-less DISCOVER (). The policies.methods block is part of the manifest body and is covered by the manifest signature () when present. Agents that need only an endpoint inventory (which methods on which paths are exposed, with one-line descriptions) MAY invoke DISCOVER /methods () for a lightweight response that bypasses the full manifest weight.

Status Codes AGTP-METHODS contributes the following structural rejection codes to the AGTP status code space. The codes are registered with IANA per . The negotiation codes 261, 263, 461, 463, and 464 are specified in . AGTP-METHODS Status Codes

Code	Name	Returned When
262	Authorization Required	Authorization the agent could obtain is absent: credential establishment, wildcards consent, authenticated identity, or an Authority-Scope the agent's identity document grants but the request omitted from its claim (`scope-required`). A request that exceeds the authority the document grants is 455 Scope Violation per , never 262.
404	Not Found	The path is absent from this server
405	Method Not Allowed	Method and path are valid but server policy exposes no endpoint at that combination
459	Method Violation	The method is in neither the floor nor the AGTP-METHODS catalog nor the server's custom methods
460	Endpoint Violation	The path violates AGTP-METHODS path grammar (method-name leakage)

The four structural rejection codes 404, 405, 459, 460 form the contract-level failure surface. Each code is independently actionable and addresses a distinct failure mode:

404: the path is absent. The agent has the wrong resource locator.
405: the method and path are individually valid but the server exposes no endpoint at that combination. The agent MAY try a different method on the same path, or MAY PROPOSE the combination per unless it is a policy refusal.
459: the method itself is the problem. The agent MUST pick a different method from the AGTP-METHODS catalog. PROPOSE cannot succeed because the method name is illegal.
460: the path violates path grammar. The agent MUST restructure the path so that no path segment matches an approved method name.

These codes are never interchangeable. A 405 is a policy decision; a 459 is a vocabulary violation; a 460 is a structural violation. Servers MUST return the most specific code that applies. The 405 response body MUST carry the following fields:

allowed_methods_for_path:: The list of methods the server's policy or registry exposes on the path. Permits the agent to retry with a method the server actually accepts.
redirects_for_path:: A map of method-to-method redirects from policies.methods () that apply to this path. MAY be empty if the server has no redirects configured. Permits the agent to learn that, for example, BOOK is redirected to RESERVE on this server, and to retry under the canonical method directly.

These fields together let the agent recover from a 405 without an out-of-band catalog lookup or a second round-trip.

Handler Binding The handler binding declares how an endpoint's behavior is realized at runtime. AGTP-METHODS defines three binding types.

Composition A composition handler exposes a derived endpoint whose behavior is realized by orchestrating one or more other endpoints declared on the same server. No application code is invoked; the server's synthesis runtime executes the composition steps in order, threading outputs into subsequent inputs while preserving authority through the composition. Composition handlers reference a named recipe: The recipe value is the name of a recipe defined in the server's recipe registry (typically in an agtp-recipes.toml file loaded at server startup). Recipes are reusable units of composition logic; one recipe MAY back multiple endpoints. Inline composition steps are unsupported in v00; every composition endpoint MUST reference a registered recipe by name. A composition handler whose recipe value fails to resolve to a recipe registered at the server MUST cause registration to fail. Composition endpoints MUST include composition_failed in their errors list. This is the named error returned when a composition step fails (any underlying step returns a non-success status). The error response carries details with the recipe identifier, the failed step number, the underlying step method, the underlying status, and any captured outputs from preceding steps. Authority MUST be preserved through the composition. The agent's Authority-Scope MUST be sufficient for every underlying endpoint the composition invokes. Authority sufficiency is verified at registration time when possible (when step paths are statically known) and at invocation time for parameterized step paths. This is the binding type most commonly produced by PROPOSE + synthesis (): a synthesized endpoint is a composition of existing endpoints, and MAY be promoted to a permanent declared composition by adding its recipe to the server's recipe registry and writing an endpoint TOML that references it by name.

Registered Function The endpoint binds to a Python function loaded by the server. The function value is a dotted path identifying the module and the callable within it; the server imports the module at registration time and resolves the function attribute. Single dotted-path form is normative; separate module and function fields are supported in v00. The function receives the validated input and returns the response payload. The handler API contract is specified separately (see the reference implementation's handler documentation). This is the binding type used for endpoints whose logic exceeds what composition can express.

