Repository navigation
feat: external OIDC tokens for self-hosted instances #5090
Description
Is your feature request related to a problem? Please describe.
v4.8.3 added auth.getOIDCToken() and @trigger.dev/sdk/oidc/aws, but only Trigger.dev Cloud can mint the tokens. getExternalOidcController() returns null unless isCloud() is true, and the minter is loaded from the closed @triggerdotdev/plugins/oidc module. On a self-hosted instance POST /api/v1/auth/oidc/token returns 503. The open-source repo already has the rest: the OIDC_ISSUER_URL issuer, the /.well-known/jwks.json and discovery routes, the signing key service, run-token verification and the ExternalOidcPlugin interface.
Self-hosted operators are often the ones with the strictest security requirements (FedRAMP, ITAR, HIPAA). A task can only get its credentials from the environment's variables. Every run in an environment receives the same static secrets: database and API admin credentials, cloud keys and vendor tokens. Any task, or any dependency bundled into the image, can read all of them. OIDC would let a task exchange a short-lived token, valid only for one audience, for credentials scoped to that task: AWS roles through STS, token exchange with Keycloak or another IdP, Vault, or GCP and Azure workload identity. The static secrets could then leave the environment.
Describe the solution you'd like to see
- Enable external OIDC on self-hosted instances when
OIDC_ISSUER_URLis set and run tokens are on, with an open-source minter. A self-hosted operator can already provide that minter through the existing plugin interface if only theisCloud()gate is removed. - Document the token's claim set. Include
run_id,task_id,attemptanddeployment_idalongsidesub, so a verifier can authorize per task and attribute activity to a run. - Optionally, let operators add claims, such as AWS session tags (
https://aws.amazon.com/tags). That would let a multi-tenant application drive attribute-based access control from data recorded when the run was triggered.
Describe alternate solutions
- Workload identity on run pods (IRSA, EKS Pod Identity): the identity belongs to the worker group's service account, so every task in the environment gets the same role. It is AWS-only, and it doesn't identify the run or task.
- Sending the internal run token (
run+jwt) to our own exchange service: its signing keys aren't published, and its audience istrigger:apiwithadminscope. The receiving service ends up holding a Trigger API credential for the whole environment. - Shipping our own
@triggerdotdev/plugins/oidcmodule: theisCloud()check still blocks it.
Additional information
Related: #1851 (OIDC on Cloud runners). I'm happy to help with a PR.
Activity