overbae/), the React Console (frontend/), the Python SDK and CLI (overmind/), and the Modal workers for training and serving. The hosted product at overmindlab.ai runs this same code.
Requirements
Docker with Compose, uv for anything Python outside the containers, and Bun for the Console. Training and serving need a Modal or Baseten account and are optional.Start the stack
http://localhost:8000 with hot reload, four Celery workers and beat, and Grafana on http://localhost:3001. The API container runs migrations and seeds the managed evaluators on every start. The Console is a Vite app served separately:
frontend/.env.development already points the Console at http://localhost:8000. Sign-in is Clerk: the frontend env ships a development publishable key and the API a matching test secret (CLERK_API_SECRET_KEY, CLERK_AUTHORIZED_PARTIES, defaulting to http://localhost:5173); for a deployment of your own, create a Clerk application and set both. Then create a project and follow the Quickstart — with overmind init --env local, so the CLI and MCP point at your API.
For a demo workspace with thirty days of traffic, datasets, runs and served models, run the seed:
Environment reference
.env.example is the operator contract. Every variable is read by docker-compose.yml or overbae/settings.py.
Required
Postgres and Redis default to the Compose services (
POSTGRES_*, CELERY_BROKER_URL, CELERY_RESULT_BACKEND).
The Data Workshop agent
The dataset agent runs on the first key it finds, in this order:CURSOR_API_KEY (a Cursor Composer session), OPENROUTER_API_KEY, then OPENAI_API_KEY, ANTHROPIC_API_KEY, or GEMINI_API_KEY with a native tool-calling loop. Both engines take the same tools and emit the same events. With none set the workshop page opens and the agent reports that it is not configured. OPENAI_API_KEY also serves embeddings for statistical evaluators.
Training and serving
Without a backend, the Console’s Training and Inference pages open, the catalog lists, and a launch fails at submission with the reason.
Billing modes
LeaveSTRIPE_SECRET_KEY empty and the platform runs in metering mode: every charge still lands on the credits ledger and shows under Settings, but there is no remaining-credit cap, no 402, no Free/Pro quota, and no signup grant. Set STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, STRIPE_PRO_PRICE_ID, and STRIPE_CREDIT_PRICE_ID to turn on the hosted product’s plans, top-ups, and the credit gate; point Stripe’s invoice.paid webhook at POST /api/billing/webhook/.
Optional
GRAFANA_ADMIN_PASSWORD and AWS_PROFILE for the Grafana service; SLACK_FEEDBACK_WEBHOOK_URL to forward Console feedback to a channel; TOGETHER_API_KEY for the retired Together backend that existing job rows still render.
Topology
The API is Django plus DRF under uvicorn; MCP is mounted on the same ASGI app at/api/mcp/. Celery runs four workers over five queues: control (default, orchestration and beat), io and io_traces (one threads worker; trace scoring rides io_traces), batch (landing datasets, uploads) and interactive (dataset runs and agent turns). Workers restart on any .py change under overbae/, so a running agent turn dies when you edit backend code beneath it. Every dataset frame lives under MEDIA_ROOT/datasets/<id>/cells/ as Parquet on the app_data volume.
Developing on it
AGENTS.md at the repo root is the playbook a coding agent reads before touching a subsystem; .claude/skills/ holds the subsystem maps. Pull requests go against main; the SDK publishes to PyPI on overmind-v* tags.