github-talent-mcp

MCP server that searches, scores, and ranks GitHub developers for technical recruiting.

Works with Claude (Code & Desktop) and GitHub Copilot (CLI & desktop app) — any MCP client that speaks stdio.

Demo

https://github.com/user-attachments/assets/b2dbe9e0-26ee-4849-861a-4b5cb268facc

Sourcing candidates for a real Anthropic JD, live in Claude Cowork.

https://github.com/user-attachments/assets/2dfd82b4-3eb5-4f2b-bc0a-2580b95043e4

Profile deep dive

Get the full developer profile and activity score for torvalds on GitHub

Claude calls get_developer_profile("torvalds") and returns:

Field	Value
Activity Score	150 (reputation floor applied)
Location	Portland, OR
Followers	293,321
Stars Received	235,068
Primary Language	C (98.1%)
Commits (90d)	0
PRs (90d)	0
Notable Repos	linux (183K stars), libdc-for-dirk, subsurface-for-dirk, uemacs, pesern-resolve
Profile README	No
Hireable	No

Torvalds has zero recent GitHub activity because kernel development flows through mailing lists, not GitHub PRs. The reputation floor (293K followers) overrides the behavioral score and sets it to 150.

Repo contributor ranking

Get the top contributors to huggingface/transformers and rank them for a founding ML engineer role at an AI startup

Claude calls get_repo_contributors("huggingface/transformers") → rank_candidates on the top 24 contributors:

Rank	Developer	Combined Score	Activity	Relevance	Strengths
1	stas00	83.4	150	72	4,553 stars, contributes to major OSS, MIT-licensed repos
2	cyyever	80.8	120	64	1,217 followers, active contributor, profile README
3	Cyrilvallez	77.2	120	56	Active: 13 commits + 57 PRs in 90 days, strong OSS presence
4	ArthurZucker	74.4	120	48	37 PRs in 90 days, contributes to huggingface/transformers
5	ydshieh	72.0	120	40	Active: 9 commits + 40 PRs in 90 days

Combined score = activity × 0.4 + relevance × 0.6. Relevance is keyword overlap with the job description (ML, AI, startup, engineer, etc.).

Installation

1. Install uv

The server runs through uvx, which downloads and launches it for you — no clone, no virtualenv, and you get updates automatically.

brew install uv

No Homebrew? curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create a GitHub personal access token

Without a token GitHub allows 60 requests per hour, and a single candidate profile costs 6-15 of them. You will run out mid-search and profiles will come back empty. With a token you get 5,000/hour.

Go to github.com/settings/tokens and create a fine-grained or classic token with these scopes:

Scope	Why
read:user	Read user profiles and search users
public_repo	Read public repo data, languages, contributors

Copy the token — you cannot view it again after leaving the page.

3. Connect it

Put the token itself in the config, not ${GITHUB_TOKEN}. Desktop apps are launched by the operating system, not by your shell, so they never read .zshrc and an environment variable reference expands to nothing. The server then starts fine, runs unauthenticated, and quietly fails a few candidates in. A .env file has the same problem unless the config also sets cwd to the project directory, because it is read relative to the working directory.

GitHub Copilot (CLI and desktop app)

Both share one config. Paste this in a terminal — it fills in your token for you:

mkdir -p ~/.copilot
TOKEN=$(gh auth token)   # or: TOKEN=github_pat_xxxxxxxx
cat > ~/.copilot/mcp-config.json <<EOF
{
  "mcpServers": {
    "github-talent": {
      "type": "local",
      "command": "uvx",
      "args": ["github-talent-mcp"],
      "env": { "GITHUB_TOKEN": "$TOKEN" },
      "tools": ["*"]
    }
  }
}
EOF
chmod 600 ~/.copilot/mcp-config.json

Quit Copilot completely and reopen it, then run /mcp show — you should see 9 tools under github-talent. The app also accepts servers under Settings → MCP if you would rather not touch a file.

If uvx is not found, give its full path as command (which uvx prints it).

