RepoSniffer

There's a repo for that. Let RepoSniffer find it.

AI-first GitHub repo discovery. Describe a feature in plain language — "markdown editor with live preview" — and RepoSniffer returns a ranked, adoption-grade list of open-source projects that implement it, with evidence and quality signals (stars, activity, license, archived status).

Built to be the grounding layer for coding agents: stop hallucinating repos, get a verifiable "best of kind" answer with reasons.

Quickstart

# CLI
uvx reposniffer "markdown editor live preview" --language python --top-k 5

# MCP server (stdio) — wire into opencode, Claude Code, Codex, Cursor, ...
uvx --from reposniffer reposniffer-mcp

Set GITHUB_TOKEN to raise search rate limits (authenticated = 30 req/min vs ~10).

Wire into your agent

opencode — add to opencode.json:

{
  "mcp": {
    "reposniffer": {
      "type": "local",
      "command": ["uvx", "--from", "reposniffer", "reposniffer-mcp"],
      "environment": { "GITHUB_TOKEN": "ghp_..." }
    }
  }
}

Claude Code: claude mcp add reposniffer -- uvx --from reposniffer reposniffer-mcp Codex/Cursor: add an MCP server pointing at uvx --from reposniffer reposniffer-mcp (stdio).

MCP tools

Every result carries as_of (a freshness timestamp agents can cite), a flags list (archived, no-license, strong-copyleft, stale, ...), a license_category (permissive / weak-copyleft / strong-copyleft / unknown), and a targeted snippet showing why the repo matched.

Architecture

1. Coarse candidate fetch — GitHub Search API (in:readme, language/license/stars filters).
2. Hybrid rerank — embed each candidate's description + README front matter (not the whole README, to avoid dilution), cosine vs embedded query, plus a lexical-overlap boost for literal matches.
3. Quality scoring — popularity (log stars), activity (pushed_at half-life), license category, archived penalty; weights differ by intent (adopt vs study).
4. Adoption safety — permissive/weak/strong-copyleft classification flags GPL/AGPL repos before you depend on them.
5. Local SQLite cache — repos, READMEs, embeddings, query results → fast repeat queries, index grows over time.

Embeddings are pluggable: default is a zero-config local fastembed ONNX model (no torch, no API key); set REPOSNIFFER_EMBED_BACKEND=api plus an OpenAI-compatible endpoint for stronger quality.

Eval

Ground-truth queries live in eval/queries.py (feature → known-good repos). Run with a token (each query fetches ~50 READMEs):

GITHUB_TOKEN=ghp_... uv run python -m eval.run

Reports hit@1 / hit@3 / hit@5. Current live result: hit@1 0.50, hit@3 0.83, hit@5 0.83.

Known limitation: the candidate stage depends on GitHub Search API relevance, which can fail to recall canonical repos with weak descriptions/READMEs (e.g. Kozea/WeasyPrint — description is just "The awesome document factory"). Semantic rerank can only rank what the candidate fetch surfaces.

Project layout

src/reposniffer/
  config.py        # env-driven settings
  cache.py         # sqlite store (repos, readmes, embeddings, query cache)
  engine/
    github.py      # GitHub REST client + search query builder + text/snippet utils
    embed.py       # Embedder protocol: local fastembed + OpenAI-compatible API
    score.py       # quality + license-category + lexical scoring
    search.py      # orchestration (Engine)
  mcp/server.py    # MCPServer (mcp 2.x)
  cli.py           # Typer CLI
eval/              # golden query → repo eval harness
tests/             # offline (fake transport + fake embedder)

Development

uv sync
uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest

CI (lint/format/type/tests) runs on every push and PR. Publishing to PyPI happens on v* tags via trusted publishing — enable it once on the PyPI project settings, then: git tag v0.1.0 && git push --tags.

Note: mcp 2.x is used — MCPServer (FastMCP was renamed in mcp 2.0). Pin mcp<2 if you need the v1 API.

License

MIT