External Service The endpoint binds to an HTTPS-reachable external service. The server proxies the request to the external endpoint and returns the response. Useful for AGTP frontends to existing HTTP-based services. The external service sits outside the AGTP contract; the AGTP server presents a contract-compliant facade. External service fields:

url:: The fully-qualified URL of the upstream HTTP endpoint. The URL MUST use the https:// scheme. Plain HTTP MUST NOT be accepted at registration time.
method:: The HTTP method to invoke on the upstream. One of GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS. REQUIRED.
headers:: Optional map of HTTP headers to send to the upstream on every request. Values MAY include ${VAR} placeholders that are resolved against process environment variables at server startup. Unresolved placeholders MUST cause registration to fail.
input_transform:: Optional map renaming AGTP request fields to the upstream's expected field names. Maps AGTP-side name to upstream-side name. Fields absent from the map pass through unchanged. v00 supports key-rename only; value transformation is anticipated in a future revision.
output_transform:: Optional map renaming upstream response fields to AGTP-side names. The map is inverted at response time: the upstream-side name is the source, the AGTP-side name is the target. Fields absent from the map pass through unchanged.
error_map:: Map of upstream HTTP status code (as a string) to AGTP error code declared in the endpoint's errors list. Servers SHOULD configure error_map for every external_service binding; without it, upstream HTTP errors surface to agents as generic upstream_error. Every value in error_map MUST appear in the endpoint's errors; registration MUST fail otherwise.
timeout_seconds:: Positive number specifying the upstream request timeout. Servers SHOULD set this explicitly. Default if absent: 30 seconds. Production deployments SHOULD NOT rely on the default; explicit timeouts make upstream-dependency posture inspectable.

External service endpoints MUST declare the following five upstream-failure error codes in their errors list, in addition to any business-level errors: Mandatory External Service Error Codes

Error Code	Returned When
`upstream_timeout`	Upstream failed to respond within `timeout_seconds`
`upstream_connection_error`	DNS, connection, or TLS failure reaching the upstream
`upstream_malformed_response`	Upstream returned a 2xx response whose body is invalid JSON
`upstream_authentication_failed`	Upstream returned 401 or 403 outside `error_map`
`upstream_error`	Upstream returned 4xx or 5xx outside `error_map`

Without these declarations, transport-level upstream failures would surface to the calling agent as opaque server errors. Mandating them makes the failure surface inspectable. Agent identity is NOT propagated to the upstream by default. This is a security default: an upstream HTTP service has no mechanism to validate AGTP agent credentials, and forwarding them would expose principal identity to systems outside the AGTP trust boundary. Servers wishing to propagate identity to a specific upstream MAY do so via explicit header configuration in the headers map (e.g., a server-issued bearer token), but MUST NOT forward AGTP-Agent-ID or AGTP-Principal-ID by default.

Conformance

Single Tier A conforming AGTP server implements:

The full endpoint primitive ().
Method validation against the AGTP-METHODS approved catalog at the server's declared catalog_version.
Path validation against the AGTP-METHODS path grammar.
Method policy enforcement at dispatch time per the policies.methods sub-block of the manifest ().
A server manifest retrievable via DISCOVER without an Agent-ID header, carrying agtp_version, agtp_api_version, document_version, catalog_version, and catalog_versions_supported per .
The endpoint projection of : manifest responses expose handler.type only; the full handler binding stays server-side.
PROPOSE handling per when policies.synthesis_enabled is true; refusal with 463 and reason synthesis-disabled otherwise.
The structural rejection status codes 404, 405, 459, 460, plus the contract-negotiation codes 261, 262, 263, 463.
The semantic block on every published endpoint.
Schema validation for input and output, with the strictness asymmetry of (inputs strict, outputs permissive by default).
The AGTP-Catalog-Warning response header on every response whose method is deprecated in the catalog version in use.
The AGTP-Endpoint-Warning response header on every response from a deprecated endpoint ().
The DISCOVER /methods built-in endpoint defined in , returning a compact inventory of the server's registered endpoints.

There is no halfway tier. Servers that implement only method dispatch without endpoints are nonconformant; they are something else, perhaps a method-based RPC layer that happens to use AGTP framing.

