CostPilot MCP

Track AI spend, catch prompt waste, and audit AI coding setups from any MCP-compatible client.

CostPilot MCP connects Claude Desktop, Claude Code, Cursor, Windsurf, and other AI tools to your CostPilot dashboard. It exposes tools for request-level cost tracking, prompt optimization, spend insights, and local setup audits.

Install

Run the MCP server with npx:

npx -y costpilot-mcp

Node.js 18 or newer is required.

Claude Desktop

Add CostPilot to your Claude Desktop MCP config:

{
  "mcpServers": {
    "costpilot": {
      "command": "npx",
      "args": ["-y", "costpilot-mcp"],
      "env": {
        "COSTPILOT_KEY": "pilot_live_your_key_here"
      }
    }
  }
}

On Windows, use npx.cmd as the command because many MCP hosts spawn commands directly without going through a shell:

{
  "mcpServers": {
    "costpilot": {
      "command": "npx.cmd",
      "args": ["-y", "costpilot-mcp"],
      "env": {
        "COSTPILOT_KEY": "pilot_live_your_key_here"
      }
    }
  }
}

On Windows, the config file is usually:

%APPDATA%\Claude\claude_desktop_config.json

On macOS, the config file is usually:

~/Library/Application Support/Claude/claude_desktop_config.json

Restart Claude Desktop after saving the config.

Environment

COSTPILOT_KEY=pilot_live_your_key_here

Optional:

COSTPILOT_API_URL=https://api.aicostpilot.online
COSTPILOT_SKIP_TRACKING=true

Use COSTPILOT_SKIP_TRACKING=true when another integration, such as a Claude Code hook or gateway, already records requests. In that mode the MCP server exposes optimization, insights, and audit tools without double-counting usage.

Tools

costpilot_start : Starts a tracked request and estimates input cost before the model answers.

costpilot_track : Completes a tracked request and records final output cost.

costpilot_optimize : Reviews long prompts for token waste and practical savings.

costpilot_insights : Returns recent AI spend, request volume, token usage, and tool breakdown.

costpilot_audit : Audits local AI tool configs for bloated rules files, unused MCP servers, uncapped shell output, and other token-wasting setup issues.

Example Prompts

After connecting the MCP server, try:

How much have I spent on AI this month?

Audit my Claude and Cursor setup for wasted tokens.

Can you optimize this long prompt before answering?

Troubleshooting

If the server prints COSTPILOT_KEY is required, add a valid CostPilot API key to the MCP config env block.

If Claude Desktop does not show the tools, restart Claude Desktop and confirm the package command is exactly:

"args": ["-y", "costpilot-mcp"]

On Windows, also confirm:

"command": "npx.cmd"

If you already track Claude Code with a CostPilot hook, set:

"COSTPILOT_SKIP_TRACKING": "true"

Links

• Dashboard: https://aicostpilot.online
• Package: https://www.npmjs.com/package/costpilot-mcp