Skip to content

Agent Mesh

docs: Google APIs with a federated credential

Antonio Ojea25a1efde8b1c
diff --git a/site/content/docs/guides/google-cloud-credentials.md b/site/content/docs/guides/google-cloud-credentials.md
new file mode 100644
index 00000000..4426906b
--- /dev/null
+++ b/site/content/docs/guides/google-cloud-credentials.md
@@ -0,0 +1,185 @@
+---
+title: "Google APIs with a federated credential"
+linkTitle: "Google Cloud credentials"
+weight: 3
+description: "Give mesh members access to Google APIs through an egress node that exchanges the control plane's token at Google STS. No Google credential is stored anywhere."
+---
+
+A member of the mesh calls a Google API: Gemini on Vertex AI, an agent on
+Agent Engine, BigQuery, or Cloud Storage. This guide puts an `agentmesh-node`
+in front of the API as the enforcement point. Workload Identity Federation
+supplies the credential.
+
+The node authorizes each request on the caller, the method and the path.
+Then the node mints a short-lived token at the control plane, exchanges it
+at Google STS, and injects the Google access token. No service account key
+exists. The member holds no Google credential.
+
+How it works:
+
+```text
+member POST /mesh/<serving peer>/egress/aiplatform.googleapis.com/v1/.../gemini-2.5-flash:generateContent
+ X-Mesh-Authentication: ****** member's Biscuit>
+node authorizes: role grants egress://aiplatform.googleapis.com, POST, this path
+ POST <control plane>/sts/token -> ES256 JWT: iss=<control plane>, sub=<member>, aud=<WIF provider>
+ POST https://sts.googleapis.com/v1/token -> federated token
+ POST iamcredentials ...:generateAccessToken (impersonate) -> access token
+ POST https://aiplatform.googleapis.com/v1/.../gemini-2.5-flash:generateContent
+ Authorization: ****** token>
+```
+
+You need a control plane that Google can reach over HTTPS, at
+`/.well-known/openid-configuration` and `/jwks`. You need a node that
+serves the destination. You need a Google Cloud project where you can
+create a Workload Identity Pool and a service account. The [Cloud Run
+guide](/docs/guides/cloud-run/) gives you such a control plane.
+
+## 1. Trust the control plane in Google Cloud
+
+The control plane is an OIDC issuer. Its issuer URL must use `https`.
+Behind a TLS-terminating proxy such as Cloud Run, the control plane does
+not take the scheme from forwarded headers. Give it the public URL:
+`--external-url=https://<hub>.run.app` on `agentmesh-one`, or
+`--sts-issuer-url` on `agentmesh-control-plane`. Check the issuer with:
+
+```bash
+curl -sS https://<hub>.run.app/.well-known/openid-configuration | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["issuer"], d["jwks_uri"])'
+```
+
+Create the pool, the provider and the service account. The allowed
+audience of the provider is its own resource name. The control plane puts
+that name in `aud`.
+
+```bash
+PROJECT=<project>
+PN=$(gcloud projects describe $PROJECT --format='value(projectNumber)')
+HUB=https://<hub>.run.app
+gcloud services enable sts.googleapis.com iamcredentials.googleapis.com aiplatform.googleapis.com --project $PROJECT
+
+gcloud iam workload-identity-pools create agentmesh --project $PROJECT --location global
+AUD="//iam.googleapis.com/projects/$PN/locations/global/workloadIdentityPools/agentmesh/providers/hub-cp"
+gcloud iam workload-identity-pools providers create-oidc hub-cp --project $PROJECT --location global \
+ --workload-identity-pool agentmesh --issuer-uri "$HUB" --allowed-audiences "$AUD" \
+ --attribute-mapping "google.subject=assertion.sub,attribute.actor=assertion.act.sub"
+
+gcloud iam service-accounts create mesh-gemini --project $PROJECT
+SA=mesh-gemini@$PROJECT.iam.gserviceaccount.com
+gcloud projects add-iam-policy-binding $PROJECT --member serviceAccount:$SA --role roles/aiplatform.user --condition None
+gcloud iam service-accounts add-iam-policy-binding $SA --project $PROJECT --role roles/iam.workloadIdentityUser \
+ --member "principalSet://iam.googleapis.com/projects/$PN/locations/global/workloadIdentityPools/agentmesh/*"
+```
+
+The last binding lets every identity that the control plane authenticates
+impersonate the service account. The mesh policy decides which members
+reach the destination, and with which methods and paths. To narrow the
+binding on the Google side as well, map a claim to an attribute
+(`attribute.role=assertion.mesh_roles[0]`). Then bind
+`principalSet://.../attribute.role/<role>` instead of `*`.
+
+The `sub` claim in the token is the principal of the member: a SPIFFE ID
+for a workload, or `user:<email>` for a person. The `act.sub` claim is the
+peer ID of the node that served the request. Google STS maps `sub` to
+`google.subject`; the attribute mapping above keeps `act.sub` as
+`attribute.actor`.
+
+## 2. Write the policy
+
+The `egress` entry names the destination, the nodes that serve it, and the
