@consumpai/mcp-server

Let any MCP-capable agent (Claude Code, Cursor, Windsurf, …) query its own AI spend — and act on it. This is Moat Layer 4 of ConsumpAI: the integration ecosystem.

Install & configure

export CONSUMPAI_API_KEY=ck_live_...
export CONSUMPAI_BASE_URL=https://app.consumpai.com   # optional

Register the server with Claude Code / Cursor:

{
  "mcpServers": {
    "consumpai": {
      "command": "npx",
      "args": ["-y", "@consumpai/mcp-server"],
      "env": { "CONSUMPAI_API_KEY": "ck_live_..." }
    }
  }
}

Tools

| Tool | Description | |:---|:---| | get_usage | Aggregated cost, tokens, forecast and anomalies for 7d/30d/90d. | | list_projects | Every tracked project ranked by spend, with ids. | | get_benchmark | Anonymous industry percentile, rank and peer recommendations. | | get_learning_recommendations | Optimizations proven on profile-similar projects. | | run_optimization | Run the engine, persist opportunities, return ranked savings. |

Design notes

• Zero dependencies — JSON-RPC 2.0 over newline-delimited stdio using only the Node.js standard library, so the package is tiny and auditable.
• Least privilege — only the API key you provide is used; run_optimization requires write scope.
• Fail-safe — transport/parse errors return JSON-RPC errors; tool failures return isError: true without crashing the server.

Development

npm install
npm run build       # bundle to dist/
npm run check:types # tsc --noEmit