Quickstart

Self-host AI Cost Gate

Buy Pro once, clone the private runtime, and put a local cost gate in front of your agents — attribution, budgets, and auto-stop without shipping prompts to a metering SaaS.

AI Cost Gate (CostBrake) is a self-hosted, metadata-only LLM cost gateway: Docker Compose brings up a FastAPI gateway (default :8000) and a dashboard (default :3000). Point OpenAI-compatible clients at the gateway with provider/model ids and project/agent metadata.

1. 1. Buy Pro and accept the GitHub invite

   Complete the one-time Pro checkout and enter the GitHub username that should receive private collaborator access to CostBrake/ai-cost-gate. Accept the invite before cloning.

   Buy Pro source

2. 2. Clone and start with Docker Compose

   Clone the private Pro repo and start both services. SQLite is the default backend — Postgres is optional via ACG_PG_DATABASE_URL.

   git clone git@github.com:CostBrake/ai-cost-gate.git
   cd ai-cost-gate
   cp .env.example .env
   # Edit .env if needed — provider keys can also be set in the dashboard after boot
   docker compose -f infra/docker-compose.yml up -d

   Gateway → http://localhost:8000 · Dashboard → http://localhost:3000

3. 3. Confirm the gateway is healthy

   Hit the health endpoint, then open the dashboard to add provider keys and review requests.

   curl http://localhost:8000/health
   # -> {"status":"ok"}

4. 4. Point agents at the local OpenAI-compatible base URL

   Set client base_url to http://localhost:8000/v1. Choose upstream with model like openai/gpt-4o-mini or anthropic/claude-3-5-sonnet-latest, and pass metadata.project / metadata.agent for attribution.

   curl -s http://localhost:8000/v1/chat/completions \
     -H "Content-Type: application/json" \
     -H "Authorization: Bearer $ACG_API_KEY" \
     -d '{
       "model": "openai/gpt-4o-mini",
       "messages": [{"role": "user", "content": "ping"}],
       "metadata": {"project": "demo", "agent": "quickstart"}
     }'

   Use OpenRouter-style model ids (provider/model). Upstream provider keys are configured in the dashboard — do not send them on /v1 requests. If ACG_API_KEY is unset, gateway auth on /v1 is skipped.

5. 5. Set a budget and verify auto-stop

   Create a daily/weekly/monthly budget for a noisy agent. When the critical threshold is hit, in-scope requests return HTTP 429 until the period rolls over, the budget is raised, or the event is acknowledged.