Context MCP Server

Overview

A self-hosted documentation RAG system that provides intelligent search across curated technical documentation. Built to replace cluttered tools like Context7 with a focused, manageable set of docs you actually use.

The Problem

Documentation search tools accumulate clutter:

• Context7 has thousands of docs you never use
• Finding relevant information takes too long
• No control over what’s indexed
• Stale documentation causes incorrect answers

AI assistants need access to current, curated documentation.

Solution

Complete control over your documentation index:

Key Capabilities

Hybrid Search

Three search modes for different needs:

Intelligent Scraping Pipeline

Multi-stage content processing:

1. URL Discovery - Firecrawl /map or browser crawling
2. Content Extraction - Firecrawl (primary), Browser Rendering, or fetch
3. Regex Cleanup - Remove nav, footers, UI chrome
4. AI Cleanup - Workers AI (Qwen 32B) removes marketing fluff
5. Semantic Chunking - Split at H2/H3 boundaries
6. Embedding - BGE-base-en-v1.5 via Workers AI
7. Indexing - Store in Vectorize for similarity search

Auto-Updates

Configurable refresh schedules:

• Hourly - Fast-changing API docs
• Daily - Most documentation
• Weekly - Stable reference material

Change detection via ETag/Last-Modified headers and content hashing.

MCP Tools (18 total)

Search:

• search_docs - Hybrid semantic + keyword search
• lookup_api - Exact function/class name lookup
• search_code - Find code examples

Source Management:

• list_sources - View indexed documentation
• add_source - Add new documentation URL
• update_source - Modify source config
• delete_source - Remove documentation
• scrape_source - Trigger re-indexing

Content Retrieval:

• get_chunk - Retrieve specific chunk by ID
• get_full_page - Get complete page content
• get_page_context - Get chunk with surrounding context

Discovery:

• get_related_chunks - Find similar content
• suggest_related - AI-powered suggestions
• compare_chunks - Compare two chunks

Architecture

Cloudflare Workers
├── Admin UI (React SPA)
├── REST API (Hono)
├── MCP Server (Durable Objects)
└── Queue Handler (async processing)

Storage:
├── D1 (metadata, sources, pages, chunks)
├── Vectorize (768-dim embeddings)
├── R2 (markdown exports)
└── KV (OAuth sessions)

Features

• 18 MCP Tools - Complete documentation access
• Hybrid Search - Vector + keyword for accuracy
• AI Cleanup - < 1% cruft rate in indexed content
• Auto-Updates - Cron-triggered re-indexing
• Admin UI - Visual source management
• OAuth - Google authentication with allowlist
• Markdown Export - Download docs as combined markdown

Quality Metrics

Production results from indexed sources:

Two-stage cleanup (regex + AI) achieves < 1% cruft across all sources.

Use Cases

Claude Code Integration - Add @context MCP server for instant doc search

API Lookups - Find exact function signatures and parameters

Learning New Frameworks - Search concepts across multiple doc sources

Code Examples - Find implementation patterns and snippets

Version Tracking - Keep docs current with auto-updates