Validation Timing Conforming servers perform two distinct categories of validation, which SHOULD occur at different times. Manifest validation at startup. Before accepting traffic, a conforming server SHOULD validate every endpoint declared in its manifest:

The method appears in the AGTP-METHODS approved catalog (or is a custom method declared in the server's policies.methods).
The path conforms to the path grammar of .
The semantic block carries every required field.
The input and output schemas are syntactically valid JSON Schema documents.
The handler binding () is resolvable: a composition's referenced steps exist; a registered_function's module and function are importable; an external_service's URL is well-formed and reachable (or marked deferred).

A server that detects manifest validation failures at startup SHOULD reject the offending endpoints with clear logs and SHOULD NOT silently ignore them. Servers MAY choose to fail startup entirely on manifest validation failure or to start serving with the valid subset of endpoints; the choice is a deployment policy decision. Request validation at runtime. Once serving, the server validates each inbound request against the endpoint's input schema before invoking the handler:

Input body validates against the endpoint's input_schema. On failure, the server returns 422 Unprocessable Entity with a schema validation error report.
The agent's Authority-Scope covers the endpoint's required_scopes declarations. On failure, the server returns 455 Scope Violation per .
The verb-and-path combination is exposed by the server's current policy. On failure, the server returns the appropriate structural rejection code (404, 405, 459, or 460).

Validation MUST complete before handler invocation. Servers MUST NOT invoke the handler with input that has yet to satisfy schema validation; handler authors may therefore assume their context structure is well-formed. The startup-vs-runtime split is a recommendation, with no strict requirement behind it. Servers MAY defer startup validation if their endpoint catalog is dynamic (e.g., a PROPOSE-driven server that only synthesizes endpoints on demand). Such servers MUST still validate at the moment an endpoint is registered or synthesized, just later than process start.

Schema Strictness Schema validation is asymmetric between inputs and outputs. Input schemas MUST be validated with additionalProperties: false. Fields undeclared in the schema MUST cause validation to fail with 422. This catches typos and version mismatches early and prevents agents from sending unrecognized fields under the mistaken assumption that the server will silently ignore them. Output schemas SHOULD be validated with additionalProperties: true. A handler that returns more fields than the schema declared is forward-compatible: agents can ignore extra fields. A handler that returns missing required fields fails validation either way. Servers MAY opt into strict output validation as a deployment policy. This asymmetry is normative because it has interoperability consequences. An agent submitting a request to two different servers can rely on both rejecting unrecognized input fields. An agent consuming a response from two different servers can rely on forward-compatible field handling without the version of one server breaking integration with the other.

Conformance Test Suite The AGTP-METHODS conformance test suite is maintained at: The test suite exercises every conformance requirement in this section. Servers claiming AGTP-METHODS conformance MUST pass the test suite at the version they declare in their manifest's agtp_api_version field.

Relationship to OpenAPI OpenAPI and AGTP-METHODS perform similar work for different consumers. OpenAPI describes HTTP APIs for human developers and code generators. AGTP-METHODS describes agent-grade endpoints for AI agents. The two formats overlap substantially: both have method, path, request schema, response schema, errors, and descriptions. AGTP-METHODS leaves compete with OpenAPI; the two formats are complementary, and AGTP-METHODS deployments commonly interoperate with OpenAPI-described HTTP services in three distinct ways.

OpenAPI as an Authoring Source A developer with an existing OpenAPI specification can convert it to AGTP-METHODS endpoint declarations. The HTTP method maps to an AGTP verb (with translation: GET maps to QUERY or DISCOVER depending on intent; POST maps to EXECUTE, BOOK, PURCHASE, or another action verb depending on the operation). The path remains. The request schema becomes input_schema. The response schema becomes output_schema. Errors become the errors list. The OpenAPI description becomes the foundation of the semantic block's intent. Conversion tooling MAY automate this translation with reasonable defaults, flagging cases where verb mapping requires human judgment. The AGTP-METHODS approved method catalog () is the canonical source for HTTP-verb-to-AGTP-verb conversions: tooling MUST select target verbs from the list and MUST NOT invent ad-hoc names, so that converted endpoints interoperate with native AGTP-METHODS endpoints. Servers SHOULD NOT treat machine-converted endpoints as authoritative until a human reviewer has confirmed the semantic block fields, particularly intent, outcome, and impact. Conversion tooling has no reliable way to infer impact from HTTP semantics; that is judgment work.

