reposcout

LLM-driven GitHub repo discovery and ranking, exposed as a Model Context Protocol (MCP) server.

When an AI coding agent needs "the best open-source library for X", it usually falls back on stale training-data recall. reposcout gives the agent a fresh, evidence-backed view of the GitHub ecosystem: it searches by intent, pulls READMEs and metadata, computes deterministic popularity / maintenance / completeness signals, and combines them with the agent's own relevance judgment into a ranked, explainable shortlist.

It ships as a stdio MCP server with three composable tools, plus a companion Claude Code skill that orchestrates them.

How it works

Discovery is split into three tools so the deterministic math stays server-side and the judgment stays with the LLM driving the loop:

| Tool | What it does | |------|--------------| | repo_search | Runs several GitHub search queries built from the objective, unions and dedupes them, and returns compact records with computed popularity_score and maintenance_score. | | repo_enrich | For a shortlist, fetches the cleaned + truncated README plus completeness signals (license, homepage, description, topic count, README size). Results are cached locally. | | repo_rank | Combines four sub-scores — relevance, popularity, maintenance, completeness — into a weighted ranking. The agent supplies relevance + completeness; reposcout owns popularity + maintenance and the final math. |

Scoring is deterministic: log-scaled popularity (star/fork caps), exponential-decay maintenance (180-day half-life) with an open-issues penalty. Default weights are relevance 0.4 / popularity 0.2 / maintenance 0.2 / completeness 0.2 (normalized, overridable per call). Archived repos are excluded by default.

Requirements

• Node.js >= 22 (uses the built-in node:sqlite).
• A GitHub token. reposcout reads GITHUB_TOKEN or GH_TOKEN; if neither is set it falls back to gh auth token. So if your gh CLI is logged in (gh auth login), no extra setup is needed.

Install

reposcout runs as a local stdio MCP server — point any MCP client at it.

Claude Code

claude mcp add reposcout -- npx -y @rakeshroushan/reposcout

Claude Desktop / Cursor / other MCP clients

Add to the client's MCP config (for Claude Desktop, claude_desktop_config.json):

{
  "mcpServers": {
    "reposcout": {
      "command": "npx",
      "args": ["-y", "@rakeshroushan/reposcout"],
      "env": { "GITHUB_TOKEN": "ghp_..." }
    }
  }
}

Omit env to use your gh CLI login instead.

From source

git clone https://github.com/Rakesh1002/reposcout
cd reposcout
pnpm install
pnpm build

Then point the client at the built server:

{
  "mcpServers": {
    "reposcout": {
      "command": "node",
      "args": ["/absolute/path/to/reposcout/dist/server.js"]
    }
  }
}

Using it

Once connected, state an objective in plain language:

Find the best TypeScript library to generate OpenAPI types from Zod schemas — min 100 stars, actively maintained.

The agent expands that into complementary GitHub queries, calls repo_search, triages a shortlist, calls repo_enrich, scores relevance + completeness from the READMEs, and calls repo_rank to return a ranked shortlist with one-line reasoning per repo. The companion reposcout Claude Code skill encodes that pipeline so you never touch raw JSON.

Development

pnpm dev        # run the server from source (tsx)
pnpm test       # node --test over tests/*.test.ts
pnpm typecheck  # tsc --noEmit
pnpm lint       # biome check
pnpm build      # tsc -> dist/

The cache lives at ~/.reposcout/cache.sqlite (24h TTL).

License

MIT — see LICENSE.