mcp-context-cost

How much of your agent’s context window does an MCP server eat before it does anything?

We measure 107 popular MCP servers; 87 have a number today, and every failure is listed with its reason. Ranked on the wire, the spread is 1,700×: from postgres at 32 tokens to github at 54,622 tokens, before the agent takes a single action. Of that, an Anthropic request carries 10,735 tokens as tool definitions, and Claude counts those at 18,728. Second-heaviest is comfyui-mcp at 50,268 on the wire, 49,877 carried, — on Claude.

• What moved — each server’s most recent cost movement, dated, and which half of the server moved
• The dashboard — the ranked chart; open any row for its breakdown
• Server pages — per-server: where the tokens are, tool by tool
• The leaderboard — every candidate listed, failures included with reasons
• Methodology v1.0 — what the number is, how to reproduce it in five lines, what it is not
• Get the badge — one line in your README, backed by a published measurement

The badge

A shields.io endpoint badge whose number links to a versioned methodology and to the raw tools/list capture it was counted from. Anyone can re-derive it from that capture in five lines, and disagree with it in one — METHODOLOGY §reproduce it has them.

That is the point. Numbers you can’t audit are marketing; numbers you can refute are measurements.

This site is open source. Improve this page.