Claude Code

claude mcp add github-talent --env GITHUB_TOKEN=github_pat_xxxxxxxx -- uvx github-talent-mcp

Restart Claude Code and verify with /mcp.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "github-talent": {
      "command": "uvx",
      "args": ["github-talent-mcp"],
      "env": {
        "GITHUB_TOKEN": "github_pat_xxxxxxxx"
      }
    }
  }
}

Restart Claude Desktop.

Checking it actually works

Run a search and look at the size of each get_developer_profile result. A real profile is 120-170 lines. Three lines means the call failed — almost always a missing or unreadable token. Every tool returning three lines while the server still shows as connected is the signature of running unauthenticated.

Running from source

Only needed if you want to modify the server:

git clone https://github.com/carolinacherry/github-talent-mcp.git
cd github-talent-mcp
uv sync

Then use uv run --directory /path/to/github-talent-mcp github-talent-mcp as the command in any config above.

Try It

Once installed, paste these prompts to verify everything works:

Basic search:

Find Python developers in Raleigh active in the last 60 days

Profile deep dive:

Get the full developer profile and activity score for torvalds on GitHub

Full workflow:

Find 10 ML engineers in San Francisco active in the last 30 days, then rank them for a senior LLM inference engineer role

Repo contributors:

Get the top contributors to huggingface/transformers and rank them for a founding ML engineer role at an AI startup

JD scoring:

Score these candidates against this job description: [paste JD]. Candidates: tiangolo, karpathy, hwchase17

Compare candidates:

Compare tiangolo and hwchase17 for a Senior Python AI Engineer role

Bulk scoring:

Score these 10 GitHub usernames and give me a ranked table: [paste list]

Outreach:

Generate a casual recruiter message for tiangolo about a Senior Python role at Acme. My name is Daniel.

Interview-first sourcing

Vague prompts produce vague shortlists, so the server is built to interview you before it searches. Ask it to "find candidates for a role" and it calls plan_search first — it detects the role family and asks targeted follow-ups (seniority, must-have skills, location, dealbreakers) and, most importantly, for the job description: paste the full text, or share a public link and paste what it shows. It only sources once it has real criteria.

Try it: "Find me senior security engineers." → the assistant should ask for the JD and your must-haves before running anything.

Want a fast, repeatable run instead? Give it everything up front — "Rank these 15 usernames against this JD: …" — or pin the sourcing to specific repos, and it'll skip the interview.

Tools

Tool	Description
plan_search	Intake step — parses a sourcing request, detects the role family, and returns targeted follow-up questions (including: paste the JD or share a public link) to ask before searching. Call this first.
search_developers	Search GitHub users by language, location, activity, followers. For topic-based sourcing, use get_repo_contributors on relevant repos instead.
get_developer_profile	Deep profile enrichment: languages, stars, commits + PRs, OSS contributions, license breakdown, profile README, and activity score with breakdown.
rank_candidates	Rank usernames against a job description. Returns sorted candidates with combined score, strengths, gaps, and reasoning.
score_against_jd	Score candidates against a JD with per-dimension breakdown (tech stack, experience level, OSS signal, leadership). Returns gaps and personalized interview questions.
compare_candidates	Side-by-side comparison of 2-5 candidates. Shows dimension winners and a recommendation. Optionally scored against a JD.
bulk_score	Score up to 100 GitHub usernames in one call. Returns a ranked markdown table or CSV. Supports optional JD matching.
generate_outreach	Generate personalized recruiter messages (short/medium/detailed) that reference the candidate's actual repos and contributions. Requires your company name and sender name. Casual or formal tone.
get_repo_contributors	Top contributors for any repo. Accepts owner/repo or full URL. The fastest way to source for a specific domain.

Scoring

The activity score combines two layers: behavioral signals (what you did recently) and a reputation floor (what you've built over time).

Behavioral Score (0-205)

Signal	Max Points	How
Commits + PRs (last 90 days)	60	Push commits + PR opens (PRs weighted x3). Captures both push-based and PR-based workflows.
Stars on repos	40	Personal repo stars + stars on repos you contribute to. Org repo maintainers get credit.
Profile README	20	Presence of a profile README (github.com/username/username).
Followers	20	Capped at 20.
Repos with descriptions	20	Ratio of repos that have descriptions. Signal of care and polish.
Permissive license repos	15	Has at least one repo with MIT, Apache-2.0, BSD, ISC, or Unlicense.
Major OSS contributions	30	PRs, pushes, or issues on repos you don't own. Capped at 3 repos (10 pts each).

