Overview
The Massed Compute MCP server lets AI assistants like Claude, Cursor, Codex, and ChatGPT manage your Massed Compute account on your behalf: GPU inventory, instances, billing, SSH keys, and setup recipes.
No SDK and no custom code. Your AI client speaks the Model Context Protocol, and our server translates. Authenticate with an API token or an OAuth 2.1 grant (the same header either way).
What you can do with it
- "What's running right now?" List instances, regions, and hourly cost without leaving your editor.
- "Spin up an H100 for this notebook." Pick a GPU for the workload, deploy it, wait for boot, and get the SSH command.
- "Set up Jupyter / Ollama / vLLM for me." Search published setup recipes and follow the encoded image + SKU + steps.
- "What's the cheapest way to fine-tune Llama 3 70B?" Get a recommendation from live inventory and pricing.
Server details
| Field | Value |
|---|---|
| Endpoint | https://vm.massedcompute.com/api/mcp |
| Auth | Authorization: Bearer <credential> on every JSON-RPC POST |
| API token | From API Settings — user_uuid.api_key |
| OAuth token | Issued after consent — mcat_… access tokens. Review grants under Connected Apps |
| Tools | 17 (see Tools reference) |
Clients that speak MCP OAuth discover it from a 401 + WWW-Authenticate on this endpoint. You do not paste a token in that path — you sign in and approve access.
Before you connect
API token path
- Create an API key at vm.massedcompute.com/settings/api.
- Pick the right scope. Read-only for look-ups (fleet, billing, SSH keys, recipes). Full to create or terminate instances and manage SSH keys.
OAuth path
- Point a compatible client at
https://vm.massedcompute.com/api/mcpwith no Bearer header. - Complete the Massed Compute sign-in and consent screen (full access vs read-only).
- Manage or revoke the app later under Connected Apps.
Next steps
- Set up your client — Claude Code, Cursor, Codex, and Claude Desktop (with screenshots)
- Tools reference — all tools with example prompts
- Agent Skills — downloadable workflows for GPU selection, launches, cost audits, and safe cleanup
- Troubleshooting — common issues and fixes