GFI Scout

An MCP server and standalone CLI that finds open source issues where beginners actually succeed — not just any issue tagged good first issue.

mcp-name: io.github.Rajveerx11/gfi-scout

Why this exists

Most "good first issue" finders are glorified GitHub search wrappers. They happily hand you issues from abandoned repos, issues already claimed by three other contributors, and issues maintainers will never review.

GFI Scout ranks results by likelihood of success — repo health, merge rate, maintainer responsiveness, issue freshness, and setup complexity all feed a composite beginner_score (0-100). The dead repos sink to the bottom.

It ships as a Model Context Protocol server, plus a standalone CLI/TUI for terminal-first workflows.

Related MCP server: github-talent-mcp

Features

• 🔎 find_issues — repo-first language + topic + star-range discovery with scored issue results

• 🩺 check_repo_health — merge rate, last commit, CONTRIBUTING/CoC/CI probes → A-F grade

• ⏱️ check_issue_status — assignment, linked PRs, staleness, maintainer confirmation → AVAILABLE / LIKELY_TAKEN / STALE verdict

• 📘 get_contribution_guide — pulls and summarises CONTRIBUTING.md, detects toolchain, estimates setup complexity

• Terminal commands via gfi-scout-cli and an interactive gfi-scout-tui

• ⚡ Parallel GitHub API fan-out (asyncio.gather) + per-namespace TTL cache for repository search, issue listing, and repo-health probes

• 🎛️ All scoring weights and thresholds live in src/gfi_scout/data/scoring_weights.json — no magic numbers in code

• 🧪 100+ tests (unit + integration), mypy --strict clean, ruff clean

Requirements

Quick start

# Clone
git clone https://github.com/Rajveerx11/gfi-scout.git
cd gfi-scout

# Install uv (skip if you have it)
# macOS / Linux:        curl -LsSf https://astral.sh/uv/install.sh | sh
# Windows (PowerShell): irm https://astral.sh/uv/install.ps1 | iex

# Install dependencies (creates .venv automatically)
uv sync

# Optional: add a GitHub token (60 req/h without one, 5,000 req/h with)
cp .env.example .env
# edit .env and paste your GitHub token

# Run the MCP server (stdio transport)
uv run gfi-scout

# Or expose a local Streamable HTTP MCP endpoint
uv run gfi-scout --transport streamable-http --host 127.0.0.1 --port 8000

# Or use the standalone CLI/TUI
uv run gfi-scout-cli find python --min-stars 500
uv run gfi-scout-tui

Install as a global uv tool

If you only want to run the MCP server / CLI and don't plan to hack on the code, install it once as a global tool. No checkout, no venv to keep around:

# Install (or update) directly from GitHub
uv tool install --force --from git+https://github.com/Rajveerx11/gfi-scout gfi-scout

# Then the binaries are on $PATH:
gfi-scout                # MCP server (stdio)
gfi-scout-cli find python --min-stars 500
gfi-scout-tui

# Upgrade later:
uv tool install --force --from git+https://github.com/Rajveerx11/gfi-scout gfi-scout

A GITHUB_TOKEN in the environment (or a .env in the directory you run from) is optional — without one you run at GitHub's unauthenticated 60 requests/hour limit.

Connecting to an MCP client

Claude Desktop

Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "gfi-scout": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/gfi-scout", "gfi-scout"],
      "env": {
        "GITHUB_TOKEN": "ghp_your_token_here"
      }
    }
  }
}

Restart Claude Desktop, then try:

"Find me Python good first issues with at least 500 stars."

"Is this issue actually available? https://github.com/fastapi/fastapi/issues/12345"

"What's the setup complexity for pallets/flask?"

Other clients

Cursor, Windsurf, and VS Code Copilot each support MCP servers — point them at the same uv run command. Detailed steps in docs/SETUP.md.

MCP tools

See docs/TOOLS_REFERENCE.md for full parameter and return-shape specs.

CLI and TUI

uv run gfi-scout-cli find python --min-stars 500 --max-results 10
uv run gfi-scout-cli health fastapi/fastapi
uv run gfi-scout-cli status https://github.com/fastapi/fastapi/issues/12345
uv run gfi-scout-cli guide pallets/flask
uv run gfi-scout-tui

Every command supports --output json for scripts. See docs/CLI.md.

How scoring works

beginner_score = repo_health        × 0.30
               + issue_freshness    × 0.20
               + issue_clarity      × 0.15
               + merge_friendliness × 0.25
               + setup_complexity_inv × 0.10

Every weight and threshold is loaded from src/gfi_scout/data/scoring_weights.json. Want to retune the ranker? Edit the JSON and re-run — no code changes.

Full breakdown in docs/SCORING_ALGORITHM.md.

Documentation

Development

uv sync                              # install everything
uv run pytest                        # 106 tests in ~2 s
uv run ruff check src/ tests/        # lint
uv run ruff format src/ tests/       # format
uv run mypy src/                     # strict type-check
uv run mcp dev src/gfi_scout/server.py  # MCP Inspector

uv.lock is committed — every contributor gets identical dependency versions.

Project layout

src/gfi_scout/
├── server.py          # FastMCP entry, tool registration
├── cli.py             # Standalone CLI + terminal UI
├── config.py          # Env loading
├── runtime.py         # Shared cache/client wiring
├── tools/             # One file per MCP tool
│   ├── find_issues.py
│   ├── check_repo_health.py
│   ├── check_issue_status.py
│   └── get_contribution_guide.py
├── services/          # GitHub client, scoring, cache
├── models/            # Pydantic models
└── utils/             # Pure helpers (validators, rate limiter, logger)

tests/                 # mirrors src/ layout
docs/                  # markdown docs
scripts/               # dev automation (setup.sh, seed_cache.py)

The scoring config lives inside the package at src/gfi_scout/data/scoring_weights.json so it ships with the installed wheel — no separate top-level config/ directory. (The runtime settings module gfi_scout/config.py is unrelated; data/ holds JSON, config.py reads env vars.)

Layer rules and folder contracts: docs/ARCHITECTURE.md.

Troubleshooting

scoring config not found: .../Lib/config/scoring_weights.json

You're on an old install (≤ v0.1.0) where the scoring config wasn't bundled into the wheel. Fix:

uv tool install --force --from git+https://github.com/Rajveerx11/gfi-scout gfi-scout

If a long-running MCP server process holds the install directory open on Windows (Access is denied during reinstall), stop the host (Claude Desktop / Claude Code / Cursor) or kill the gfi-scout Python process first, then re-run the command.

Results are slow or you hit rate limit exceeded quickly

You're probably running without a token (60 requests/hour). Set GITHUB_TOKEN — from the environment or a .env file in the working directory — to get 5,000 requests/hour. For uv tool installs, either export it in your shell profile or set it in the MCP client's env block (see Connecting to an MCP client above).

Contributing

Issues and PRs welcome — practising what we preach. Start at CONTRIBUTING.md. Good first issues are labelled on the tracker.

By participating you agree to the Code of Conduct.

Security

Found a security issue? Please don't open a public issue — see SECURITY.md for the disclosure process.

GFI Scout only ever needs public_repo (read-only) scope on your GitHub token.

License

MIT — because the whole point is helping people contribute to open source.

Built with frustration, then determination. Because finding your first open source contribution shouldn't require a PhD in "how to navigate GitHub."