AAuth Rich Resource Requests (R3)

4.1. Resource Metadata Extensions

R3 extends the /.well-known/aauth-resource.json document defined in AAuth Protocol ([I-D.hardt-aauth-protocol]):¶

{
  "resource": "https://calendar.example.com",
  "r3_vocabularies": {
    "urn:aauth:vocabulary:mcp": "https://calendar.example.com/mcp",
    "urn:aauth:vocabulary:openapi": "https://calendar.example.com/openapi.json"
  }
}

r3_vocabularies (OPTIONAL). A JSON object mapping vocabulary URIs to their discovery endpoints. Keys MUST be vocabulary URIs from the urn:aauth:vocabulary: namespace for standard vocabularies defined in this document, or third-party URI namespaces for proprietary vocabularies. Values are vocabulary-specific discovery endpoints (the MCP server URL, the OpenAPI spec URL, the gRPC reflection endpoint, etc.). A resource MAY advertise multiple vocabularies simultaneously.¶

4.2. Standard Vocabularies

This document defines seven standard vocabularies. Third parties MAY define additional vocabularies using their own URI namespaces. Each vocabulary defines: the vocabulary URI, the structure of operation requests, how the resource maps operations to R3 documents, and the discovery endpoint.¶

Standard vocabularies use the urn:aauth:vocabulary: namespace.¶

{
  "vocabulary": "urn:aauth:vocabulary:mcp",
  "operations": [
    { "tool": "create_calendar_event" },
    { "tool": "modify_calendar_event" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:openapi",
  "operations": [
    { "operationId": "createEvent" },
    { "operationId": "updateEvent" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:grpc",
  "operations": [
    { "method": "calendar.CalendarService/CreateEvent" },
    { "method": "calendar.CalendarService/UpdateEvent" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:graphql",
  "operations": [
    { "operation": "CreateCalendarEvent", "type": "mutation" },
    { "operation": "GetCalendarEvents", "type": "query" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:asyncapi",
  "operations": [
    { "operationId": "publishCalendarUpdate", "action": "send" },
    { "operationId": "receiveCalendarEvent", "action": "receive" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:wsdl",
  "operations": [
    { "operation": "CreateCalendarEvent", "service": "CalendarService" }
  ]
}

{
  "vocabulary": "urn:aauth:vocabulary:odata",
  "operations": [
    { "operation": "Events", "methods": ["GET", "POST", "PATCH"] },
    { "operation": "Events/SendCancellation" }
  ]
}

5.1. Request

The agent sends r3_operations in the authorization endpoint request:¶

POST /authorize HTTP/1.1
Host: calendar.example.com
Content-Type: application/json
Signature-Input: sig=("@method" "@authority" "@path" "signature-key");created=1741824000
Signature: sig=:...signature bytes...:
Signature-Key: sig=jwt;jwt="eyJhbGc..."

{
  "r3_operations": {
    "vocabulary": "urn:aauth:vocabulary:openapi",
    "operations": [
      { "operationId": "createEvent" },
      { "operationId": "updateEvent" }
    ]
  }
}

r3_operations (OPTIONAL). An object containing:¶

• vocabulary (REQUIRED). A URI identifying the vocabulary. MUST be supported by the resource as advertised in r3_vocabularies.¶
• operations (REQUIRED). An array of operation requests. Structure is vocabulary-specific; see {{mcp-vocabulary}} through {{odata-vocabulary}}.¶

When r3_operations is present, the resource maps the declared operations to an appropriate R3 document and includes r3_uri and r3_s256 in the resource token. When r3_operations is absent, the resource MAY still include R3 claims in the resource token based on its own policy.¶

5.2. Response

The resource returns a resource token as defined in AAuth Protocol ([I-D.hardt-aauth-protocol]), extended with R3 claims:¶

{
  "resource_token": "eyJhbGciOiJFUzI1NiJ9..."
}

The resource token contains r3_uri and r3_s256 identifying the R3 document that covers the requested operations. The resource's internal mapping from operations to R3 documents is opaque to the agent.¶

5.3. Operations Spanning Multiple Definitions

A resource MAY organize its authorization definitions into multiple R3 documents internally — for example, one document per scope or operation group. A single resource token carries exactly one r3_uri/r3_s256 pair, and the resulting auth token therefore pins exactly one R3 document.¶

When an agent's r3_operations request includes operations that the resource maps to more than one of its internal definitions, the resource MUST compose a single R3 document that covers all of the requested operations and reference that composed document in the resource token:¶

