GPU Compute / Integrations

MCP (agent compute)

badgr-mcp exposes Badgr GPU compute as MCP tools so coding agents (Claude, Cursor, and others) can provision GPU jobs and model endpoints directly from chat. Hard budget and runtime limits prevent runaway billing.

Install

npm install -g badgr-mcp

Configure in Claude Desktop

Add to your Claude Desktop config at ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "badgr": {
      "command": "badgr-mcp",
      "env": {
        "BADGR_API_KEY": "your-api-key",
        "BADGR_MAX_PRICE": "5.0",
        "BADGR_MAX_RUNTIME": "60"
      }
    }
  }
}

BADGR_MAX_PRICE caps the hourly spend per job (default $5/hr). BADGR_MAX_RUNTIME auto-terminates endpoints after N minutes (default 60). Agents cannot exceed these limits even if they try.

Available tools

badgr_runRun a one-off GPU job. Returns deployment_id, status, receipt_id, logs_url.

badgr_serveProvision a persistent OpenAI-compatible endpoint. Requires max_runtime_minutes to prevent runaway billing.

badgr_statusList active GPU deployments for this account.

badgr_logsFetch logs for a deployment by ID.

badgr_downTerminate a deployment and stop billing.

Example prompts

Run a training job

“Run python train.py on an A100. Use the pytorch 2.3 image. Max $2/hr.”

Serve a model

“Serve Llama-3.1-8B on an L40S for 30 minutes so I can test it.”

Check status

“What GPU deployments are currently running?”

Safety limits

These environment variables set hard ceilings that agents cannot override:

BADGR_MAX_PRICE$5/hrHard cap on max_price_per_hour for any job or endpoint

BADGR_MAX_RUNTIME60 minHard cap on max_runtime_minutes for badgr_serve