GratefulAgents Starts on Loopback, and Exposure Should Follow the Deployment

GratefulAgents is a self-hosted control plane for running AI-agent workflows against repositories. The controller reconciles platform and trigger resources, exposes a dashboard, and can receive GitHub webhooks. A local Kind deployment binds the dashboard to 127.0.0.1:8090 and prints a command to retrieve the generated administrator password.

ZoomEye returned zero for app="GratefulAgents", zero for the product title, and zero for the HTTP body. The application query is selected and all three probes succeeded. The result is consistent with a product that starts as a local control plane, but it does not prove that every deployment remains on loopback or stays out of an index.

A local default is a useful control

Loopback binding prevents a dashboard from accepting connections from another host. It does not prevent a reverse proxy from publishing the application later. The deployment should record the intended interface, the proxy configuration, the authentication method, and the administrator recovery process.

The local setup uses a generated administrator password. That password should not remain the only recovery path. Store it in a secret manager, require an administrator login, and define how it is rotated. If an identity provider is added, map groups to roles and keep the recovery account under separate control.

Separate dashboard access from integration credentials

The project's connection-secret documentation says dashboard access alone does not grant permission to use a stored integration credential. That distinction should appear in the authorization model. A user who can view a run may not need the GitHub token that the run uses. A user who can create a project should not automatically receive every provider key.

Inventory repository tokens, GitHub App private keys, OAuth tokens, webhook secrets, Kubernetes credentials, and provider keys. Record which runtime and which project can use each one. Check the service account that launches the agent container and the permissions available to the repository checkout.

Test the deployment you actually run

For a local deployment, verify that the dashboard is reachable only from loopback and that no proxy route publishes it. For a Kubernetes deployment, inspect the Service, Ingress or Gateway, NetworkPolicy, TLS configuration, authentication mode, and administrator password retrieval path.

The zero external count is not an all-clear. Use DNS and proxy inventory to find a dashboard that the search engine did not index. Then test unauthenticated access, an ordinary user role, and an expired session.

Monitor the agent run boundary

Agent runs can read repositories, write patches, use network tools, and produce artifacts. Review the image, runtime profile, mounted files, network egress, and tool policy. A run should not inherit a broad human token when a scoped repository credential is enough.

If a dashboard or secret path was exposed, revoke provider keys, webhook secrets, and Kubernetes credentials, then review runs that used them. The useful evidence is the link between a project, a run, a credential, and the repository or model action.

References