• operations is the union of the requested operations, expressed in the request's vocabulary.¶
• display describes the combined access. The resource MAY merge the display sections of the underlying definitions (for example, concatenating implications and unioning data_accessed) or author a purpose-built summary for the combination.¶

The composed document is served at a fresh content-addressed r3_uri exactly like any other R3 document (see {{content-addressing}}): the resource builds it on the fly and persists the serialized bytes so the hash remains stable across the AS and PS fetches. Because R3 documents are content-addressed, an identical combination of operations reduces to the same hash and MAY be cached and reused across requests.¶

This composition is opaque to the agent. The agent sees one r3_uri/r3_s256 in the resource token regardless of how many internal definitions the requested operations were drawn from, and the auth token's r3_granted and r3_conditional claims express the granted operations against that single composed document.¶

6.1. Fields

version (RECOMMENDED). A string identifying the version of this R3 document. The combination of r3_uri + SHA-256 hash provides content-addressing independent of this field; version is for human readability.¶

vocabulary (REQUIRED). The vocabulary URI identifying how operations are expressed. MUST match one of the vocabularies the resource advertises in r3_vocabularies.¶

operations (REQUIRED). An array of operations covered by this R3 document, using the vocabulary-specific structure defined in {{mcp-vocabulary}} through {{odata-vocabulary}}. This is the same format used in the agent's r3_operations request and in the auth token's r3_granted and r3_conditional claims.¶

display (RECOMMENDED). Human-readable descriptions of the consequences of granting this access. The resource describes what it does, not what the agent intends:¶

• summary (REQUIRED if display present). A short plain-language description suitable for a consent UI or auth agent.¶
• implications (OPTIONAL). Side effects of granting this access: emails sent, records modified, costs incurred.¶
• data_accessed (OPTIONAL). What data becomes visible to the caller.¶
• irreversible (OPTIONAL). Plain-language description of actions that cannot be undone.¶

6.2. Content Addressing

The R3 hash (r3_s256) is computed as the SHA-256 hash of the bytes of the R3 document as served by the resource, base64url-encoded without padding.¶

The resource's serialization is the document. There is no canonicalization step — verifiers hash the bytes received over the wire, not a normalized form. Resources MUST serialize the R3 document once and serve those exact bytes verbatim on every request for the same r3_uri. Re-serialization between hash computation and serving (e.g. middleware that parses and re-stringifies JSON, CDN minification, response framework helpers that reorder keys) will produce different bytes and break hash verification. Resources that build R3 documents on the fly SHOULD persist the serialized bytes (e.g. in a key-value store keyed by r3_uri or r3_s256) rather than re-build the document per request.¶

The r3_s256 hash is the document's identity, not the URI. The AS caches documents by hash. If a resource updates the document at the same URI, existing auth tokens still reference the previous hash (which the AS has cached). New resource tokens reference the new hash. This enables:¶

• Infinite caching by the AS. A document that verifies against its hash need never be re-fetched.¶
• Permanent audit records. An auth token carrying r3_s256 identifies the exact authorization semantics that were approved, regardless of subsequent changes at the same URI.¶

6.3. Resource Token Extensions

R3 extends the resource token defined in AAuth Protocol ([I-D.hardt-aauth-protocol]) (a JWT with typ: resource+jwt) with two additional payload claims. When a resource includes R3 information, it MUST include both.¶

Base claims (from AAuth Protocol): - iss: Resource URL - dwk: aauth-resource.json - aud: Auth server URL - jti: Unique token identifier - agent: Agent identifier - agent_jkt: JWK Thumbprint of the agent's signing key - iat: Issued at timestamp - exp: Expiration timestamp - scope: Requested scopes (optional)¶

R3 extension claims: - r3_uri (REQUIRED for R3): The URI where the AS can fetch the R3 document. The AS authenticates itself using an HTTP Message Signature. - r3_s256 (REQUIRED for R3): The SHA-256 hash of the R3 document at r3_uri, base64url-encoded without padding.¶

{
  "typ": "resource+jwt",
  "alg": "EdDSA",
  "kid": "resource-key-1"
}

{
  "iss": "https://calendar.example.com",
  "dwk": "aauth-resource.json",
  "aud": "https://as.example.com",
  "jti": "rt-8f3a2b",
  "agent": "assistant@agent.example",
  "agent_jkt": "NzbLsXh8uDCcd-6MNwXF4W_7noWXFZAfHkxZsRGC9Xs",
  "r3_uri": "https://calendar.example.com/r3/a1b2c3d4",
  "r3_s256": "aBcDeFgHiJkLmNoPqRsTuVwXyZ0123456789abcd",
  "iat": 1741824000,
  "exp": 1741824300
}

