massed-compute-mcp
Model Context Protocol server for Massed Compute — list GPU inventory, launch and manage VM instances, manage SSH keys, audit billing.
Documentation
Massed Compute MCP Server
A Model Context Protocol server that lets AI assistants — Claude, Cursor, Codex, ChatGPT, and other MCP-compatible clients — interact with your Massed Compute account: browse GPU inventory, launch and manage VMs, audit billing, discover setup recipes.
Contents
Overview
The packages published from this repo are a verbatim proxy of the public MCP endpoint at `https://vm.massedcompute.com/api/mcp` — zero added surface. Every JSON-RPC message (`initialize`, `tools/list`, `tools/call`, `resources/list`, `resources/read`, `ping`, …) is forwarded unchanged with your stored API key injected as a Bearer header, and responses come back unchanged. Nothing is added, renamed, filtered, or rewritten on the way through:
- The hosted endpoint owns the catalog. Tools, resources, schemas, and descriptions come from the live endpoint — the local package never goes stale when new capabilities (like setup recipes) ship upstream.
- The hosted endpoint owns security. Key-scope enforcement (read-only keys hide destructive tools) and credential redaction happen server-side, identically for every client.
- What the local package adds is exactly what a stdio client needs: a stdio↔streamable-HTTP bridge, OS-conventional storage for your API key (`0600` on POSIX), and one-shot client wiring via `init`.
Issue a read-only key for analysis-only assistants and the endpoint rejects destructive tool calls (launch, restart, terminate, SSH-key changes) at call time. Works with Claude Code, Claude Desktop, Cursor, and Codex. Run `massed-compute-mcp tools` to print the live catalog — GPU inventory, VM images, instance lifecycle, coupons, billing, SSH keys, and setup recipes.
Beyond raw tools, Massed Compute publishes Agent Skills — markdown workflow templates for common operations like GPU selection and cost auditing. Full docs at vm-docs.massedcompute.com/docs/category/mcp.
Installation
Step 1. Get your API key
Open vm.massedcompute.com/settings/api, create a key (read-only for analysis-only assistants; full-access to allow launch / restart / terminate / SSH-key changes), copy it.
Step 2. Pick an install path
- Hosted endpoint — point your client at the streamable-HTTP URL, zero install
- Local CLI — runs on your machine, key stored in your OS config dir
Hosted endpoint
The local packages proxy this endpoint verbatim, so pointing your client straight at it gives you the identical catalog with nothing to install. Pick the snippet for your client:
Claude Code
claude mcp add --transport http massed-compute \
https://vm.massedcompute.com/api/mcp \
--header "Authorization: Bearer MC_TOKEN"Cursor — `~/.cursor/mcp.json`
{
"mcpServers": {
"massed-compute": {
"url": "https://vm.massedcompute.com/api/mcp",
"headers": { "Authorization": "Bearer MC_TOKEN" }
}
}
}Codex — `~/.codex/config.toml`
[mcp_servers.massed-compute]
url = "https://vm.massedcompute.com/api/mcp"
bearer_token_env_var = "MC_TOKEN"
enabled = trueClaude Desktop
Claude Desktop does not yet speak streamable-HTTP MCP, so use `mcp-remote` as a stdio↔HTTP bridge. Config at `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or `%APPDATA%\Claude\claude_desktop_config.json` (Windows):
{
"mcpServers": {
"massed-compute": {
"command": "npx",
"args": [
"-y", "mcp-remote",
"https://vm.massedcompute.com/api/mcp",
"--header", "Authorization: Bearer MC_TOKEN"
]
}
}
}Local CLI
Install via whichever ecosystem you prefer:
npm install -g massed-compute-mcp # Node >= 20
# or
pip install massed-compute-mcp # Python >= 3.10
# or
uv tool install massed-compute-mcp # fast Python install via uvRun one-shot setup:
massed-compute-mcp init`init` prompts for the key, validates it upstream, stores it at `0600` (POSIX), detects installed MCP clients (Claude Code, Cursor, Claude Desktop, Codex), and offers to wire each one. A timestamped backup is taken before any client edit. Restart wired clients to pick up the tools.
Config file location:
- Linux: `$XDG_CONFIG_HOME/massed-compute/config.json` (falls back to `~/.config/...`)
- macOS: `~/Library/Application Support/massed-compute/config.json`
- Windows: `%APPDATA%\massed-compute\config.json`
Non-interactive (CI / scripts):
# Key from env, auto-wire every detected client
MASSED_COMPUTE_API_KEY= massed-compute-mcp init --yes
# Key from a file, wire only specific clients
massed-compute-mcp init --token-file ~/keys/mc --yes --clients claude-code,cursor
# Store the key only; don't touch any client config
massed-compute-mcp init --yes --no-install-clientsAdd or remove a client later:
massed-compute-mcp install-client cursor # claude-desktop | codex | claude-code
massed-compute-mcp uninstall-client cursor`install-client` is idempotent — re-running when the entry already matches is a silent no-op.
Verifying the connection
massed-compute-mcp doctor`doctor` confirms the stored key still works, prints the tool catalog, and shows copy-pasteable snippets for any clients you didn't auto-wire. Or just ask your assistant *"Validate my Massed Compute API key."* — a `{ message: "Valid Token" }` response confirms the wiring.
CLI reference
| Command | What it does |
|---|---|
| `massed-compute-mcp` *(no args)* | Run the MCP server over stdio |
| `massed-compute-mcp init` | First-run setup: prompt, validate, store, wire clients |
| `massed-compute-mcp doctor` | Verify the stored key and print client snippets |
| `massed-compute-mcp install-client ` | Wire a single client (`claude-code`, `cursor`, `claude-desktop`, `codex`) |
| `massed-compute-mcp uninstall-client ` | Remove our entry from a client config |
| `massed-compute-mcp config show` | Print resolved config path, masked key, resolution chain |
| `massed-compute-mcp logout` | Delete the stored API key |
| `massed-compute-mcp tools [--json]` | Fetch and print the live tool catalog from the hosted endpoint |
| `massed-compute-mcp version` | Print the version |
Key resolution
When the server starts, the API key is taken from the first source that provides one:
1. `--token ` CLI flag
2. `--token-file ` CLI flag (first line, trimmed)
3. `MASSED_COMPUTE_API_KEY` environment variable
4. Stored config file written by `init`
If none of those are set, the server exits non-zero with a pointer to `massed-compute-mcp init`. Override the upstream with `MASSED_COMPUTE_API_BASE_URL` (default `https://vm.massedcompute.com`).
Resources
License
MIT — see LICENSE.
Frequently asked questions
What is massed-compute-mcp?
massed-compute-mcp is Model Context Protocol server for Massed Compute — list GPU inventory, launch and manage VM instances, manage SSH keys, audit billing.
How do I install massed-compute-mcp?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is massed-compute-mcp open source?
Yes — it is hosted on GitHub at https://github.com/Massed-Compute/massed-compute-mcp.
Related MCP tools
Model Context Protocol Servers
The Open-Source Multimodal AI Agent Stack: Connecting Cutting-Edge AI Models and Agent Infra
A MCP for Claude Desktop / Claude Code / Windsurf / Cursor to build n8n workflows for you
MCP server to provide Figma layout information to AI coding agents like Cursor
The world's best AI personal assistant for email. Open source app to help you reach inbox zero fast.
Instant is the best backend for AI-coded apps. You get auth, permissions, storage, presence, and streams — everything you need to ship apps your users will love.
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP