Getting Started

tokencost (tokencost-dev) is an MCP server that provides real-time LLM pricing data. It runs as a local process and communicates over stdio.

Prerequisites

Node.js 18+ — download

That’s it. No API keys, no accounts, no configuration files.

Claude Code

The fastest way to get started:

claude mcp add tokencost-dev -- npx -y tokencost-dev

This registers the MCP server and it will be available in all future Claude Code sessions.

Claude Desktop

Open Claude > Settings > Developer > Edit Config to open claude_desktop_config.json, then add:

{
  "mcpServers": {
    "tokencost-dev": {
      "command": "npx",
      "args": ["-y", "tokencost-dev"]
    }
  }
}

Save the file and restart Claude Desktop.

VS Code (GitHub Copilot)

Create or edit .vscode/mcp.json in your workspace:

{
  "servers": {
    "tokencost-dev": {
      "command": "npx",
      "args": ["-y", "tokencost-dev"]
    }
  }
}

VS Code will detect the file and make the tools available in Copilot chat (agent mode).

Cursor

Add to your Cursor MCP config (.cursor/mcp.json):

{
  "mcpServers": {
    "tokencost-dev": {
      "command": "npx",
      "args": ["-y", "tokencost-dev"]
    }
  }
}

Windsurf

Open Windsurf Settings > Cascade > MCP Servers, or edit ~/.codeium/windsurf/mcp_config.json directly:

{
  "mcpServers": {
    "tokencost-dev": {
      "command": "npx",
      "args": ["-y", "tokencost-dev"]
    }
  }
}

Other MCP clients

The server config is the same for any MCP client that supports stdio transport:

{
  "mcpServers": {
    "tokencost-dev": {
      "command": "npx",
      "args": ["-y", "tokencost-dev"]
    }
  }
}

Consult your client’s documentation for where to place this configuration.

Your first query

Once installed, just ask your AI assistant a pricing question in natural language:

“How much does Claude Sonnet 4.5 cost per million tokens?”

The assistant will call the get_model_details tool and return something like:

Model: claude-sonnet-4-5
Provider: anthropic
Mode: chat

Pricing (per 1M tokens):
  Input:  $3.00
  Output: $15.00

Context Window:
  Max Input:  200K
  Max Output: 8K

Capabilities: vision, function_calling, parallel_function_calling

How it works

On first use, the server fetches pricing data from the LiteLLM community registry
Data is cached in-memory for 24 hours (with a disk fallback)
Your AI assistant calls one of the 4 tools via the MCP protocol
Results are returned as formatted text

No data leaves your machine — the only network request is fetching the public pricing registry.