Skip to main content

Caller identity and OIDC

Mecatl can require an OIDC bearer token on every gRPC and HTTP request and use the verified identity to isolate application resources. This is more than request authentication: sessions, schedules, teams, memory entries, and persisted child sessions are owned by the caller that created them.

Availability

Caller identity is an opt-in server feature for mecated and mecak8s. It protects both wire surfaces with the same validator and ownership rules. mecatui can either send an operator-supplied static bearer with --auth-token, or enroll a remote target with mecatui login and obtain, validate, and refresh its own OIDC credential. With no explicit credential, an existing saved enrollment is preserved; any clean enrollment miss is attempted credential-free and the server decides whether caller authentication is required. Use connect-only --anonymous to intentionally bypass saved credentials when no static token is selected. A token supplied through --auth-token or MECATL_AUTH_TOKEN wins if both are present; --anonymous has no environment equivalent. Remote targets use verified TLS automatically; a static bearer is never sent over explicit non-loopback plaintext, and saved OIDC authentication always requires verified TLS. Those are distinct client modes; the server still only validates the bearer presented on each request.

With OIDC disabled, the server preserves the single-shared-deployment behavior: there is no caller subject and anyone who can reach the API is treated as the same caller. Do not expose that posture as a multi-user endpoint.

What identity means

The durable owner is the verified (issuer, subject) pair from the token:

{
"owner": {
"issuer": "https://idp.example.com",
"subject": "opaque-user-id",
"grant_type": "user",
"name": "Alice"
}
}

issuer and subject are the identity key. name is only a cosmetic snapshot of the token's name claim and can change later. The owner is written at resource creation from the verified request context, never from a request-body field.

With ownership enforcement enabled:

  • callers list only their own sessions, schedules, teams, and fires;
  • a caller cannot prompt, resume, cancel, fork, or delete another caller's session or persisted child;
  • live event streams and durable event-log reads enforce the same owner check;
  • identical logical keys, such as a memory key or schedule name, can exist for different owners without collision; and
  • foreign or ownerless resources look like not found, preventing resource enumeration.

Subagents, Parallel branches, Team members, and scheduled fires inherit the appropriate owner. A fork inherits the source session's owner and still checks that the caller owns the source. Resources created before identity was enabled are ownerless and become unavailable once enforcement is turned on; there is no automatic first-reader adoption.

Ownership is not the same as event actor attribution. A scheduled system action can act on a resource while the resource remains owned by its user. The actor stamp is durable event-log metadata, not a client-visible authorization field.

Configure OIDC

The essential server settings are:

FlagPurpose
--oidc-issuerHTTPS issuer URL, compared byte-for-byte with the token's iss; setting it enables caller identity.
--oidc-audienceRequired accepted aud value; prevents accepting tokens minted for another service.
--oidc-jwks-uriOptional pinned JWKS endpoint; otherwise discovery obtains it from the issuer.
--oidc-max-jwks-stalenessMaximum age of a last-good signing-key cache during an IdP outage. 0 deliberately removes the bound.
--oidc-resourceOptional canonical external HTTPS protected-resource URL (RFC 9728).
--oidc-client-idOptional public mecatui client-registration hint; a mecatl extension, not an RFC 9728 field.
--oidc-scopesOptional CSV scope list advertised as scopes_supported. It is a narrow operator-configured public-client request allowlist, not server authorization policy. Discovered login requests an advertised list exactly; when omitted, it requests the fixed openid,profile,offline_access baseline.

A typical configuration uses the issuer and audience together:

mecated serve \
--oidc-issuer https://idp.example.com/realms/operators \
--oidc-audience mecatl \
--workspace /srv/mecatl/workspace

--oidc-jwks-uri is useful for air-gapped or pinned-key deployments. Use a real HTTPS IdP in production. The validator rejects an HTTP issuer and refuses JWKS endpoints resolving to private, loopback, link-local, or metadata addresses. Initial configuration or key-fetch failure fails closed rather than starting an unauthenticated service.

The production validator is a delegated, maintained OIDC/JWT library; Mecatl does not hand-roll signature verification. A successful JWKS fetch is cached in process. During a short IdP outage, the last-good keys may continue to work until the staleness limit; after that the service returns 503 until it can refresh. A malformed, expired, wrong-issuer, or wrong-audience token returns 401. The cache is not persisted, so a restarted process fetches keys again.

