ADR-019: MCP Client identity -- production extension point (mTLS / workload identity)

Status: Documented (spike), not implemented in the RI (EP-01-T04, could). Related: this ADR is the canonical record of this decision. See RI/docs/AS-BUILT.md and RI/docs/interfaces/identity-oidc.md for implementation and validation context. Also relevant: README.md §19, RF-02, Axiom 2.

Context

The RI's identity boundary (src/emcp_bus/identity/authn.py, EP-01-T03) is OAuth 2.1 client_credentials against Keycloak (EP-01-T01): every MCP Client authenticates with a client_id + secret, gets a short-lived JWT, and the Gateway validates that JWT's signature/issuer/audience/expiry against the IdP's JWKS before trusting the azp claim as the caller's identity.

This is a deliberate, confirmed decision for the RI (not a placeholder): OAuth 2.1 satisfies every property README.md §19 actually requires of an identity mechanism -- short-lived, revocable, audience-bound, verifiable without a shared long-term secret in the request path. Real enterprise deployments commonly go further, using mTLS (mutual TLS client certificates) or a cloud provider's workload identity federation (e.g. SPIFFE/SPIRE, AWS IAM roles for service accounts, GCP Workload Identity, Azure Managed Identity) instead of -- or in addition to -- a bearer JWT. This ADR documents how that extension plugs into the RI's existing seam, without requiring one, so a production fork isn't blocked on a redesign.

Decision

Do not implement mTLS/workload identity in the RI. Document the extension point instead, so EP-01-T03's contract is provably stable under it.

The extension point

TokenValidator.validate(token: str) -> AuthenticatedClient is the only function in the RI that produces client identity for authorization purposes (emcp_bus.gateway.server.on_list_tools/on_call_tool both call get_access_token(), which is populated by the SDK's TokenVerifier protocol wrapping this same validator -- see emcp_bus.identity.token_verifier.MCPTokenVerifier). Everything downstream of it -- EntitlementManager.resolve, the PDP, the audit/correlation layer (EP-08) -- consumes only the resulting AuthenticatedClient.client_id. None of that code cares how the identity was established.

A production deployment adding mTLS or workload identity would:

Implement a second identity source honoring the same output contract:
```
class MTLSClientValidator:
    def validate(self, request: ...) -> AuthenticatedClient: ...
```
fed from the terminating proxy/load balancer's verified client certificate (a X-Client-Cert-CN style header from a trusted mTLS terminator, or the ASGI scope["extensions"]["tls"] a compliant server exposes) instead of a JWT. AuthenticatedClient.client_id would be the certificate's Subject/SAN, mapped 1:1 to a ClientRegistration.client_id exactly the way a JWT's azp is today -- no other model in the RI changes: ClientRegistration, ClientProfile, EntitlementManager, the PDP contract, and the audit schema are all already identity-mechanism agnostic.
Provide an alternate build_auth()-equivalent in emcp_bus/gateway/server.py that wires the new validator in instead of MCPTokenVerifier, gated by deployment config -- the same shape as today's build_auth() returning (None, None) for the "no OIDC configured" mode. A hybrid deployment (mTLS for one class of backend service, OAuth for interactive agents) is possible by making the TokenVerifier/equivalent selection per-listener rather than per-process, which the SDK's streamable_http_app(auth=..., token_verifier=...) parameters already support per app instance.
Cloud workload identity (SPIFFE/SPIRE, IRSA, Workload Identity, Managed Identity) is a variant of the same shape: the "verify a credential the caller cannot forge, and extract a stable identifier from it" contract is identical, only the credential format and verification library differ. None of it requires touching EntitlementManager, the PDP, the Fabric router (EP-06-T01), or the audit/correlation layer (EP-08-T02/T03).

What must not change

AuthenticatedClient shape (client_id, issuer, expires_at): every downstream consumer types against this, not against a JWT-specific claim set.
Fail-closed contract: any identity mechanism's failure path must raise (an AuthenticationError-shaped exception) rather than fall back to an unauthenticated or agent-declared identity (README.md §5, Axiom 2) -- this is the actual security property EP-01-T03's acceptance criteria test, and it is mechanism-independent.
No new privileged path for clientInfo: whatever the identity mechanism, only a value verified independently of the MCP request body may ever populate client_id.

Consequences

Confirmed for the RI: reduces EP-01's implementation effort without weakening any of README.md §19's desirable identity properties (all are satisfied by OAuth 2.1 client_credentials against Keycloak).
A production fork adding mTLS/workload identity changes exactly one module (identity/) and one wiring function (build_auth-equivalent); every other component in the RI (Entitlement Manager, PDP, Fabric, Registry, audit/correlation, approval) requires zero changes.

5.3 KiB Raw Blame History