OpenAPI as an Export Target AGTP servers MAY render their endpoint catalog as an OpenAPI document for HTTP consumers that have yet to speak AGTP. The reverse translation: AGTP verb to HTTP method (mappings are imperfect; some AGTP verbs have no clean HTTP analog and are best exposed via the legacy verb compatibility mode of ); path remains; schemas pass through unchanged. The exported OpenAPI document is a compatibility surface; the authoritative description of the service is the AGTP contract. Servers exporting OpenAPI SHOULD mark the export as derived from the AGTP-METHODS manifest and SHOULD publish a link to the manifest in the OpenAPI info section.

OpenAPI as a Coexistence Layer A server MAY run AGTP and HTTP side by side. The AGTP endpoint's handler is the same code that the HTTP endpoint invokes. OpenAPI documents the HTTP face; the AGTP-METHODS manifest documents the agent face. The handler is unaware which protocol invoked it. This deployment posture is signaled in the manifest by populating the apis array () with descriptors pointing at the server's OpenAPI documents. Coexistence is the path most large organizations will take during transition: AGTP is added as a parallel face without requiring modification to existing HTTP infrastructure. Implementation guidance for coexistence deployments is provided in .

What AGTP-METHODS Does Not Specify AGTP-METHODS explicitly excludes the following operational concerns, which remain middleware territory:

Rate limiting strategies and policies.
Observability mechanisms (metrics, logs, traces).
Caching strategies beyond manifest caching.
Custom authentication beyond AGTP agent identity.
Request and response transformation.
Server deployment topology and scaling.
HTTP middleware integration patterns.

These are addressed by reference middleware repositories that ship separately and non-normatively. The contract-plumbing distinction is intentional: the protocol governs contracts; middleware governs operations.

Security Considerations

Verb-Path Tampering A malicious server could declare endpoints whose semantic blocks misrepresent the actual behavior. Agents that rely on semantic blocks for authorization decisions MUST also enforce Authority-Scope checks against actual outcomes, whatever the declared outcomes say.

Method Policy Tampering The method policy is carried as the policies.methods sub-block of the server manifest and is covered by the manifest signature () when present. Agents that retrieve the manifest over AGTP transport (which mandates TLS 1.3 or higher per ) and verify the manifest signature when present are protected against policy substitution. Agents that accept manifest content without signature verification are vulnerable to substitution of the method policy along with the rest of the manifest.

Wildcard Abuse The wildcards: true consent on the agent identity document and wildcards_accepted: true on the server policy are mutual consents that enable ad-hoc method invocation. Agents SHOULD NOT declare wildcards: true by default; servers SHOULD NOT declare wildcards_accepted: true for any operation with non-trivial impact tier. The 262 Authorization Required response with body type wildcards-required is the safe default.

Verb List Trust Agents trust the AGTP-METHODS approved method catalog as published at https://agtp.io/api/methods.json. The catalog MUST be served over HTTPS with content integrity verification. Agents MAY pin the catalog version they have validated and refuse upgrades without explicit operator approval.

IANA Considerations This document defines the following AGTP-METHODS-specific allocations. The corresponding AGTP IANA registry entries are recorded in the IANA Considerations of .

AGTP-METHODS Status Code Assignments The following codes are registered in the AGTP Status Code Registry with AGTP-METHODS as the authoritative reference:

262 Authorization Required
405 Method Not Allowed (AGTP-METHODS semantics)
459 Method Violation
460 Endpoint Violation

Codes 261, 263, and 463 are registered by .

Media Type Registrations This document defines two media types and requests their registration in the IANA Media Types registry. Full registration templates are provided in the appendices. The master registry of all AGTP-family media types is maintained in . Media Types Defined by AGTP-METHODS

Media Type	Use	IANA Status
`application/vnd.agtp.manifest+json`	AGTP server manifest format	Planned (this document)
`application/vnd.agtp.endpoint+json`	AGTP-METHODS endpoint definition format	Planned (this document)

application/vnd.agtp.manifest+json: The AGTP server manifest format. Carried on responses to a server-level DISCOVER request (DISCOVER without an Agent-ID header). The registration application will be filed concurrent with publication of this document.
application/vnd.agtp.endpoint+json: The AGTP-METHODS endpoint definition format. Also used in PROPOSE request bodies and in 263 Proposal Approved response bodies per . The registration application will be filed concurrent with publication of this document.

