OmniDocs MCP Server

OmniDocs MCP

OmniDocs MCP is an intelligent Model Context Protocol (MCP) server that empowers AI agents to instantly search, index, summarize, and inject live framework documentation directly into their context window.

Stop hallucinating code for outdated framework versions. Let OmniDocs fetch the exact documentation your AI needs, the moment it needs it.

[AI Agent] Calling get_library_docs("react", "useActionState usage")
[OmniDocs] 🔍 Library 'react' not indexed. Crawling react.dev...
[OmniDocs] 🧠 Chunking 150 pages and computing embeddings (ONNX)...
[OmniDocs] ⚡ Returning top 5 semantic chunks (Dense + BM25)
[AI Agent] Receives 1,500 highly-relevant tokens. Writes perfect code.

🤔 Why OmniDocs? (The Comparison)

✨ Core Features

• Deep HTML Crawling: Employs an Indexer & Sniper architecture to map deep documentation sites via XML Sitemaps or pure HTML-crawling, returning dense Tables of Contents for agents to navigate.
• Local RAG & Semantic Search: Embeds documentation locally using ONNX (via fastembed) and chunks it semantically. Exposes a natural language query interface so agents receive precise, high-density excerpts instead of full massive pages.
• Local Manifest Auto-Discovery: Point OmniDocs at any package.json or requirements.txt. It will seamlessly communicate with the NPM/PyPI registries to auto-discover library documentation URLs and register them in its tracking file.
• Persistent Disk Caching: Prevents excessive redundant scraping and LLM token usage by storing fetched markdown via diskcache, allowing the user to configure granular TTLs (Time-To-Live).

🏗 Architecture & How it Works

OmniDocs operates as a middleware server between an AI Agent and official documentation websites. Instead of the AI browsing the web blindly, it uses OmniDocs to precisely retrieve, parse, chunk, embed, and cache documentation locally.

Core Modules

1. Server CLI (server.py): The main entry point. Exposes get_library_docs which agents use to ask natural language questions.
2. Fetcher (fetcher.py): Handles outbound HTTP requests, crawling Sitemaps and pure HTML. Uses BeautifulSoup to strip away navbars and footers, and markdownify to convert perfectly to Markdown.
3. Chunker (chunker.py): Splits massive Markdown pages into smaller, semantically coherent 512-token chunks, keeping Markdown headers intact so the context isn't lost.
4. Vector Store (vector_store.py): Embeds chunks locally using the fastembed ONNX model and stores them persistently in ChromaDB. Uses a hybrid retrieval method (Dense Vector Search + BM25 keyword re-ranking) for maximum accuracy on exact API names.
5. Cache Layer (cache.py): Uses diskcache to store the raw downloaded Markdown on the local hard drive to prevent redundant network requests.
6. Auto-Discovery (discovery.py): Parses local package.json or requirements.txt files to auto-register libraries.

🔄 Retrieval Workflow

When an AI encounters a library it doesn't know, it just issues a natural language query, and the following flow occurs:

sequenceDiagram
    participant AI as AI Agent
    participant MCP as OmniDocs
    participant VectorDB as ChromaDB
    participant Web as fetcher.py
    
    AI->>MCP: Call `get_library_docs("react", "useActionState usage")`
    MCP->>VectorDB: Check if 'react' is indexed
    
    alt Not Indexed
        MCP->>Web: Crawl entire doc site & convert to Markdown
        Web-->>MCP: Return Markdown pages
        MCP->>MCP: Chunk pages & compute local embeddings
        MCP->>VectorDB: Store chunks & vectors
    end
    
    MCP->>VectorDB: Perform hybrid search (Dense + BM25) for query
    VectorDB-->>MCP: Top 5 semantic chunks
    MCP-->>AI: Return pure, precise Markdown context

🚀 Quick Start

Prerequisites

• Python 3.10+ (Required for FastMCP and ChromaDB)
• OS: Windows, macOS, or Linux
• Hardware: Runs entirely