Resource tokens MAY include both scope (as defined in AAuth Protocol ([I-D.hardt-aauth-protocol])) and R3 claims. When both are present, the AS MUST enforce both independently.¶

7.1. AS Processing

When the AS receives a resource token containing r3_uri and r3_s256, it MUST:¶

1. Validate the resource token signature per AAuth Protocol ([I-D.hardt-aauth-protocol]).¶
2. Fetch the R3 document at r3_uri. The AS MAY use a cached copy if the cache entry was stored with the same r3_s256 value.¶
3. Compute the SHA-256 hash of the bytes received and compare it to r3_s256. If the hashes do not match, the AS MUST reject the resource token.¶
4. Record r3_uri and r3_s256 in its audit log alongside the token issuance event, the agent identifier, and the timestamp.¶
5. Use the operations section for policy evaluation.¶
6. Include r3_uri, r3_s256, r3_granted, and (if applicable) r3_conditional in the issued auth token.¶

7.2. Caching

Because R3 documents are content-addressed, the AS MAY cache them by r3_s256 to avoid redundant fetches. When serving a cached entry, the AS MUST verify that the stored document produces the expected hash. The AS is not required to retain R3 documents beyond their immediate use in token issuance — the AS's audit log records r3_uri and r3_s256, which is sufficient for later verification by re-fetching.¶

8.1. Resource Enforcement

The resource matches each incoming API call against the auth token claims:¶

1. Match in r3_granted: serve the request.¶
2. Match in r3_conditional: build a per-call proposal ({{per-call-proposals}}) and return AAuth-Requirement with a resource token referencing it. The AS evaluates the specific call against the proposed parameters.¶
3. No match: reject the request.¶

No token introspection or R3 document fetch is needed at enforcement time. The resource uses the vocabulary it already understands.¶

When r3_operations was not used (the agent received the resource token via a 401 rather than the authorization endpoint), the AS populates r3_granted and r3_conditional based on the operations defined in the R3 document and its own policy. The AS decides which operations to grant outright and which to make conditional.¶

9.1. Proposal Document

In addition to the R3 document fields ({{r3-document}}), a per-call proposal carries:¶

operations (REQUIRED). The single conditional operation being invoked, in the resource's vocabulary.¶

parameters (REQUIRED). The concrete parameters of the call, machine-readable, for the AS to evaluate and the resource to bind. A large or sensitive value MAY be represented by a digest object in place of the inline value:¶

• s256 (REQUIRED). BASE64URL(SHA-256(value-bytes)) of the parameter value as it will be presented at call time.¶
• excerpt (OPTIONAL). A short, human-readable excerpt of the value for display.¶
• media_type (OPTIONAL). The media type of the value.¶

display (RECOMMENDED). Per-call human context for the user's approval decision. In addition to the structured fields defined in {{r3-document}}, a proposal's display MAY include a detail Markdown string. The following sections are RECOMMENDED as a convention (not normative requirements): ## Action, ## To / ## Recipient, ## Details, ## Content (excerpt), ## Irreversible.¶

{
  "vocabulary": "urn:aauth:vocabulary:mcp",
  "operations": [ { "tool": "send_email" } ],
  "parameters": {
    "to": "mom@example.com",
    "subject": "Dinner Sunday?",
    "body": { "s256": "aBcD…", "excerpt": "Hi Mom, are you free…", "media_type": "text/plain" }
  },
  "display": {
    "summary": "Send an email as you",
    "detail": "## Action\nSend an email\n\n## To\nmom@example.com\n\n## Details\nSubject: Dinner Sunday?\n\n## Content\nHi Mom, are you free…"
  }
}

9.2. Flow

1. Conditional challenge. The agent invokes an r3_conditional operation. The resource builds the proposal, persists it keyed by its r3_s256, and returns AAuth-Requirement with a resource token whose r3_uri/r3_s256 reference the proposal. The token carries only the reference, not the parameters.¶
2. Approval. The AS fetches the proposal and evaluates parameters per policy; the PS renders display for user consent. On approval, the AS issues a per-call auth token that echoes the proposal's r3_uri/r3_s256 and lists the now-approved operation in r3_granted.¶
3. Enforced retry. The agent retries the actual call with the per-call auth token. The resource recovers the proposal from its store via r3_s256 and MUST verify that the agent's actual parameters match the approved proposal. For any parameter represented as a digest, the resource MUST verify that BASE64URL(SHA-256(presented-value)) equals the stored s256. If anything differs, the resource MUST reject the call. An approval to email one recipient cannot be replayed against another.¶