Until the IANA registrations complete, the reference implementation MAY continue to emit application/json for manifest and endpoint responses. Implementations are encouraged to emit the AGTP-specific types as soon as they are accepted by IANA so that content negotiation and MIME-based routing can distinguish AGTP traffic from generic JSON.

AGTP-METHODS Method Catalog Reference This document does NOT request establishment of an IANA AGTP-METHODS Verb List Registry. As specified in , the verb list is maintained as an open living artifact at https://agtp.io/api/methods.json, in contrast to a closed IANA registry, so that the catalog can evolve at the cadence of agent deployment, which outpaces the cadence of IETF process. The structural status codes, the media types, and the response headers below are appropriate for IANA registration; the method catalog is the exception.

AGTP-METHODS Response Headers This document requests registration of the following response headers in the IANA HTTP Field Name Registry (or equivalent AGTP-specific registry where applicable):

AGTP-Catalog-Warning:: Advisory header surfaced on responses where the request method is deprecated in the catalog version the server is using. Format and semantics specified in .
AGTP-Methods-Version:: Header surfacing the AGTP-METHODS specification version a manifest or endpoint definition conforms to. Servers MAY continue to emit AGTP-API-Version with the same value during the transition from ; clients SHOULD accept either.

Open Items The following items are explicitly out of scope for this revision and are anticipated in future revisions:

Mandatory manifest signing. v00 specifies manifest_signature as SHOULD; a future revision will make it MUST. The signing algorithm, canonical encoding rules for JWS over the manifest body, key-publication mechanism, and rotation semantics will be specified separately in . Reference implementations are expected to ship signing support in alignment with that revision.
Cross-endpoint shared schema definitions. v00 requires schemas to be inline in each endpoint's input_schema and output_schema. A future revision may add a schemas block to the manifest carrying named, reusable JSON Schema fragments referenceable by $ref from endpoint schemas. This is a manifest-size and authoring-convenience concern; the wire-level contract is unchanged.
Config-driven document_version. The reference implementation hardcodes document_version to a single value; an operator workflow that bumps document_version on every meaningful manifest change is anticipated. The wire contract permits any string; the operator workflow is implementation-specific.
The full text of the method catalog (425 methods at catalog version 1.0.0). v00 references the published JSON at agtp.io/api/methods.json without embedding the catalog inline. A future revision may include a curated subset as a non-normative appendix; the authoritative catalog remains the published JSON per .
Multi-version concurrent catalog support. v00 specifies catalog_versions_supported as a list, but the runtime semantics for a server processing a request against a non-default catalog version are unspecified. Most servers list a single version.
Conformance test suite specification (currently described by reference to the agtp.io implementation).
Authentication passthrough for external_service handlers (forwarding agent identity to upstream services). Out of scope by security default in v00; future revisions may specify opt-in passthrough patterns.
Hot reload of endpoints and policies.methods configuration without server restart.
Wildcard opt-in implementation. The spec defines the consent model (wildcards: true on the agent identity document and wildcards_accepted: true in server policy) and the 262 Authorization Required response with body type wildcards-required for unsatisfied consent. The default behavior in v00 is refuse: ad-hoc method invocations against undeclared endpoints are rejected as 459 Method Violation when the method is absent from the AGTP-METHODS catalog, or as 405 Method Not Allowed when the method is in the catalog but no endpoint exists for the method-and-path combination. The opt-in path that activates 262 with wildcards-required (when an agent or server has consented to wildcards but the other side has withheld it) is anticipated as an implementation extension in a future revision.

Changes from AGTP-METHODS v01 This is the first revision under the AGTP-METHODS name. Changes from :

Document renamed. The title, abbreviation, and every self-reference now read AGTP-METHODS; the subtitle names the vocabulary as the AGTP verbs.
Runtime contract negotiation extracted. The PROPOSE and Synthesis chapter, the DISCOVER /patterns and DISCOVER /contracts built-ins, status codes 261, 263, 461, 463, and 464, the RCNS headers, and the related open items are now specified in . The tier taxonomy, the manifest policy fields, the composition handler, and the reserved path list keep their references to RCNS and point there.
Method Properties () and Method Overview () added, giving the vocabulary the structure HTTP gives its methods: named properties, a floor table, and the catalog as the registry beyond the floor.
Catalog baseline stated: 425 methods in nine categories at catalog version 1.0.0, with the eighteen-method floor as the embedded set and the five legacy HTTP verb mappings.
AGTP-API-Version response header renamed AGTP-Methods-Version, with the old name accepted during transition.
References updated to base AGTP v09 and the current companion revisions.

