mouseketool

Before you build something new, find out whether it already exists.

When you tell your coding agent you want to build a new tool, app or library, mouseketool searches GitHub for projects that already do it. It checks them for real: their health, their license, and the code behind each README's claims. Then it answers with a verdict against your requirements: reuse, adapt, use as a base, or build from scratch.

How it works

• An MCP server does the mechanical work and makes no judgment:
  • many GitHub searches in one call;
  • repository inspection: README, files, activity, releases, issues, contributors and license, plus plain health signals;
  • file reads.
• An Agent Skill teaches your agent the method:
  1. separate the must-haves from the nice-to-haves;
  2. search from several angles;
  3. inspect the mature candidates;
  4. confirm in the code what the verdict relies on;
  5. answer verdict first, with pros and cons per candidate.

The skill triggers by itself when you want a new reusable tool to exist, and stays silent on small tasks such as a one-off script, a fix or a new flag. mouseketool has no LLM of its own: your agent's model does the judging.

Install

You need Node 24 or later, pnpm, and a GitHub token: either GITHUB_TOKEN, or being logged in with gh auth login.

git clone https://github.com/iDiagoValeta/mouseketool.git
cd mouseketool
pnpm install --frozen-lockfile
pnpm build

Claude Code. The repository is a plugin:

claude --plugin-dir /path/to/mouseketool

opencode.

1. Copy skills/mouseketool/ into a skills directory, such as .opencode/skills/ or ~/.config/opencode/skills/.
2. Add the server to opencode.json, as in opencode.example.json, with the absolute path to dist/server.mjs.

The token is sent only to api.github.com. It is never logged and never appears in a tool result. Searches are queued to GitHub's limit of 30 per minute.

Evaluation

evals/ runs the whole method headless, against real models and the real GitHub API. It has four kinds of case:

• 20 positive ideas, built on real projects, each in a vague and a defined version;
• 6 ideas covered only in part, where the right answer is base with a mature project;
• 6 invented ideas, where the right answer is to build;
• 6 small tasks, where the skill must stay silent.

A recommendation counts as a hit only if all three hold:

• the project came from that run's search, not from the model's memory;
• the host read its code;
• a separate judge confirms it meets every must-have.

Any project that meets them counts, not only the one a case was built on.

pnpm eval --calibrate   # check the judge and the cases first
pnpm eval --pass 0      # a full pass

Results of 2026-11-04 (58 runs), within every threshold:

| Host | Defined | Vague | Partial | Invented | Small tasks | |---|---|---|---|---|---| | Claude Code, Sonnet 5.5 | 92% | 83% | 100% | 100% | 100% | | opencode, deepseek-flash | 100% | 100% | 50% | 100% | 100% |

Every run, with the host's full answer and the judge's verdicts, is in evals/results/2026-11-04/.

Development

pnpm check   # typecheck, lint (including the hexagonal layer rules) and tests
pnpm build   # bundle the server into dist/server.mjs

The code is hexagonal:

• src/domain and src/application import no adapter or IO module;
• src/infrastructure holds the GitHub, run-log and MCP adapters;
• src/composition-root.ts wires them.

See AGENTS.md for the conventions. GitHub is the only source for now.

License

MIT