+federation. The role lists the destination and narrows it to the method
+and the path:
+
+```json
+{
+ "roles": [
+ {"name": "mesh:role:node", "allowed_services": ["*"], "allowed_publish": ["*"], "allowed_targets": ["*"], "allowed_labels": ["*"]},
+ {"name": "mesh:role:router", "allowed_services": ["*"], "allowed_targets": ["*"]},
+ {"name": "sandbox",
+ "allowed_services": ["system://mesh.catalog", "egress://aiplatform.googleapis.com"],
+ "allowed_targets": ["*"],
+ "http": [{"service": "egress://aiplatform.googleapis.com", "methods": ["POST"],
+ "paths": ["/v1/projects/<project>/locations/us-central1/publishers/google/models/*"]}]}
+ ],
+ "bindings": [
+ {"role": "sandbox", "members": ["user:spiffe://openshell.local/openshell/sandbox/*"]}
+ ],
+ "egress": [
+ {"name": "aiplatform.googleapis.com",
+ "served_by": ["site=office"],
+ "broker": {"oidc_federation": {
+ "token_endpoint": "https://sts.googleapis.com/v1/token",
+ "audience": "//iam.googleapis.com/projects/<project number>/locations/global/workloadIdentityPools/agentmesh/providers/hub-cp",
+ "impersonate": "mesh-gemini@<project>.iam.gserviceaccount.com",
+ "scopes": ["https://www.googleapis.com/auth/cloud-platform"]}}}
+ ]
+}
+```
+
+- `served_by` selects the nodes by role or label. Each selected node logs
+ `[Egress] Serving egress://aiplatform.googleapis.com` when it takes the
+ assignment.
+- APIs that do not accept the federated principal directly need
+ `impersonate`. Vertex AI is one of them.
+- `paths` is a prefix match under the destination. The grant above allows
+ every publisher model in one region. It allows no other path on the host.
+- One hostname can serve more than one product. `aiplatform.googleapis.com`
+ serves Gemini (`publishers/google/models/*`) and Agent Engine
+ (`reasoningEngines/*`). A role has one `http` entry per service. To grant
+ both, put both paths in the `paths` list of that entry.
+
+Push the policy with `POST /policies` and the admin token, or through the
+Cloud Run secret. The nodes log the assignment within their sync interval.
+
+## 3. Call the API
+
+Send the request from the node of a member, with the credential of that
+member. Address the serving node by peer ID:
+
+```bash
+PEER=<peer id of a node that serves the destination>
+curl -sS -H "X-Mesh-Authentication: ******" -H "Content-Type: application/json" \
+ "http://127.0.0.1:8080/mesh/$PEER/egress/aiplatform.googleapis.com/v1/projects/<project>/locations/us-central1/publishers/google/models/gemini-2.5-flash:generateContent" \
+ -d '{"contents":[{"role":"user","parts":[{"text":"Answer with one word: what is 2+3?"}]}]}'
+```
+
+Gemini returns the answer. A `GET` on the same path returns `403` with
+`Proxy-Status: agentmesh-node; error=http_request_denied`. A `POST` to
+`/v1/projects/<project>/locations/us-central1/endpoints` returns the same
+`403`.
+
+The serving node writes one line per decision:
+
+```text
+Audit Traceability {"target": "egress://aiplatform.googleapis.com", "method": "POST",
+ "path": "/v1/projects/<project>/locations/us-central1/publishers/google/models/gemini-2.5-flash:generateContent",
+ "host": "aiplatform.googleapis.com", "port": 443, "decision": "allow",
+ "user": "spiffe://openshell.local/openshell/sandbox/0e4e35e4-…", "role": "sandbox", "actor_node": "12D3KooW…"}
+Audit Traceability {"target": "egress://aiplatform.googleapis.com", "method": "GET", ..., "decision": "deny",
+ "reason": "biscuit: denied by no matching policies"}
+```
+
+The control plane logs the `/sts/token` request. The Google audit log shows
+one `GenerateAccessToken` on the service account for each new token, with
+`lifetime: 300s`. That entry names the service account. The identity of
+the member is in the audit line of the node.
+
+## 4. From a sandbox
+
+An OpenShell sandbox or a Substrate actor reaches the destination through
+its edge node (see [OpenShell or Substrate](/docs/guides/openshell-and-substrate/)).
+The enforcement point of the runtime injects the Biscuit of the member. The
+edge node forwards the request to the serving node. The allowlist of the
+runtime must permit the path `/mesh/**` on the edge node. The examples
+permit it. `make gemini NAME=<sandbox> GCP_PROJECT=<project>` in each
+example runs the three requests above from inside the sandbox or the actor.
+
+## Limits
+
+- The node caches the Google token per caller credential. An entry is valid
+ for at most the lifetime of the token that the node minted at the control
+ plane, 5 minutes by default. Then the node performs the exchange again. A
+ change of the role of a member takes effect at the next mesh credential of
+ that member, as for any destination.
+- A `TaskAuthorizationRule` on the Biscuit of the caller narrows `scopes`,
+ methods and paths further. It cannot widen them.
+- The destination is one hostname. See [Egress
+ destinations](/docs/guides/egress-destinations/) for the general limits.