The v00 to v01 change history of is carried in that document. Wire format compatibility: apart from the response header rename, nothing on the wire changes; a server conformant to is conformant to this document.

Acknowledgments This document was developed in response to implementation experience with the AGTP v06 through v09 builds. The architectural decision to unify the contract layer under a single companion specification was articulated in the AGTP-METHODS Position Paper (May 2026). The empirical evidence for the action-intent semantic class is documented in .

References Normative References Key words for use in RFCs to Indicate Requirement Levels In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. Uniform Resource Identifier (URI): Generic Syntax A Uniform Resource Identifier (URI) is a compact sequence of characters that identifies an abstract or physical resource. This specification defines the generic URI syntax and a process for resolving URI references that might be in relative form, along with guidelines and security considerations for the use of URIs on the Internet. The URI syntax defines a grammar that is a superset of all valid URIs, allowing an implementation to parse the common components of a URI reference without knowing the scheme-specific requirements of every possible identifier. This specification does not define a generative grammar for URIs; that task is performed by the individual specifications of each URI scheme. [STANDARDS-TRACK] HTTP Semantics The Hypertext Transfer Protocol (HTTP) is a stateless application-level protocol for distributed, collaborative, hypertext information systems. This document describes the overall architecture of HTTP, establishes common terminology, and defines aspects of the protocol that are shared by all versions. In this definition are core protocol elements, extensibility mechanisms, and the "http" and "https" Uniform Resource Identifier (URI) schemes. This document updates RFC 3864 and obsoletes RFCs 2818, 7231, 7232, 7233, 7235, 7538, 7615, 7694, and portions of 7230. Agent Transfer Protocol (AGTP) JSON Schema: A Media Type for Describing JSON Documents Informative References URI Template A URI Template is a compact sequence of characters for describing a range of Uniform Resource Identifiers through variable expansion. This specification defines the URI Template syntax and the process for expanding a URI Template into a URI reference, along with guidelines for the use of URI Templates on the Internet. [STANDARDS-TRACK] AGTP-RCNS: Runtime Contract Negotiation Substrate AGTP Composition Profiles AGTP Transport Bindings: TCP/TLS and QUIC AGTP Agent Certificate Extension AGTP Identifier Chain AGTP Trust and Verification Specification AGTP-API: Verbs, Paths, Endpoints, and Synthesis Semantic Method Naming and LLM Agent Accuracy: A Controlled Benchmark of REST/CRUD versus Agentive API Interface Design

Implementation Guidance (Informative) This appendix is non-normative. It describes how a typical AGTP-METHODS server is built and operated, in support of the normative spec. Implementations are free to depart from these patterns; they are documented here to give developers a starting point and to make the intended deployment posture concrete.

Endpoint Authoring A common pattern for endpoint authoring uses one declaration file per endpoint, dropped into a directory the server scans at startup:

A human-editable declaration format (such as TOML, YAML, or JSON) carries the endpoint's verb, path, semantic block, schemas, errors, and handler reference.
A handler function in the server's implementation language carries the business logic, with a uniform signature that takes a context object and returns a structured response or named error.
The server scans the endpoints directory at startup, validates every declaration per , imports the referenced handler modules, and registers the valid endpoints in memory.

The reference Python implementation of an AGTP-METHODS server uses TOML for endpoint declarations and a context → response | error handler signature. The exact TOML schema and handler patterns are documented at: The TOML format is one serialization of the endpoint primitive defined normatively in . Other implementations are free to use other serializations (YAML, JSON, programmatic registration, database-backed registries) provided the in-memory endpoint structure conforms to this specification.

Runtime Dispatch Pattern A typical request lifecycle in a conforming server:

The server receives an AGTP request and parses the verb and path.
The dispatcher matches verb-and-path against the endpoint registry.
On match, the dispatcher validates the request body against the endpoint's input_schema.
The dispatcher checks the agent's Authority-Scope against the endpoint's required_scopes declaration.
The dispatcher constructs a context object containing validated input and agent identity, and invokes the handler.
The handler executes business logic and returns a structured response or named error.
The dispatcher translates the result into an AGTP response.

