Start

Credentials map

One table for “what do I paste where?” — Admin login, Agent Protocol clients, runners, and connector secrets are different trust domains.

Credential	Where it lives	Who uses it	Admin login?
API key with `admin`	`auth.keys` in `langgraph.json` / Helm	Operators, curl to `/admin-api`	Yes
`auth.admin_keys`	Same config object	Break-glass Admin only	Yes (Admin only)
API key read/write	`auth.keys`	App backends (Agent Protocol)	No
JWT with `admin` claim	Your IdP	SSO operators	Yes
`RUNNER_TOKEN_*`	Env on control plane + runners	Runner ↔ plane gRPC/HTTP	No
`RUNNER_TENANTS_*`	Env on control plane	Tenant allow-list per runner kind	No
Connector `${ENV}` / `secret_ref`	Secrets / env / Vault	Plane vends run-bound sessions	No
Postgres / Redis URLs	Env / Secret	Control plane process	No

How to implement (Supported)

Mint a long random operator key; put it in a secret store; reference via env substitution in config if you use ${VAR}.
Put admin on that key (or use admin_keys).
Issue separate read/write keys per app / tenant — never share the Admin key with product backends.
Set runner tokens per kind; when client auth is on, also set RUNNER_TENANTS_* (fail-closed otherwise).
Keep connector secrets out of git; use ${NAME} or auth.secret_ref in connector YAML (see Secrets).

Rotating Admin — deploy new key in config/secret, restart/reload plane, update operator vault, revoke old.
Leaked read/write key — can create runs in its tenant; cannot open Admin if it lacks admin.
Leaked runner token — can pull jobs for allowed kinds/tenants; still not Admin.