@swarmdo/gpu-mcp

MCP server that gives Claude Code — or any MCP client — the ability to rent GPUs and run serverless GPU inference. Ask your agent to find an H100, deploy a vLLM endpoint, or work out why an endpoint is wedged, and it can actually do it.

Implements the Model Context Protocol. 43 tools over the SwarmDo GPU platform.

Install

Claude Code:

claude mcp add swarmdo-gpu -- npx -y @swarmdo/gpu-mcp

Or add it to any MCP client's config:

{
  "mcpServers": {
    "swarmdo-gpu": {
      "command": "npx",
      "args": ["-y", "@swarmdo/gpu-mcp"],
      "env": { "SWARMDO_API_KEY": "sk_live_…" }
    }
  }
}

Get an API key from your dashboard at gpu.swarmdo.com. Without one, only the public catalog tools work.

What you can ask for

"Find the cheapest A100 in us-east and tell me what it costs per hour."

"Deploy Qwen2.5-7B as a serverless endpoint that scales to zero, then send it a test prompt."

"My endpoint has been queuing jobs for ten minutes and nothing is completing. What's wrong?"

That last one is the reason this exists. The agent can pull serverless_status for a derived state — serving, starting, stuck, throttled, failing, idle — read masked logs, and then act: purge a wedged queue, restart workers, or scale the endpoint.

Tools

GPUs — search_gpus, rent_gpu, list_rentals, instance_telemetry, simulate_telemetry

Serverless — deploy_serverless, list_serverless, serverless_status, serverless_scale, serverless_suspend, serverless_resume, serverless_logs, serverless_purge_queue, serverless_restart, delete_serverless, list_prebaked

Models — list_models, deploy_model, list_managed_models, run_managed_model

Batch & data — list_jobs, cancel_job, list_datasets, create_dataset, list_storage, create_volume

Cost control — get_balance, idle_status, set_idle_policy, run_idle_sweep, cache_status, prewarm_cache

Reliability — sla_status, set_sla_tier, settle_sla, fleet_trust, list_alerts, create_alert

Account — list_notifications, list_activity

Cost and safety

These tools spend real money. rent_gpu starts an hourly charge, and deploy_serverless bills per GPU-second once a worker is running — including the time a worker spends pulling an image and loading weights before it serves anything.

Worth knowing before you let an agent loose:

• Set a spend cap when renting, and an idle policy (set_idle_policy) so forgotten instances stop themselves.
• minWorkers: 0 scales to zero and costs nothing while idle, at the price of a cold start.
• get_balance is cheap to call — have the agent check it before provisioning.
• Review your MCP client's approval settings. Provisioning tools are the ones you want to approve by hand.

Provider masking

SwarmDo brokers capacity across several upstream clouds. Tool results never reveal which one is behind a given instance or endpoint — ids, hostnames and prices are all SwarmDo's, and the routing underneath can change without anything you built noticing.

License

MIT — see LICENSE. The licence covers this client library only; it grants no rights to the SwarmDo platform itself, which is governed by your account's terms of service. "SwarmDo" and the SwarmDo logo are trademarks and are not licensed for use as your own product's name or branding.

Links

• Docs: gpu.swarmdo.com/docs
• Dashboard: gpu.swarmdo.com
• TypeScript SDK (same platform, for code rather than agents): @swarmdo/gpu-sdk

Requires Node 20+.