This pattern keeps contract enforcement in the dispatcher and business logic in the handler. Handlers perform no schema validation, authority checks, or status code translation; the dispatcher does those uniformly across all endpoints.

Tooling The reference implementation publishes tooling at agtp.io/api/tools:

An OpenAPI → AGTP-METHODS conversion tool that reads an OpenAPI 3.x specification and produces a directory of endpoint declarations with external_service handlers pointing at the original HTTP endpoints. Where method mapping requires judgment, the tool flags the endpoint for human review with suggestions from the AGTP-METHODS approved method catalog.
An AGTP-METHODS → OpenAPI export tool for servers that need to publish HTTP-side documentation derived from their AGTP manifest.
A manifest validator that checks server manifests against this specification.
A conformance test runner that exercises the conformance requirements of against a deployed server.
An agtp-catalog-diff tool that diffs two catalog versions and optionally scans a deployment directory to identify endpoints, recipes, and policies.methods entries that would break under a candidate catalog upgrade. See .

These tools are non-normative reference implementations. Other parties MAY publish equivalent tools.

Operator Notes on Catalog Versioning The catalog version a server declares in its manifest is the version the server validates incoming method names against. Operators upgrading the on-disk verb catalog (typically by regenerating methods.json from a new catalog release) SHOULD restart the server to pick up the new version. Long-running servers that have yet to be restarted after a catalog upgrade may continue validating against the previous version they loaded at startup. The reference agtp-catalog-diff tool is the recommended way to preview the impact of a catalog upgrade against a deployment directory before restarting. Run it as a deployment gate to surface endpoint-removal conflicts, recipe-step references to removed methods, and policies.methods entries that would be silently skipped under the new version.

Migration Paths (Informative) This appendix is non-normative. It describes deployment strategies for organizations adopting AGTP-METHODS on top of existing HTTP services, and for organizations building AGTP-METHODS services from scratch.

Wrap-and-Expose The lowest-effort path for an organization with an existing HTTP API. The HTTP API stays where it is. The organization writes AGTP-METHODS endpoint declarations whose handler binding is external_service (), pointing at the existing HTTP endpoints. AGTP-METHODS becomes a thin agent-facing facade over the existing HTTP backend. Request translation maps AGTP-METHODS input fields to HTTP request shapes; response translation maps HTTP responses back to AGTP-METHODS output shapes. The OpenAPI → AGTP-METHODS conversion tool () produces this kind of deployment by default: it reads an OpenAPI specification and generates a directory of endpoint declarations with external_service handlers. A 200-endpoint HTTP service can be made AGTP-discoverable in a single working session. Wrap-and-expose deployments SHOULD populate the manifest's apis array () with the HTTP API descriptor so agents that need the underlying HTTP documentation can locate it.

Translate-and-Rehost A medium-effort path. The organization reimplements its HTTP API as native AGTP-METHODS endpoints, moving the handler logic into the AGTP-METHODS server's implementation language directly. The HTTP layer goes away, or is retained only for legacy clients that have yet to migrate. The conversion tool generates endpoint declarations and handler stubs from the existing OpenAPI specification, leaving the organization to fill in the business logic. The work is conversion of an existing API. Translate-and-rehost deployments leave the apis array empty once the HTTP layer retires; during transition, the array continues to reference the HTTP face alongside the native AGTP endpoints.

Coexist-Permanently No migration. The HTTP API stays as the primary interface; AGTP-METHODS is added as a parallel face for agent traffic. New agent-facing features go in the AGTP-METHODS layer; existing HTTP traffic continues unaffected. Two interfaces, one backend, indefinite coexistence. Most large enterprises will take this path when regulatory, contractual, or customer-relationship commitments make a forced migration impractical. AGTP-METHODS is additive in this posture and replaces nothing. Coexist-permanently deployments populate apis with HTTP API descriptors that include http_documentation URLs pointing at the HTTP-side OpenAPI specifications.

Path Selection The right path depends on organizational constraints; the protocol has no preference. Wrap-and-expose scales: it can move millions of existing APIs to AGTP-METHODS discoverability at low cost. Translate-and-rehost is the right target for organizations committing to AGTP-METHODS as the long-term primary surface. Coexist-permanently is the honest answer for organizations that want agent access without disruption to existing systems.