Reputation Floor

The behavioral score alone penalizes developers whose work doesn't produce GitHub events — Torvalds works through mailing lists, senior maintainers merge via org bots, and many engineers work in private repos.

The reputation floor ensures cumulative impact isn't erased by a quiet quarter:

Threshold	Floor
10K+ followers or 50K+ stars	150
1K+ followers or 5K+ stars	120
500+ followers or 1K+ stars	100
100+ followers or 200+ stars	80

The final score is max(behavioral_score, reputation_floor). If the floor is applied, the breakdown includes a reputation_floor field so you know.

Score Tiers

• 150+ — exceptional (top OSS maintainers, well-known engineers)
• 120-149 — strong signal, worth reaching out
• 80-119 — solid developer with meaningful public work
• 40-79 — active but limited public signal
• <40 — low signal (likely private work or junior)

Ranking

rank_candidates combines the activity score with a relevance score (0-100) based on keyword overlap between the job description and the candidate's profile (bio, languages, repo topics, README). The combined score weights relevance at 60% and activity at 40% — a high-activity developer with no overlap to the job shouldn't outrank a relevant one.

Interactive dashboard

After a search produces a shortlist, the server asks whether you want an interactive dashboard — search, skill filters, ranking, evidence, and GitHub profile links. Answer yes and your assistant builds it with its own artifact tooling (Copilot's canvas, Claude's artifacts) from the scored candidate data.

It only ever offers; nothing is built unless you say yes, and the offer is skipped when a search produced no usable profiles. Set GITHUB_TALENT_DASHBOARD_PROMPT=0 to turn it off.

If the page opens in an inline canvas, note that those panes sandbox their content and block outbound links, so the assistant is also asked to open the saved file in your browser where the GitHub links work.

Rate Limits

GitHub REST API: 5,000 requests/hour with a token, 60 without one. A single enriched profile costs 6-15 calls and a typical workflow (search + enrich 5 candidates + rank) uses ~60-100, so an unauthenticated server runs out inside one search. Profile results are cached within a session to avoid redundant calls during ranking.

Two limits are separate from that hourly budget and worth knowing:

• Search endpoints (/search/commits, /search/issues) allow only 30 requests/minute even with a token. The server treats a failure there as an unknown activity count rather than a failed profile, so a shortlist still comes back — the commit counts may just read 0.
• Secondary rate limits fire on bursts of concurrent requests and return an explicit Retry-After. The server waits exactly that long, up to 30 seconds, then gives up rather than retrying into a window that has not lifted.

Limitations & responsible use

This tool scores public GitHub activity as one signal for technical sourcing. Know its limits before you rely on it:

• Results vary between runs. It's AI-driven — the assistant decides which repos and searches to explore, so the same prompt can surface a different shortlist each time. The scoring itself is deterministic for a given set of candidates; the variation comes from sourcing. For repeatable runs, constrain the sourcing: name the repos to pull contributors from, or hand it an explicit list of usernames to rank.
• GitHub is not the whole engineer. Public activity is strong evidence of technical work but blind to private-repo and internal/enterprise contributions, and to non-GitHub ecosystems (mailing lists, GitLab, etc.). It cannot verify people-management or leadership history — confirm those off-GitHub. (The reputation floor exists precisely because low recent activity ≠ low capability.)
• Use it as a lead generator, not a filter. Public OSS visibility correlates with free time, tenure, and circumstance — not just skill — and that skews across demographics. Treat scores as a starting point for outreach and human judgment. Don't use them to automatically exclude candidates, and always pair them with equitable, role-relevant evaluation.
• Data is live and rate-limited. Scores reflect GitHub at query time and shift as activity changes; an unauthenticated server is capped at 60 requests/hour.

License

Apache License 2.0 © 2026 Daniel An. Released versions up to and including 0.4.0 remain under the MIT License; 0.4.1 onward is Apache-2.0.