Protected-resource discovery (RFC 9728)

A deployment may publish a canonical HTTPS resource profile with --oidc-resource, --oidc-client-id, and optional --oidc-scopes. RFC 9728 fields (resource, authorization_servers, and scopes_supported) are kept separate from mecatl extensions (com.stacklok.mecatl.audience and com.stacklok.mecatl.client_id). The existing issuer and audience remain the OIDC validator's source of truth. Discovery is anonymous HTTPS bootstrap and is separate from authenticated gRPC transport; it never inherits private-issuer CA exceptions. ToolHive's metadata/networking code is provenance for the client implementation, not a runtime dependency of the engine. scopes_supported is a narrow operator-configured public-client request allowlist, not server authorization policy: discovered login requests an advertised scopes_supported set exactly, or the fixed openid,profile,offline_access baseline when the member is omitted, and never expands a saved enrollment from later metadata. Discovery rejects --scopes; administrators configure oidc.scopes for other scopes. Explicit identity login retains its --scopes override. The direct configured well-known endpoint serves metadata; subordinate API 401 responses remain generic Bearer because their route and untrusted Host cannot prove the configured resource identity. With a published profile, mecatui login ADDRESS discovers and confirms the public tuple from a hostname or HTTPS resource URL. The explicit mecatui login --issuer ... form remains compatible for deployments without a profile and for private issuers.

For a static bearer, obtain a token through your identity provider and pass it to mecatui or another client. Keep it out of shell history where possible. This mode does not refresh the token:

export MECATL_AUTH_TOKEN="$(your-oidc-cli print-access-token)"
mecatui connect 127.0.0.1:8080 --auth-token "$MECATL_AUTH_TOKEN"

For managed remote OIDC, enroll once and then connect without putting a bearer on the command line:

mecatui login mecated.example.internal:443 \
--issuer https://idp.example.internal \
--client-id mecatui --audience mecatl \
--tls-ca /path/to/issuer-ca.pem --private-issuer
mecatui connect mecated.example.internal:443 \
--tls --tls-ca /path/to/server-ca.pem

The issuer CA bundle path/reference—not the CA contents—verifies issuer endpoints and is saved with the enrollment; the optional connect CA independently verifies the gRPC server. Managed credentials live in a keyring-wrapped encrypted store and refresh on later token demand. connect never opens a browser implicitly.

For non-loopback connections, mecatui refuses to send a bearer over cleartext. Use TLS (and connect --tls-ca for a private server CA). A gRPC dial can succeed before the first request is authenticated; verify that an unauthenticated first request fails before producing a model response.

Credential-free private-network connection

--anonymous means “bypass saved OIDC enrollment and send no bearer”; it is not a login mode and creates no saved state. A clean enrollment miss already gets a credential-free attempt. For remote use, verified TLS remains the default. If a Tailscale deployment deliberately uses the tailnet as shared authority and transport, explicitly select plaintext:

mecatui connect ozzllama:9080 --tls=false

Add --anonymous only to override an existing saved enrollment.

Bind mecated to one concrete Tailscale address rather than a wildcard, do not enable Funnel, and make tailnet ACLs the load-bearing reachability boundary. Use a dedicated server workspace, run with the least OS authority that can access it, and configure a restrictive rate limit. The server's non-loopback no-caller-authentication warning is expected: ordinary TLS does not authenticate callers. Never infer this posture from a private IP, hostname, or Tailscale-like name. A non-loopback client's current directory is not the workspace; the server selects and governs its workspace.

For a non-loopback target, the server's listener policy selects the workspace or no-FS profile: mecatui sends no local cwd and rejects --workspace. A loopback connection may still select a server-host path. In Kubernetes, isolate tenant workspaces separately: ownership does not make a shared pod filesystem a security boundary.

Management is separate from authentication

A valid OIDC token does not automatically grant storage-management authority. Remote cleanup, migration, and retention management require an explicit operator-tier allowlist of exact issuer/subject pairs:

storage_management:
version: 1
principals:
- issuer: https://idp.example.com/realms/operators
subject: storage-admin

An absent or empty list fails closed. Project settings, request owner claims, display names, grant types, and system-principal status cannot grant this authority. Destructive management also requires a working cross-process session lease; management permission alone is not a single-writer proof.

This distinction is intentional:

  • authentication answers “who made this request?”;
  • ownership answers “whose resource is this?”;
  • management authority answers “may this operator perform storage-wide maintenance?”; and
  • session leases answer “does this process exclusively own the mutation?”

Deployment limitations

  • OIDC configuration is supported on the server roots, not as an in-process token issuer. An embedding must provide its own authenticated boundary and owner propagation if it needs multi-user isolation.
  • The static-bearer/unmanaged mecatui path obtains no token and performs no refresh. Managed remote OIDC instead uses mecatui login and refreshes the encrypted, target-bound credential on later application token demand.
  • mecatequi and scheduled/headless jobs do not open a browser. Pre-provision a short-lived identity or use the deployment's non-interactive credential path.
  • Raw remote store and memory drivers are trusted infrastructure; caller ownership is enforced at the Mecatl service edge, not by an unauthenticated driver endpoint.
  • A workspace path is not an ownership boundary. Use separate namespaces, containers, or operating-system permissions when tenants must not share files.
  • Turning identity on is an isolation cutover. Export or migrate resources from an ownerless deployment before enabling it; there is no automatic adoption.
  • Bearer authentication and OIDC caller identity are different modes. A static --auth-token authenticates possession of one shared secret but cannot provide per-caller ownership.

For the full Kubernetes overlay, JWKS cache behavior, and troubleshooting steps, see Caller identity and ownership isolation.

Next steps