Mcpgauge

mcpgauge

Grade an MCP server's tool definitions in seconds — deterministic, offline, no account.

[](https://github.com/Michael-WhiteCapData/mcpgauge/actions/workflows/ci.yml) [](https://pypi.org/project/mcpgauge/) [](https://www.python.org/) [](LICENSE)

An agent only uses an MCP server well if its tool definitions are clear — good names, real descriptions, documented parameters. mcpgauge launches a server, reads its tools, and scores their definition quality against a [published, deterministic rubric](docs/METHODOLOGY.md). No LLM, no network beyond the server you point it at, no sign-up.

Quickstart

pipx install mcpgauge          # or: uvx mcpgauge ...
mcpgauge "uvx ollama-handoff"

Grade A  (4.18/5)  —  8 tools
mean 4.31 · min 3.72 · methodology v0

  [B] summarize_local  3.72/5
      purpose █████  params █████  behavior █████  schema █████  naming █████  concise ████·
  [A] query  4.80/5
      ...

• --json — machine-readable output for tooling.
• --min B — exit non-zero if the server grades below B (drop it in CI).
• --quiet — print only the final grade line.

Why it exists

Most MCP servers are thin wrappers with vague tool descriptions, and agents pick the wrong tool as a result. The only comparable quality score is closed and requires registering your server on a third-party site. mcpgauge is the opposite: open methodology, offline, reproducible, and CI-runnable. You can read exactly how every point is awarded in [docs/METHODOLOGY.md](docs/METHODOLOGY.md) and argue with it.

How scoring works (short version)

Each tool is scored 0–5 on six dimensions — Purpose Clarity, Parameter Semantics, Behavioral Transparency, Schema Validity, Naming, Conciseness. The server score is 0.6 · mean + 0.4 · min across tools, so one badly-described tool drags the whole server down (which is exactly how agents fail). Full rubric: [docs/METHODOLOGY.md](docs/METHODOLOGY.md).

Status

Early and honest: v0 of the rubric, stdio servers, deterministic checks only. An optional LLM "would an agent pick the right tool?" pass and remote-HTTP support are on the roadmap. Issues and rubric critiques welcome.

Development

uv pip install -e ".[dev]"
pytest
ruff check src tests

License

MIT — see [LICENSE](LICENSE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Michael-WhiteCapData
• Source: Michael-WhiteCapData/mcpgauge
• License: MIT

Install and usage instructions live in the source repository linked above.