9.3. Large and Sensitive Payloads

Representing a parameter as a digest keeps large or sensitive payloads out of every token and away from the PS: only the s256 and a short excerpt appear in the proposal. The full bytes travel directly from the agent to the resource at call time, where the resource verifies them against the digest. This is also a privacy control — the resource chooses what the PS (and through it, the user-facing surface) sees versus what stays between the agent and the resource.¶

Whether the AS or PS additionally machine-evaluates parameters (for example, auto-denying a payment over a threshold) or treats the proposal as display-for-human-consent only is deployment policy. The parameters are present in the proposal either way, so machine policy can be layered on without a format change.¶

10.1. R3 Document Access Restriction

The AS MUST authenticate itself when fetching r3_uri using an HTTP Message Signature as defined in the AAuth Protocol ([I-D.hardt-aauth-protocol]). The resource MUST reject requests not signed by its AS.¶

This prevents agents from fetching R3 documents by following the r3_uri they carry in the resource token. Since a resource has exactly one AS, the resource only needs to recognize signatures from that AS. The agent opacity property (agents carry the hash of a document they cannot read) depends on this restriction.¶

10.2. R3 Endpoint Access Control

The agent opacity property — agents carry the hash of a document they cannot read — depends entirely on the resource correctly restricting access to R3 document endpoints. If the resource fails to require a valid HTTP Message Signature from its AS on R3 document requests, or accepts signatures from keys other than its AS's, agents can fetch and read R3 documents, breaking the opacity guarantee. Implementations MUST treat R3 endpoint access control as a critical security requirement and SHOULD verify this restriction during deployment testing.¶

10.3. Hash Verification

The AS MUST verify r3_s256 against the fetched document before using it. Failure to verify allows a resource to serve different content than what was hashed in the resource token.¶

10.4. Audit Log Integrity

The AS MUST write audit log entries atomically with token issuance. An auth token issued without a corresponding audit log entry creates an undetectable gap in the observability record. Implementations SHOULD use transactional writes or equivalent mechanisms.¶

10.5. Operation Validation

For all vocabularies, the resource MUST validate declared operations against its authoritative definition (MCP tool list, OpenAPI spec, .proto file, GraphQL schema, AsyncAPI spec, WSDL document, or OData $metadata) before issuing a resource token.¶

10.6. Grant Enforcement

Resources MUST enforce r3_granted and r3_conditional claims in auth tokens. Operations in r3_granted define fully authorized access. Operations in r3_conditional MUST trigger an AAuth-Requirement before being served. The resource MUST reject API calls that do not match an operation in either claim.¶

11.1. JWT Claims Registration

This document requests registration of the following JWT claims in the IANA JSON Web Token Claims registry:¶

11.2. R3 Vocabulary Registry

This specification establishes the AAuth R3 Vocabulary Registry. The initial contents are:¶

New values may be registered following the Specification Required policy ([RFC8126]).¶

A.1. Why Not RAR

OAuth 2.0 Rich Authorization Requests ([RFC9396]) defines authorization_details as a structured extension to authorization requests. RAR is a natural reference point for this work. R3 deliberately does not use or profile RAR for the following reasons:¶

Directionality. RAR is client-declared: the agent constructs authorization_details and sends it to the AS. The agent defines what it wants. R3 is resource-declared: the resource defines what access it provides and signs that definition. The agent cannot modify or reframe it. This is the opposite directionality, and the security properties depend on it.¶

Agent opacity. In R3, the R3 document is fetched by the AS directly from the resource, restricted to AS-only access. The agent carries a hash of a document it cannot read. RAR has no equivalent because the client constructs the authorization details and necessarily knows their content.¶

Content addressing. R3 uses a content-addressed URI plus SHA-256 hash to pin the exact authorization semantics in effect at approval time. An auth token carrying r3_uri and r3_s256 is a permanent, verifiable record. RAR carries no equivalent versioning or integrity guarantee.¶

Audit trail. The AS records r3_uri in its audit log, creating a durable reference to the exact R3 document version. This is not possible with RAR's inline authorization_details structure.¶

RAR and R3 are complementary. RAR remains appropriate for client-declared authorization detail. R3 addresses the resource-declared case that RAR was not designed for.¶