Skip to main content
Give a coding agent the Compute Agent Skill so it installs the CLI, dry-runs a real example, and only then submits a GPU job.

​
Raw skill (for agents)

Agents should fetch this file: https://compute.cx/SKILL.md
The same document is also at https://compute.cx/.well-known/skills/compute/SKILL.md.
GPU runs spend prepaid credit. --yes and MCP compute_run with confirm_spend=true skip the interactive price confirmation and still bill. --dry-run and compute_dry_run print the upload plan and create no machine.

​
What to tell the agent

That is enough for Cursor, Claude Code, Codex, and similar tools that can fetch a URL.

​
MCP

After the CLI is installed, agents can call Compute as tools instead of shelling out. Cloud or hosted agents that cannot spawn a local process can instead reach https://mcp.compute.cx/mcp over Streamable HTTP with Clerk OAuth. Both servers can dry-run and submit work; the hosted server accepts the Python source inline. See MCP server for both.
This repository also ships as an Agent Plugin (plugin.json + skills/compute/SKILL.md + mcp.json) for Cursor team marketplaces and marketplace publish.

​
What the skill does

  1. Installs with curl -fsSL https://compute.cx/install.sh | sh and compute setup.
  2. Writes the tested simple_mlp.py example from First run.
  3. Runs compute run simple_mlp.py::train --gpu H100-SXM --dry-run (no machine), or MCP compute_dry_run.
  4. On your confirmation, submits with --wait --yes or compute_run + confirm_spend=true and checks teardown via compute machines.
  5. When you ask it to report an unexpected failure, calls compute_report_issue, gives you the returned rpt_ id, and follows replies with compute_get_report. The CLI equivalents are compute report and compute report status.
Do not paste API keys, session tokens, config.toml, Checkout session ids, secret values, or signed URLs into the agent chat or a report.

​
Next

MCP server is the tool surface. First run is the same example, written for a human. Support explains reports, follow-up, contact, and system status. Run is the flag reference.