Individual Internet-Draft. This page describes the design in draft-sweeney-wimse-credential-delegation, which no IETF Working Group has adopted. The draft describes intended behavior, and the Cred SDK implements part of it. The SDK’s conformance table lists what is implemented, and section 06 lists where the Cred server goes further than the draft.
a proposed profile for credential delegation in multi-agent systems
MCP standardizes how agents connect to tools. A2A standardizes agent-to-agent communication. No open standard governs how credentials flow through these connections. The Cred Protocol fills this gap — a profile of existing OAuth 2.0 standards that defines secure, auditable credential delegation for AI agents.
AES-256-GCM encrypted OAuth tokens. Referenced by opaque credential_handle only. Decrypted in-process, never returned to callers. AWS KMS key management with per-account isolation.
Issues DPoP-bound DelegationTokens. Validates capability attenuation — authority can only narrow, never widen. Generates DS-signed prf receipts per hop to prevent RFC 8693 chain splicing.
CIBA-derived backchannel consent — no redirect URI required. Immutable consent records retained minimum 90 days. Re-consent triggered on scope widening, expiry, or agent_operator change.
Validates DT and DPoP proof, retrieves credential from Vault, calls resource server with stored token, returns API response only. Synchronous revocation: root <1s, full tree <5s. Tamper-evident audit log with consent_id chain tracing.
sequenceDiagram participant S as Subject participant DS as Delegation Server participant V as Credential Vault participant A as Agent participant RS as Resource Server Note over S,V: Phase 1 - Setup once per OAuth service S->>DS: Authorize Google OAuth DS->>S: Redirect to provider consent S->>DS: OAuth callback with auth code DS->>DS: Exchange code for refresh token DS->>V: Encrypt and store token AES-256-GCM DS-->>S: Connection confirmed Note over A,DS: Phase 2 - Agent identity bootstrap A->>A: Generate Ed25519 keypair A->>A: Derive did:key from public key A->>A: Sign JWT Bearer assertion exp 5 min Note over A,DS: Phase 3 - Delegation request A->>DS: POST /delegate JWT Bearer plus capabilities DS->>DS: Validate JWT Bearer and jti replay check DS->>S: CIBA backchannel consent push S-->>DS: Approve with scoped capabilities DS->>DS: Issue DelegationToken DPoP-bound 15 min TTL DS->>DS: Sign prf receipt for chain integrity DS-->>A: DelegationToken issued Note over A,A: Agent holds DT only - not the credential Note over A,RS: Phase 4 - Credential exercise A->>DS: POST /exercise with DT and DPoP proof DS->>DS: Validate DT signature and expiry DS->>DS: Verify DPoP binding to htm and htu DS->>DS: Check operation within authorization_details DS->>V: Retrieve credential by credential_handle V-->>DS: Decrypted token in-process only DS->>RS: API call with stored access token RS-->>DS: API response DS-->>A: API response only - credential not returned DS->>V: Wipe decrypted credential from memory Note over S,DS: Phase 5 - Revocation S->>DS: DELETE /delegation/jti DS->>DS: Invalidate root DT within 1 second DS->>DS: Cascade to child DTs within 5 seconds DS-->>A: HTTP 401 delegation_revoked on next request
Agents receive DelegationTokens, not credentials. Raw OAuth tokens never cross the agent trust boundary. Prompt injection cannot exfiltrate what the agent does not hold.
Sub-delegation can only narrow capabilities — never widen. Enforced by the DS at issuance time, not by trusting agents. Authority flows downward and only shrinks.
Not "access Google Drive" but drive.files.get on folder:abc123 with max_calls: 10 until time T. Eliminates ambient authority — the confused deputy exploit surface.
Root DT invalidated in <1s. Full delegation tree cascade in <5s. Eventual consistency is not acceptable for adversarially-promptable systems.
Each hop produces a DS-signed prf receipt cross-referencing parent and child jti values. Closes the RFC 8693 §2.1-2.2 vulnerability disclosed to the OAuth WG, February 2026.
Agents derive identity from an Ed25519 keypair via did:key, with no key registration. The draft has the DS accept a DID it has not seen before. Identity is not authority: the Cred server recognizes the key and issues nothing until an operator grants it a Permission.
Wire protocol. act and may_act provide delegation semantics. Production at MS Entra, Google, Auth0, Keycloak. Deliberately underspecified — Cred defines the delegation mechanics it leaves open.
Proof-of-possession per request. Bound to HTTP method and URI. Prevents DelegationToken theft and replay without requiring changes at resource servers.
Structured capability claims via authorization_details. Operation + resource + constraints. Not coarse OAuth scopes — resource- and operation-level granularity.
Agent authentication to DS. Signed with ephemeral Ed25519 key. No client_secret required. Supports did:key as issuer. Five-minute maximum assertion lifetime.
Async consent without redirect URI. Agent-initiated flows pause awaiting Subject approval on mobile. The only consent flow that works for fully autonomous agents.
DID derived deterministically from Ed25519 public key. Self-certifying, with no CA. The Cred server authenticates any such key, and authorizes it only through a Permission.
The sections above describe the draft. The Cred server names some things differently (the draft’s exercise call is POST /api/v1/use) and adds these checks. The docs have the detail and the current limits.
An agent named by agent_did needs a Permission for the service, or it gets 403 permission_required. An operator grants one at POST /admin/permissions.
Each brokered call is classified on the exact URL, method and body forwarded. An unmapped call gets 403 operation_not_mapped, and one the scopes do not cover gets 403 scope_insufficient.
github.read, salesforce.read and notion.read are read-only delegations that the broker enforces, derived from stored consent. They are never issued as raw tokens.
Revoking a Permission or a delegation holds the agent out, with 403 agent_access_revoked, until an operator grants it again.
Each brokered call event names the agent DID, its key thumbprint and the delegation. Refusals are events too.
Raw provider tokens are refused with raw_token_disabled unless the operator turns on a deprecated setting. The enforcement checks above have deprecated opt-outs too, each logged at startup.
The reference implementation is on GitHub. Issues, security analysis, and PRs welcome — particularly from implementers and IETF participants.