Skip to main content
Compute has two MCP servers: a local stdio server bundled with the CLI, and a hosted Streamable HTTP server at mcp.compute.cx for agents that cannot spawn a local process. After install and sign-in (or, for the hosted server, an OAuth connect), an agent can list GPUs, check credit, follow status, and manage secrets; both can dry-run and submit a run: the local server reads the entrypoint file from disk, the hosted server takes the Python source inline. The Agent Skill at https://compute.cx/SKILL.md is the workflow. MCP is the tool surface.
compute_run spends prepaid credit. It requires confirm_spend=true. compute_dry_run prints the upload plan and creates no machine.

​
Install the CLI

compute mcp is not a human command. The host spawns it and speaks JSON-RPC on stdin/stdout.

​
Local (stdio)

​
Cursor

Project or user mcp.json:
The server reads COMPUTE_API_KEY or ~/.compute/config.toml. After compute setup, no extra env is required. Cursor install link (CLI must already be on PATH): Add Compute MCP

​
What the local tools do

Do not paste API keys, session tokens, config.toml, Checkout session ids, secret values, or signed URLs into the agent chat or a report.

​
Hosted (Streamable HTTP)

For cloud/hosted agents (Claude.ai- or ChatGPT-style connectors, background agents) that cannot spawn a local process:
No CLI install and no COMPUTE_API_KEY — authentication is Clerk OAuth 2.1 with PKCE. A request without a valid token gets a 401 with a WWW-Authenticate header pointing at the discovery document:
Dynamic Client Registration is enabled. The authorization server at https://clerk.compute.cx publishes a registration_endpoint, so an MCP host self-registers its client and completes the browser consent; nothing has to be pre-registered. The consent signs in the Compute account whose prepaid credit the runs will spend.
Both servers expose the full set of 21 tools, including compute_report_issue and compute_get_report. The hosted server differs in the input of compute_dry_run and compute_run: instead of a project_root, they take the Python source inline as files, an array of {path, content} holding the entrypoint and every local module it imports (at most 64 .py files, 2 MiB total); third-party packages go in pip. The API plans the import closure from those files, so the hosted server never needs the calling agent’s disk. compute_run still requires confirm_spend=true.

​
Plugin (skill + MCP)

This docs repository is also an Agent Plugin: plugin.json, mcp.json, and skills/compute/SKILL.md. Import the repo into a Cursor team marketplace, or submit it at cursor.com/marketplace/publish. The skill is also a raw file:

​
Next

Agent Skill is the spend-aware workflow. Run is the flag reference.