mcp-context-cost

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

We measured 82 popular MCP servers (65 measurable without credentials). The spread is 1,700×: from markitdown’s 64 tokens to the official GitHub server’s 54,422 tokens — more than a quarter of a 200K context window, before the agent takes a single action.

• 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

[context cost | 2,061 tokens]

A shields.io endpoint badge whose number links to a versioned methodology and a raw tools/list capture. Anyone can re-derive it:

import { getEncoding } from "js-tiktoken";
const m = JSON.parse(fs.readFileSync("measurement.json", "utf8"));
const n = getEncoding("o200k_base").encode(JSON.stringify(m.rawToolsCapture)).length;
// n === m.totalTokens, or the badge is wrong and you just proved it

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

This site is open source. Improve this page.