11-Step Verification Pipeline

Every intercepted capability call goes through this pipeline before your code runs. A fresh JWT is minted and verified on every call — no token reuse.

The Pipeline

Step-by-Step

Step	What is checked	Error on failure
1-2	Decode JWT header + payload, confirm typ = "agent+jwt"	token_invalid
3	sub matches registered agentId	agent_not_found
4	iss matches registered public key thumbprint	token_invalid
5	aud matches the requested capability name	capability_denied
6	hostThumbprint in token matches agent's registered Host	token_invalid
7	Ed25519 signature is valid	token_invalid
8	exp/iat temporal check (30s clock skew tolerance)	token_expired / token_invalid
9	JTI not seen in the 90-second replay window	token_replayed
10	Agent holds an active grant for this capability	capability_denied
11	Grant has not expired (expiresAt)	capability_denied
11b	Call arguments satisfy all grant constraints + required fields	constraint_violated

JWT Claims

Every capability call mints a JWT with these claims:

{
  "typ": "agent+jwt",
  "alg": "EdDSA",
  "sub": "<agentId>",
  "iss": "<agent JWK thumbprint>",
  "aud": "<capability name>",
  "hostThumbprint": "<Host JWK thumbprint>",
  "jti": "<random 128-bit base64url>",
  "iat": 1700000000,
  "exp": 1700000060,
  "hostname": "<agent hostname>",
  "agentName": "<agent name>"
}

The token is:

• Signed with the agent's Ed25519 private key
• Single-use — a unique jti per call (replayed within 90s = rejected)
• Short-lived — 60-second TTL
• Scoped — aud must exactly match the capability name

JTI Replay Protection

The jti (JWT ID) is cached for 90 seconds. If the same jti appears again within that window, the call is rejected with token_replayed. This prevents replay attacks even if an attacker captures a valid token.

By default, JTI tracking is in-memory. For multi-process deployments, inject a Redis adapter.

Auth Overhead

The entire pipeline runs in-process via Node.js Web Crypto. Typical warm-path overhead is under 1ms per call:

const stats = chain.getStats();
// stats.authOverhead → { avgMs: number, maxMs: number }

Every audit entry records authOverheadMs so you can monitor the cost of the auth pipeline.