Self-hosted durable control plane for LangGraph agents.
agentctl is planned as a production-oriented OSS platform for running long-lived
agent workflows with Temporal-backed execution, approval gates, operational
visibility, model/cost telemetry, and eval gates.
This repository contains the M1-M4 platform foundation:
- Java/Spring API with
/api/health, run projections, approval projections, audit timeline APIs, Flyway migrations, and Postgres wiring - Temporal Java run workflow with approval signals and projection activities
- Python
uvFastAPI agent worker with LangGraph/LangChain support-ticket draft step - Next.js dashboard shell
- Docker Compose wiring with healthchecks for API, worker, web, Temporal, Postgres, and MinIO by default, plus opt-in agent and observability profiles
- Platform PRD
- M1/M2 Foundation PRD - superseded
- M4 Agent Step Protocol PRD
- M5 Support Ticket V1 PRD
- M5 Agent Tool Execution Boundary PRD
- Support Ticket Agent PRD
- Incident Agent PRD
- GitHub Ops Ticket-to-PR Agent PRD
Copy .env.example if you want to override local defaults, then run:
docker compose up --buildThe stack will start the Java/Spring control plane, Temporal, Postgres, MinIO, Next.js dashboard, and Java Temporal worker.
Optional profiles:
docker compose --profile agents up --build
docker compose --profile observability up --build
docker compose --profile agents --profile observability up --buildThe agents profile starts the Python LangGraph agent worker and Ollama.
No Gemma model tag is hardcoded yet; the exact official tag must be verified
before making it a default. The observability profile starts OTel Collector,
Prometheus, Tempo, and Grafana.
Basic API smoke checks for the durable support-ticket draft flow:
docker compose --profile agents up --build
curl -fsS http://localhost:8080/api/health
curl -fsS -X POST http://localhost:8080/api/runs \
-H 'Content-Type: application/json' \
-d '{"agentId":"support-ticket","input":"Create a ticket for login failure"}'
curl -fsS http://localhost:8080/api/runs
curl -fsS http://localhost:8080/api/approvals/pendingRun creation starts a Temporal workflow. The Java worker calls the Python
POST /v1/agent-steps runtime, stores the agent step/model/tool projections,
creates a pending approval from the support-ticket draft response, waits for an
approval signal, and then marks the run COMPLETED or REJECTED. No fake
ticket or GitHub mutation happens in M4.
cd services/api && mvn test
cd services/agent-worker && uv run python -m unittest discover -s tests
cd services/web && npm run test:contract && npm run build
node scripts/check-compose-contract.mjs
docker compose configApache-2.0