Context Tools MCP Server

Documentation tools for AI agents — fetch library docs, search semantically, scrape websites cleanly.

Features

• 📚 Library Docs - Instant access to 1000+ libraries via Context7 API
• 🔍 Semantic Search - Natural language Q&A over documentation targets
• 🌐 Web Scraping - Clean markdown extraction from any website
• 💾 Smart Caching - Automatic deduplication and content organization
• ⚡ Token Efficient - Optimized responses to avoid context bloat
• 🎯 Resource Discovery - MCP resources expose available targets for agents

Quick Start

Recommended: npx

npx -y github:ain3sh/context-mcp

No installation required. Always uses the latest version.

From Source

git clone https://github.com/ain3sh/context-mcp.git
cd context-mcp
npm install
npm run build
npm start

Installation

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "context-mcp": {
      "command": "npx",
      "args": ["-y", "github:ain3sh/context-mcp"],
      "env": {
        "GEMINI_API_KEY": "your_api_key_here"
      }
    }
  }
}

Claude Code (Project-Level)

Create .mcp.json in your project root:

{
  "mcpServers": {
    "context-mcp": {
      "type": "stdio", 
      "command": "npx",
      "args": ["-y", "github:ain3sh/context-mcp"],
      "env": {
        "GEMINI_API_KEY": "${GEMINI_API_KEY}"
      }
    }
  }
}

Tools Overview

Tool	Purpose	Token Usage
fetch_docs	Get library/framework documentation	~5k-50k tokens
fetch_site	Scrape webpage → clean markdown	~2k-20k tokens
ask_docs_agent	Semantic Q&A over documentation	~500-3k tokens

---

fetch_docs

Fetch documentation for any library or framework from Context7's massive index.

Usage Examples

// Simple library fetch
fetch_docs({
  target: "react"
})

// Specific topic within a library
fetch_docs({
  target: "pytorch",
  tag: "autograd"
})

// Control documentation depth
fetch_docs({
  target: "pandas",
  depth: "high"  // ~50k tokens for comprehensive docs
})

// Browse available versions
fetch_docs({
  target: "tensorflow",
  browse_index: true
})

Parameters

Param	Type	Default	Description
target	string	required	Library name (fuzzy matching supported)
tag	string	-	Topic filter within library
depth	"low" | "medium" | "high"	"medium"	Documentation comprehensiveness (~5k/15k/50k tokens)
version	string	-	Specific version to fetch
browse_index	boolean	false	List available libraries instead of fetching docs

How It Works

Query → Context7 API → Smart matching → Filtered docs → Markdown response

Context7 maintains pre-indexed documentation for 1000+ libraries. The tool uses fuzzy matching, so "tf" finds "tensorflow", "pd" finds "pandas", etc.

---

fetch_site

Extract clean, readable content from any website using Mozilla's Readability algorithm.

Usage Examples

// Single page extraction
fetch_site({
  url: "https://arxiv.org/abs/2301.00234"
})

// Include images
fetch_site({
  url: "https://example.com/tutorial",
  images: true
})

// Batch fetch multiple URLs
fetch_site({
  url: [
    "https://site.com/page1",
    "https://site.com/page2",
    "https://site.com/page3"
  ]
})

// Force refresh cached content
fetch_site({
  url: "https://news.site/article",
  refresh: true
})

Parameters

Param	Type	Default	Description
url	string | string[]	required	URL(s) to fetch (max 10 for batch)
images	boolean	false	Fetch and save images alongside content
refresh	boolean	false	Bypass cache and re-fetch

Output Structure

Content is saved to disk and returned:

./context/
└── understanding-react-hooks/   # Auto-named from page title
    ├── CONTENT.md               # Markdown with YAML frontmatter
    └── images/
        └── diagram.jpg

How It Works

URL → Fetch → JSDOM → Readability → Turndown → Clean Markdown

Uses the same extraction algorithm as Firefox Reader View to remove ads, navigation, and clutter.

---

ask_docs_agent

AI-powered semantic search over documentation stores using Gemini File Search.

