MCP GitHub Learning Assistant

An MCP (Model Context Protocol) server that helps developers discover and learn from open-source projects on GitHub, based on their skill level and goals.

Features

8 tools to support every stage of open-source learning:

Screenshots

MCP Inspector - Tools list	MCP Inspector - analyze_repo_for_learning result

Screenshots use dark-mode crops under 500KB each.

Installation

pip install mcp-github-learning-assistant

Requires Python 3.10+ and mcp>=2.3.0.

Configuration

Set your GitHub Personal Access Token:

# Windows PowerShell
$env:GITHUB_TOKEN = "your_github_token"

# Linux / macOS
export GITHUB_TOKEN=your_github_token

Token scope: fine-grained PAT with no special permissions (public repo read-only) is enough.

Entry Point

After pip install mcp-github-learning-assistant, the console command mcp-github-learning is available:

mcp-github-learning

This is equivalent to python -m mcp_github_learning_assistant.

Usage in MCP Clients

MCP Inspector (development)

npx @modelcontextprotocol/inspector mcp-github-learning

Cherry Studio

1. Open Cherry Studio, go to Settings -> MCP
2. Add an MCP server with the command mcp-github-learning
3. Set the environment variable GITHUB_TOKEN

Claude Desktop

Edit claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-github-learning": {
      "command": "mcp-github-learning",
      "env": {
        "GITHUB_TOKEN": "your_github_token"
      }
    }
  }
}

On Windows, if the command is not found, use the full path or fall back to python -m mcp_github_learning_assistant.

DSH / Dify

Add an MCP server of type stdio:

name: mcp-github-learning
transport: stdio
command: mcp-github-learning
env:
  GITHUB_TOKEN: your_github_token

If the client does not support custom env, set GITHUB_TOKEN at the system level before launching the client.

Current limitations

These are known and intentional, not bugs. They are listed so the tool's output is not over-trusted.

• No AI/LLM layer. Scoring is deterministic: star thresholds, README byte length, and directory-name keyword matching. There are no embeddings and no model calls.
• recommend_learning_repos uses a curated seed list, currently covering Spring Boot, FastAPI, Vue, React, Electron and HarmonyOS. Any other stack returns an empty list rather than a bad guess.
• Only get_trending_repos_for_learning derives its ranking from data — it computes stars / age_in_days and re-sorts, since GitHub has no public trending API.
• No HTTP caching, retry, backoff, or pagination. Unauthenticated GitHub allows 60 requests/hour, and one recommend_learning_repos call can spend several of them. Always set GITHUB_TOKEN.
• Some parameters are accepted but not yet used (depth in get_repo_roadmap, user_level comparison weighting in compare_learning_repos). They are reserved for the next iteration.
• Test coverage is partial. 9 tests cover analyze, roadmap, trending and info; the recommender, comparer, similar and issuefind modules have none.

Development

git clone https://github.com/iyiook333/mcp-github-learning-assistant.git
cd mcp-github-learning-assistant
pip install -e .[dev]
pytest

License

MIT