Claude Code integration

What the MCP server is

The Tokeven MCP server is the pre-send advisor, running inside Claude Code (or any MCP-compatible tool). It lets Claude price a call before making it. It does not track usage: nothing it does sends a usage event to Tokeven.

Add it to Claude Code

It is listed in the MCP registry as com.tokeven/tokeven, or add it by hand:

claude mcp add tokeven -- uvx --with 'mcp>=2.0,<3.0' tokeven mcp

Or install the package with the MCP extra (pip install "tokeven[mcp]") and point your MCP configuration at the tokeven-mcp command.

tokeven mcp and tokeven-mcp are the same server. It takes no command-line arguments; anything it needs is read from the environment.

Tools

• •tokeven_estimate - prices a call on the model you name and scores cheaper alternatives across Claude, OpenAI, Gemini and Grok, with a confidence score per task type. It runs entirely on your machine and needs no account.
• •tokeven_improve - rewrites a prompt to be clearer and shorter (paid plans). The prompt is sent to Anthropic using your own ANTHROPIC_API_KEY, never to Tokeven; Tokeven only checks that your organization is on a paid plan, which needs TOKEVEN_INGEST_TOKEN set in the server's environment.

Tracking is separate

Usage reaches your dashboard only from instrumented calls: the Python wrapper (wrap() or tokeven.instrument()), the LangChain, LiteLLM and OpenAI Agents integrations, the TypeScript SDK (@tokeven/sdk), the opt-in OpenTelemetry receiver, or the self-hosted sidecar. Run tokeven status to see which are live on your machine.

Adding the MCP server does not make Claude Code's own calls appear in your dashboard, and usage on a Claude Pro, Max or Team subscription is not tracked today.