Requires: GEMINI_API_KEY environment variable (get one here)

Usage Examples

// Simple question
ask_docs_agent({
  target: "context",
  query: "How does chunking work in File Search?"
})

// Include source chunks for verification
ask_docs_agent({
  target: "my-docs",
  query: "authentication setup",
  include_chunks: true,
  top_k: 5
})

// Get structured JSON response
ask_docs_agent({
  target: "api-docs",
  query: "rate limits",
  format: "json"
})

Parameters

Param	Type	Default	Description
query	string	required	Natural language question (5-500 chars)
target	string	required	Documentation store name
top_k	int	3	Number of relevant chunks to retrieve (1-20)
include_chunks	boolean	false	Include chunk previews in response
format	"markdown" | "json"	"markdown"	Response format
metadata_filter	string	-	Advanced: List Filter syntax for file filtering

Discovering Available Targets

Query the docs://targets MCP resource to see what's available:

{
  "targets": [
    { "target": "my-api-docs", "id": "...", "createTime": "...", "updateTime": "..." }
  ],
  "total": 1
}

Use docs://targets/{name} for details on a specific target.

Response Examples

Default (markdown, no chunks): ~500-1000 tokens

# Search Results: context

**Query**: How does chunking work?

**Response**:
Files are automatically chunked when imported into a file search store...

---

**Sources** (2 files):
  - ai.google.dev_gemini-api_docs_file-search.md
  - technical_spec.md

With chunks: ~2000-3000 tokens

[... same as above, plus ...]

---

## Retrieved Context Chunks

### [1] ai.google.dev_gemini-api_docs_file-search.md

Files are automatically chunked when imported...
[truncated to 500 chars]

How It Works

Query → Gemini File Search API → Semantic retrieval → LLM synthesis → Cited response

Queries pre-indexed Gemini FileSearchStores in the cloud. The stores must be created separately (e.g., via GitHub Actions workflow).

---

Configuration

Environment Variables

Variable	Required	Description
GEMINI_API_KEY	For ask_docs_agent	Your Gemini API key for semantic search
CONTEXT7_API_KEY	No	Higher rate limits for fetch_docs (optional)
FETCH_SITE_CONTENT_DIR	No	Content storage directory (default: ./context)
LOG_LEVEL	No	debug, info, or error (default: info)

Token Efficiency

This server is optimized to minimize context usage:

Tool	Typical Usage	Maximum
fetch_docs with depth="low"	~5k tokens	10k
fetch_docs with depth="medium"	~15k tokens	30k
fetch_docs with depth="high"	~50k tokens	100k
fetch_site (single page)	~2-5k tokens	20k
ask_docs_agent (default)	~500-1k tokens	2k
ask_docs_agent with chunks	~2-3k tokens	5k

Troubleshooting

fetch_docs: "Library not found"

• Try variations of the name (e.g., "tf" vs "tensorflow")
• Use browse_index: true to see available libraries
• Check if you need a specific version

fetch_site: "Failed to extract content"

• Some sites block automated access
• Try refresh: true to bypass cache
• Check if the site requires authentication

ask_docs_agent: "Target not found"

• Query docs://targets resource to see available targets
• Verify the target name matches exactly
• Check that GEMINI_API_KEY is set correctly

ask_docs_agent: Rate limits

• Free tier: ~15 queries per minute
• Wait 60 seconds between bursts
• Consider upgrading Gemini API tier if needed

MCP client doesn't show tools

• Run npm run build to compile TypeScript
• Check MCP config JSON is valid
• Review client logs (e.g., ~/Library/Logs/Claude/mcp*.log)
• Test manually: GEMINI_API_KEY=key npx -y github:ain3sh/context-mcp

Development

Local Development

# Install dependencies
npm install

# Build TypeScript
npm run build

# Run development mode (auto-reload)
npm run dev

# Test with environment
GEMINI_API_KEY=your_key npm start

Quick Test

npm run build
timeout 5s GEMINI_API_KEY=your_key npx .

Project Structure

context-mcp/
├── src/
│   └── index.ts         # MCP server implementation
├── dist/
│   └── index.js         # Compiled output
├── context/             # Cached content (git-ignored)
├── package.json
└── tsconfig.json

License

MIT License – see LICENSE