| name | deploy-demo |
| description | Deploy the full LLM observability stack (Langfuse + ClickHouse + LibreChat + demo apps) end to end and verify it is demo-ready. Use when the user says "deploy this demo", "set this up", "get the stack running", or when a partial/failed deployment needs to be brought to a healthy state. |
Deploy the Demo
The canonical, non-interactive deploy runbook is AGENTS.md at the
repo root. Read it and follow it exactly — this skill only adds the operating posture
around it.
Procedure
-
Check prerequisites (Docker running, docker compose, jq) using the commands
in AGENTS.md > Prerequisites. Fix what you can (brew install jq / apt-get install -y jq);
only stop if Docker itself is unavailable.
-
Resolve the one required secret. The only secret needed is an Anthropic API key.
Check .env first (grep '^ANTHROPIC_API_KEY=sk' .env), then the environment.
If absent, ask the user for it — never invent one, never proceed without it.
-
Deploy:
ANTHROPIC_API_KEY=sk-ant-... ./setup.sh --seed
- Idempotent — re-running is the correct first response to most failures.
- First run takes ~5 minutes (image pulls);
--seed adds a few minutes of traces.
- Do not use
docker compose up for initial deployment; setup.sh provisions
secrets and API-side config that compose alone does not.
-
Verify with every check in AGENTS.md > Verify (health endpoints, LLM connection,
5 LibreChat agents, traces, evaluators) or ./setup.sh --status — all "Demo
Readiness" lines must be ✓. Do not declare success on a partial pass.
-
Report the URLs and credentials exactly as listed in AGENTS.md > "Report to the
user when done", and suggest the natural next steps: present it (run-demo skill,
docs/SA_FIELD_GUIDE.md) or explore it
(docs/USER_JOURNEY.md).
If something fails
Use the AGENTS.md > Troubleshooting table first, then the troubleshoot skill for
deeper diagnosis. Re-run ./setup.sh after any fix. Never run ./scripts/reset.sh
(destructive) without explicit user confirmation.
