SourceTap

SourceTap is an MVP of a Model Context Protocol (MCP server that lets your AI assistant learn and search any library directly from its GitHub repository or documentation URL.

Features

This project provides two tools:

1. query_docs(url, query): A RAG (Retrieval-Augmented Generation) tool.

   • Input: Takes a URL to a ZIP archive (e.g., a GitHub repo archive) and a search query.

   • Process:

     • Downloads the ZIP file (cached via SQLite to prevent redundant downloads).

     • Extracts .md and .mdx content.

     • Indexes the content in-memory using minsearch (TF-IDF/Keyword search).

   • Output: Returns the full content of the top 5 most relevant documentation files.

   • Use Case: Helps AI agents understand libraries that are too new, private, or obscure for their base references.

2. fetch_web_content(url): A reader tool.

   • Input: Any webpage URL.

   • Process: Proxies the request through r.jina.ai to convert HTML to clean, LLM-friendly Markdown.

   • Output: The text content of the page.

   • Use Case: Inspecting specific documentation pages, blog posts, or issue threads.

Related MCP server: Search Docs MCP

Installation

To use this tool with your AI assistant (e.g., Claude Desktop, Cline), add the following configuration to your MCP Settings file:

{
  "mcpServers": {
    "sourcetap": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/sourcetap",
        "run",
        "python",
        "main.py"
      ]
    }
  }
}

Note: Replace /absolute/path/to/sourcetap with the actual path to this directory on your machine. The uv command will automatically handle dependency installation and environment setup when the server starts.

Project Architecture

The Tech Stack

• MCP Framework: FastMCP (Python)

• Web Scraping: Jina Reader API (via httpx)

• Search Engine: minsearch (TF-IDF/Keyword search)

• Caching: SQLite with WAL mode

• Used MCP: Context7

• AI Assistant: Google Gemini 3 Flash + Antigravity IDE

Caching Strategy

The project uses SQLite for persistent caching of downloaded ZIP files.

• WAL Mode: Write-Ahead Logging enabled for better concurrent read/write performance.

Search Implementation

Uses minsearch for in-memory document search.

• Text Fields: Indexes both content and filename for comprehensive search.

• TF-IDF Scoring: Ranks documents by term frequency-inverse document frequency.

• Top-K Retrieval: Returns the 5 most relevant documents per query.

• Memory Efficient: Index is rebuilt per query (no persistent index storage).

Limitations & Possible Improvements

• Keyword-Only Search: Currently uses TF-IDF. Semantic search with embeddings (e.g., all-MiniLM-L6-v2) would enable conceptual matching.

• Full-File Retrieval: Returns entire files. Smart chunking by headers would improve precision.

• Markdown-Only: Only indexes .md and .mdx files. Code parsing (.py, .ts) would enable technical implementation queries.

• ZIP Archives: Downloads full repositories. GitHub Tree API would enable sparse downloading of only needed files.

• No Persistent Index: Index is rebuilt per query. Persistent indexing would improve performance for repeated queries.

• Single-Threaded Cache: SQLite cache is synchronous. Async cache operations would improve throughput.