SourceTap
Allows AI agents to learn and search any library directly from its GitHub repository by downloading and indexing documentation files (Markdown/MDX) and performing keyword search.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SourceTap search the docs of https://github.com/tiangolo/fastapi for 'dependency injection'"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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:
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
.mdand.mdxcontent.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.
fetch_web_content(url): A reader tool.Input: Any webpage URL.
Process: Proxies the request through
r.jina.aito 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/sourcetapwith the actual path to this directory on your machine. Theuvcommand 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
contentandfilenamefor 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
.mdand.mdxfiles. 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.
Available Tools
2 toolsfetch_web_contentC
Download content of any web page using Jina reader.
Args: url: The URL of the web page to fetch content from.
Inputs: url (required). Returns: result
query_docsA
Index and search documentation from a ZIP archive URL.
Args: url: URL to a ZIP file (e.g. GitHub archive) query: Search query
Inputs: url (required), query (required). Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 2 tools
The two tools have clearly distinct purposes: fetch_web_content retrieves any web page, while query_docs specifically indexes and searches documentation from a ZIP archive. Although both accept URLs, the actions and outputs are unambiguous.
Both tools follow the same verb_noun pattern with lower_snake_case: fetch_web_content and query_docs. The verbs differ but the structure is consistent and predictable.
With only two tools, the server feels quite thin. This is borderline for a useful MCP server, though the tools themselves are substantial and well-defined.
The server covers two distinct capabilities: web content retrieval and documentation search. Potential gaps include lacking a way to list or manage indexed docs and no broader web search beyond direct URL fetching, but the core scope is reasonably served.
Maintenance
Related MCP Connectors
Versioned documentation registry and semantic search for AI tools and coding assistants.
@latest documentation and code examples to 9000+ libraries for LLMs and AI code editors in a singl…
A cited wiki of your GitHub repo: search, read pages, find symbols and ask, with line citations.
Code intelligence for LLMs. Analyze, search, and retrieve code from any public git repository.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceScrapes and indexes documentation websites to provide AI assistants with searchable access to documentation content, API references, and code examples through configurable URL crawling.ISC
- FlicenseAqualityDmaintenanceEnables semantic search across multiple AI library documentations to keep coding assistants up-to-date.11-
- FlicenseNot gradedqualityDmaintenanceEnables retrieval and cleaning of official documentation for AI/Python libraries using search and LLM-based HTML cleaning.-
- AlicenseAqualityDmaintenanceProvides AI assistants with up-to-date documentation for popular libraries and frameworks, enabling them to generate more accurate code using less common or newly released libraries.536MIT