Verify the install

The bootstrap script runs a smoke test, but it is useful to know what to check yourself when something is off.

All services up

docker compose ps

Every service should be Up with (healthy) next to it (except the minio-init one-shot, which exits cleanly after creating the bucket).

Gateway

curl -s http://localhost:4000/health/liveliness
# → "I'm alive!"
curl -s -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  http://localhost:4000/v1/models | jq '.data[].id'
# → list of registered models

If /v1/models is empty, you probably ran bootstrap with --backend skip. Add one with ./scripts/add-model.sh.

Chat

curl -s http://localhost:3080/api/health
# → OK

Open http://localhost:3080 in a browser. The dropdown should populate within a couple of seconds. If it stays on loading..., the chat backend cannot reach the gateway — check docker compose logs chat for the exact error.

Console

curl -s http://localhost:3001/_health
# → ok

Open http://localhost:3001. If you are not signed in to the chat in the same browser, you will be redirected to /unauthorized. Sign in at The chat first.

Langfuse

Open http://localhost:3000. Log in with the admin email + password that bootstrap printed. Click Traces — after you have sent at least one chat message, you should see it here within a few seconds.

Grafana

Open http://localhost:3030 and log in (GRAFANA_ADMIN_USER / GRAFANA_ADMIN_PASSWORD). The Gateway Overview dashboard auto-loads from monitoring/grafana/dashboards/.

End-to-end smoke test

./scripts/smoke-test.sh         # quick: just gateway ↔ AI provider
./scripts/e2e-smoke-test.sh     # full: chat → gateway → AI provider → trace

The end-to-end version registers a user, sends a chat, and asserts the Langfuse trace landed with the expected hardware_id. See E2E_* env vars for what it needs in .env.

If something is broken

See Troubleshooting for the common failure modes.