mcp-checkup 0.0.1

🩺 A health check for your MCP setup — measure the context tax and hygiene of your MCP servers

🩺 mcp-checkup

A health check for your MCP setup.

One command that tells you what your MCP servers cost you — in context-window tokens and dollars — and whether they follow basic security hygiene.

---

The problem

Every MCP server you connect injects its full tool schemas into your model's context window — on every single request — before you type a word.

• GitHub's official MCP server alone consumes ~17,600 tokens per request in tool definitions (autopsy).
• Teams have burned 72% of a 200k context window on tool definitions before doing any actual work (MCP spec issue #2808).
• A security audit of publicly exposed MCP servers found 119 out of 119 sampled allowed unauthenticated access to internal tool listings.

You are paying a context tax on every request, and you probably have no idea how big it is. mcp-checkup measures it.

What it looks like

⚠️ Target UX — under active development. The output below is the design we are building toward. Today the CLI installs and prints the roadmap.

$ uvx mcp-checkup

  🩺 MCP Checkup — 4 servers, 38 tools

  WEIGHT                                    HYGIENE
  ────────────────────────────────────     ─────────────────────────────────
  github       17,612 tok   $0.053/req     ⚠ 2 tools reachable without auth
  slack         4,105 tok   $0.012/req     ✓ ok
  filesystem    2,890 tok   $0.009/req     ⚠ unrestricted root path
  custom-api    9,441 tok   $0.028/req     ⚠ 3 schemas 8× over minimal size

  Context tax: 34,048 tokens — 17% of a 200k window, before your first message.

  Run `mcp-checkup --fix` to emit compressed schemas.
  Run `mcp-checkup --fail-over 20000` in CI to stop the bloat from coming back.

Quickstart

uvx mcp-checkup          # or: pip install mcp-checkup

Point it at your client config (.mcp.json, Claude Desktop, Cursor, …) or a running server. That's the whole interface — one command, one report.

Roadmap

• Weight — per-tool token count of every server's schemas (tiktoken + Anthropic token counting)
• Weight — $ cost per session per model (GPT, Claude, Gemini pricing tables)
• Hygiene — missing auth, exposed tool listings
• Hygiene — bloated schemas (vs. minimal equivalent), suspicious / injection-prone tool descriptions
• --fix — emit compressed schemas that preserve call structure
• CI mode — --fail-over <tokens> budget gate for your pipeline
• Client config auto-discovery (Claude Desktop / Claude Code / Cursor / Windsurf)

Status

Early days — the measurement engine is being built in the open. If the context tax bothers you too: ⭐ star the repo to follow along, open an issue with your worst MCP token bill, or see CONTRIBUTING.md to help build it.

License

Apache-2.0