Deploying and integrating ABR

Environment

Health check

GET /v1/healthz (no auth) returns { ok, name, version, repository, jobProvider, captureProvider, activeCaptures }.

Use /v1/healthz, not /healthz. On Cloud Run (and many L7 frontends) only paths under /v1/ are reliably forwarded to the container for external traffic; startup probes that hit the container port directly can still use either path.

Every JSON response also carries a Server: agent-browser-runtime/<version> (+repo) header.

Local (Docker Compose)

AGENT_RUNNER_TOKEN=$(openssl rand -base64 32) \
  STRATEGIES_DIR=$PWD/examples/demo-strategy \
  docker compose up --build

curl localhost:8080/v1/healthz
curl -H "authorization: Bearer $AGENT_RUNNER_TOKEN" \
  -H 'content-type: application/json' \
  -d '{}' localhost:8080/v1/probe   # expect 400 validation (auth OK)

Steel is exposed on :3000 (API + live view) and :9223 (CDP). Only the runtime should reach them in production.

Container image

ghcr.io/praveen-palanisamy/agent-browser-runtime:<version> bundles Node, Playwright's Chromium and the runtime CLI. Extend it with your strategies:

FROM ghcr.io/praveen-palanisamy/agent-browser-runtime:latest
COPY dist/strategies.js /app/strategies.js
CMD ["serve", "--strategies", "/app/strategies.js"]

The image pins playwright-core to the base image's Playwright version (ARG PLAYWRIGHT_VERSION), verified in CI by scripts/check-playwright-pin.mjs, so no browser download happens at runtime.

Cloud Run (runtime + Steel sidecar)

Layout

• One service, two containers: the runtime (ingress on :8080) and Steel on localhost:3000.
• Set STEEL_API_URL=http://127.0.0.1:3000, AGENT_RUNNER_CAPTURE_PROVIDER=steel, AGENT_RUNNER_JOB_PROVIDER=local.
• Interactive captures are stateful (a browser stays open while the user signs in). Prefer minScale=0 when idle and scale up for capture sessions, or pin minScale=maxScale=1 only while capture UX is in active use.
• Ingress must be public for the live view; every /v1/* API route (except health) is bearer-token protected. Store AGENT_RUNNER_TOKEN in Secret Manager and grant the Cloud Run service account roles/secretmanager.secretAccessor on that secret. Annotate the revision with run.googleapis.com/secrets.
• Steel sidecar: listen on HOST=0.0.0.0 (not localhost — Cloud Run probes need a reachable port), ≥ 2–4 GiB memory, in-memory volume at /app/.cache.

Images in Artifact Registry (not GHCR)

Cloud Run often fails to import GHCR images (ContainerImageImportFailed). Mirror Steel into your project's Artifact Registry and reference that:

podman pull --platform linux/amd64 ghcr.io/steel-dev/steel-browser-api:latest
# authenticate, then:
podman tag … REGION-docker.pkg.dev/PROJECT/REPO/steel-browser-api:latest
podman push …

Podman vs Docker auth

gcloud auth configure-docker only wires Docker's credential helper. With Podman, log in explicitly:

gcloud auth print-access-token | \
  podman login -u oauth2accesstoken --password-stdin REGION-docker.pkg.dev

(Verber Studio's scripts/lib/artifact-registry-auth.cjs does this automatically.)

CPU quota (important)

An always-on sidecar pair (minScale=1, 2 CPU runner + 2 CPU Steel ≈ 4 reserved CPUs) counts against regional Cloud Run CPU quota alongside every Gen2 Cloud Function in the same GCP project. That often blocks firebase deploy --only functions with:

Quota exceeded for total allowable CPU per project per region.

Recommended: host ABR in a dedicated GCP project (separate 20 vCPU regional pool). Orchestrators (e.g. Firebase Functions in another project) only need AGENT_RUNNER_URL + a shared AGENT_RUNNER_TOKEN. Peak need for one warm instance with the slim template (1+1 CPU, minScale=0) is ~2 vCPU — one interactive Steel OSS capture at a time.

Economical default: keep the agent runner at minScale=0 (or delete the service) until capture is needed; run headless posts/probes only when the service is scaled up. Prefer requesting a quota increase only after product traffic justifies it.

To free quota quickly:

gcloud run services delete SERVICE_NAME --region=REGION --quiet
# and/or prune unused revisions:
# (host app) npm run cleanup:cloudrun-revisions

Dedicated project layout (portable)

flowchart LR
  subgraph studio [App / Functions project]
    Fn[Orchestrator]
  end
  subgraph abr [ABR project]
    Runner[Cloud Run ABR + Steel]
    AR[Artifact Registry]
  end
  Fn -->|Bearer AGENT_RUNNER_TOKEN| Runner

Images, Secret Manager (AGENT_RUNNER_TOKEN), and Cloud Run live in the ABR project. The app project only stores the public URL and the same token value. Swap hosts later (Fly/Railway/GCE) by changing URL + DNS — protocol stays GET /v1/healthz and bearer-protected /v1/*.

Optional providers (Infisical path /abr for agents): STEEL_API_KEY (Steel Cloud), KERNEL_API_KEY / KERNEL_API_URL (Kernel microVMs when AGENT_RUNNER_*_PROVIDER=kernel).

Smoke test

curl -sS "$AGENT_RUNNER_URL/v1/healthz"
# authenticated probe should return 400 (missing fields), not 401:
curl -sS -X POST "$AGENT_RUNNER_URL/v1/probe" \
  -H "authorization: Bearer $AGENT_RUNNER_TOKEN" \
  -H 'content-type: application/json' -d '{}'

Scaling beyond one instance

• Move captures to Kernel (AGENT_RUNNER_CAPTURE_PROVIDER=kernel): each capture gets its own microVM with a live view URL, so instances become interchangeable.
• Or run several runner instances with sticky capture routing or an external registry that maps both the management ID and the distinct live-view token to the owning instance. The two credentials intentionally cannot be correlated by clients.

Ways to consume ABR

Releases are cut automatically from main; every semver tag publishes all three channels with provenance. Follow the Releases page or watch the changelog.