Vast.ai MCP Server

An MCP (Model Context Protocol) server that exposes Vast.ai cloud operations as tools: list GPUs, search offers, create volumes, create instances, and view billing.

Setup

Installed as a package via pip (recommended):

pip install vastai-mcp          # stdio only
pip install "vastai-mcp[http]"  # adds the HTTP transport's deps (uvicorn, starlette)
export VAST_API_KEY=your_api_key_here

Or from source, for local development:

pip install -e ".[http]"

Get an API key at https://cloud.vast.ai (Settings -> API Keys).

Run (stdio, for MCP clients)

Installing the package puts a vastai-mcp command on your PATH:

vastai-mcp

Claude Desktop config example

{
  "mcpServers": {
    "vastai": {
      "command": "vastai-mcp",
      "env": { "VAST_API_KEY": "your_api_key" }
    }
  }
}

(See mcp.json.example in this repo for a copy-pasteable template — copy it to mcp.json and fill in your key; mcp.json itself is gitignored since it holds a real secret.)

Run (HTTP, for remote/multi-user deployment)

vastai-mcp-http

The HTTP server holds no Vast.ai key of its own. Every request to POST /mcp must include the caller's own key in an X-Vast-Api-Key header — that key is used only for that request, so rentals and billing land on the caller's Vast.ai account, not the server operator's. Requests without the header get a 401. GET /health needs no auth.

Deploy behind HTTPS (e.g. Caddy/nginx/Cloudflare in front) — remote MCP clients such as ChatGPT connectors require TLS.

Deploy (Docker)

docker build -t vastai-mcp-http .
docker run --rm -p 8000:8000 vastai-mcp-http
curl http://localhost:8000/health

This does not need publishing to PyPI, or even the pip-packaging machinery at all — the image just installs the runtime deps from requirements.txt and runs python -m vastai_mcp.server_http directly (no console-script entry point involved inside the container). Push the built image to your cloud provider's registry (or build straight from this repo if the provider supports that, e.g. Fly.io, Railway, Render), and put a TLS-terminating proxy or the platform's built-in HTTPS in front of port 8000 before pointing a ChatGPT connector at it.

Tools

Typical workflow

1. list_gpus to see what's available.
2. search_offers(gpu_name="RTX 4090", max_price=1.0, limit=10) to find a machine.
3. billing_summary to check current spend.
4. create_instance(offer_id=123, max_hourly_price=1.0, volume={"size_gb": 100, "mount_path": "/data"}) to launch (refuses if the offer's live price exceeds max_hourly_price).