prime-intellect-mcp

Let Claude Code rent, drive, and terminate Prime Intellect GPU pods on its own — with hard spend caps you control.

📺 30-second demo: (replace with video / GIF on launch — see docs/DEMO_SCRIPT.md)

What this is

An MCP server that connects Claude Code (or any MCP client) to your Prime Intellect account. With it, the agent can:

• 🔍 Find the cheapest GPU pod that matches your requirements
• 💸 Quote a price before committing money
• 🛒 Provision the pod (only after you say confirm=True)
• 🖥️ SSH into it (the connection string is handed to the agent's own Bash tool)
• 🛑 Terminate it when work is done — and warn loudly if you forget

Built for one workflow: telling Claude "rent the cheapest H100, run my training script, then kill it" and not waking up to a $400 bill.

Install in 60 seconds

You only need this much to start renting GPUs through Claude Code:

1. Get a Prime Intellect API key

Click here to generate one → set permissions:

Copy the key — it starts with pit_….

2. Add the server to Claude Code

Open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or your project's .mcp.json, and paste:

{
  "mcpServers": {
    "prime-intellect": {
      "command": "uvx",
      "args": ["prime-intellect-mcp"],
      "env": {
        "PRIME_API_KEY": "pit_PASTE_YOURS_HERE",
        "PRIME_MAX_HOURLY_USD": "5",
        "PRIME_MAX_TOTAL_USD": "40"
      }
    }
  }
}

That's it. Restart Claude Code and ask: "What GPUs are available right now under $1/hr?"

Don't have uvx? Install it with curl -LsSf https://astral.sh/uv/install.sh | sh (or brew install uv). It's a one-liner installer for the uv package manager and you'll never have to manage a virtualenv again.

✨ Add SSH (optional, +2 min) — needed for Claude to actually run code on the pod

The server above can already provision/inspect/terminate pods. But to have Claude Code SSH into a running pod and execute commands on it, Prime Intellect needs to know your machine's public SSH key.

3. Find or generate an SSH key on your machine

ls ~/.ssh/*.pub          # if you have id_ed25519.pub or similar, you're set
# otherwise:
ssh-keygen -t ed25519 -C "you@example.com"   # press Enter through the prompts

4. Register the public key with Prime Intellect

cat ~/.ssh/id_ed25519.pub    # or whichever .pub file you have

Copy the output (one line starting with ssh-ed25519 …), then paste it into the Add SSH key form at app.primeintellect.ai/dashboard/ssh-keys.

That's it. Future pods will have your public key in authorized_keys, and Claude Code's Bash tool can SSH straight in:

ssh ubuntu@<pod-ip-from-pod_status> "nvidia-smi"

Coming in v0.2: a register_ssh_key MCP tool that does step 4 from inside Claude (no browser visit). See the issue tracker to follow along.

What Claude can now do (the 9 tools)

Safety: nothing provisions silently

Three layers, in order:

1. Quote first. pod_quote returns a price + a 60-second token. No side effects. The dollar amount is now in the agent's context.
2. Explicit confirm. pod_create (and pod_terminate) requires confirm=True. Without it, you get a dry-run preview.
3. Hard env-var caps. PRIME_MAX_HOURLY_USD blocks any pod above the rate. PRIME_MAX_TOTAL_USD blocks any (rate × max_lifetime_hours) above the budget. Wallet balance is also enforced. None of these caps can be overridden by tool arguments — they're read at every call.

Defaults: PRIME_MAX_HOURLY_USD=5, PRIME_MAX_TOTAL_USD=40. Set them in your config's env block.

Every pod_create / pod_terminate is appended as JSON to ~/.prime-intellect-mcp/audit.log, so you have a complete history of what the agent did with your money.

Example prompts (paste these into Claude Code)

List the cheapest 1×H100 pods available right now. Show me the top 3 by hourly price.

Quote a 1×A100 80GB with 100GB disk, 8 vCPU, 64GB RAM. Don't provision yet —
just show me what it would cost.

I need to fine-tune a 7B model overnight. Find the cheapest 1×H100 with 200GB
disk, max $40 total budget, max 12 hours. Provision it, give me the SSH command,
and remind me to terminate when I'm done.

Check if I have any running pods I forgot about and show me their hourly cost.

Terminate pod abc123. Confirm before doing it.

Troubleshooting

$ env | grep PRIME

Why another one?

There's a prime-mcp-server 0.1.2 on PyPI. It's a thin proof-of-concept; this isn't a fork. Differences for unattended overnight use:

Local development

git clone https://github.com/kvrancic/prime-intellect-mcp
cd prime-intellect-mcp
uv sync
uv run pytest -m "not live"        # 32 fast tests, no network, no spend
uv run ruff check .
uv run mypy src

Live smoke test (provisions cheapest available GPU, runs nvidia-smi, terminates; ~$0.05 spend):

PRIME_API_KEY=pit_... PRIME_LIVE_TEST=1 PRIME_LIVE_MAX_HOURLY=0.60 \
PRIME_MAX_HOURLY_USD=0.60 PRIME_MAX_TOTAL_USD=2.00 \
uv run pytest tests/test_smoke_live.py -v -s

Roadmap

• v0.2 — register_ssh_key MCP tool (kill the dashboard step), Sandboxes (prime-sandboxes SDK), Environments Hub
• v0.3 — Optional auto-terminate daemon (server-side enforcement of max_lifetime_hours); cost telemetry
• v1.0+ — Hosted/OAuth deployment when Prime Intellect ships OAuth; submission to Anthropic connector directory

Acknowledgements

• Prime Intellect for the prime Python SDK that does 90% of the work
• MIT 6.S192 for the credits that made testing this possible
• Anthropic for MCP
• FastMCP for the framework

License

MIT — see LICENSE.

Contributing

Issues and PRs welcome. Please run uv run pytest -m "not live" and uv run ruff check . before submitting.