Web Search MCP Server for use with Local LLMs

A TypeScript MCP (Model Context Protocol) server that provides comprehensive web search capabilities using direct connections (no API keys required) with multiple tools for different use cases.

Features

• Multi-Engine Web Search: Prioritises Bing > Brave > DuckDuckGo for optimal reliability and performance
• Full Page Content Extraction: Fetches and extracts complete page content from search results
• Multiple Search Tools: Four specialised tools for different use cases
• Smart Request Strategy: Switches between Playwright browsers and fast axios requests to ensure results are returned
• Concurrent Processing: Extracts content from multiple pages simultaneously

How It Works

The server provides four specialised tools for different web search needs:

1. full-web-search (Main Tool)

When a comprehensive search is requested, the server uses an optimised search strategy:

1. Browser-based Bing Search - Primary method using dedicated Chromium instance
2. Browser-based Brave Search - Secondary option using dedicated Firefox instance
3. Axios DuckDuckGo Search - Final fallback using traditional HTTP
4. Dedicated browser isolation: Each search engine gets its own browser instance with automatic cleanup
5. Content extraction: Tries axios first, then falls back to browser with human behavior simulation
6. Concurrent processing: Extracts content from multiple pages simultaneously with timeout protection
7. HTTP/2 error recovery: Automatically falls back to HTTP/1.1 when protocol errors occur

2. get-web-search-summaries (Lightweight Alternative)

For quick search results without full content extraction:

1. Performs the same optimised multi-engine search as full-web-search
2. Returns only the search result snippets/descriptions
3. Does not follow links to extract full page content

3. get-single-web-page-content (Utility Tool)

For extracting content from a specific webpage:

1. Takes a single URL as input
2. Follows the URL and extracts the main page content
3. Removes navigation, ads, and other non-content elements

4. clear-cache (Cache Management)

Flushes all cached search results and extracted page content from disk:

1. Deletes all files under CACHE_DIR
2. Returns the number of entries removed
3. Useful when you want fresh results instead of serving cached data

Compatibility

This MCP server has been developed and tested with LM Studio and LibreChat. It has not been tested with other MCP clients.

Model Compatibility

Important: Prioritise using more recent models designated for tool use.

Older models (even those with tool use specified) may not work or may work erratically. This seems to be the case with Llama and Deepseek. Qwen3 and Gemma 3 currently have the best restults.

• ✅ Works well with: Qwen3
• ✅ Works well with: Gemma 3
• ✅ Works with: Llama 3.2
• ✅ Works with: Recent Llama 3.1 (e.g 3.1 swallow-8B)
• ✅ Works with: Recent Deepseek R1 (e.g 0528 works)
• ⚠️ May have issues with: Some versions of Llama and Deepseek R1
• ❌ May not work with: Older versions of Llama and Deepseek R1

Installation (Recommended)

Requirements:

• Node.js 18.0.0 or higher
• npm 8.0.0 or higher

1. Download the latest release zip file from the Releases page

2. Extract the zip file to a location on your system (e.g., ~/mcp-servers/web-search-mcp/)

3. Open a terminal in the extracted folder and run:

   npm install
   npx playwright install
   npm run build

   This will create a node_modules folder with all required dependencies, install Playwright browsers, and build the project.

   Note: You must run npm install in the root of the extracted folder (not in dist/).

4. Configure your mcp.json to point to the extracted dist/index.js file:

{
  "mcpServers": {
    "web-search": {
      "command": "node",
      "args": ["/path/to/extracted/web-search-mcp/dist/index.js"]
    }
  }
}

Example paths:

• macOS/Linux: ~/mcp-servers/web-search-mcp/dist/index.js
• Windows: C:\\mcp-servers\\web-search-mcp\\dist\\index.js

Global Installation Without Cloning (npm)

If you publish this fork to npm, you can run the MCP server on any machine without cloning the repository.

1. Publish the package

npm login
npm publish

The package is configured for public publishing with automatic build at pack time.

2. Configure your MCP client

Example mcp.json:

{
  "mcpServers": {
    "web-search": {
      "command": "npx",
      "args": ["-y", "@cyberbobjr/web-search-mcp"]
    }
  }
}

With this configuration, the client downloads and runs the server automatically, without git clone or git pull.

In LibreChat, you can include the MCP server in the librechat.yaml. If you are running LibreChat in Docker, you must first mount your local directory in docker-compose.override.yml.

in docker-compose.override.yml:

services:
  api:
    volumes:
    - type: bind
      source: /path/to/your/mcp/directory
      target: /app/mcp

in librechat.yaml:

mcpServers:
  web-search:
    type: stdio
    command: node
    args:
    - /app/mcp/web-search-mcp/dist/index.js
    serverInstructions: true

Troubleshooting:

• If npm install fails, try updating Node.js to version 18+ and npm to version 8+
• If npm run build fails, ensure you have the latest Node.js version installed
• For older Node.js versions, you may need to use an older release of this project
• Content Length Issues: If you experience odd behavior due to content length limits, try setting "MAX_CONTENT_LENGTH": "10000", or another value, in your mcp.json environment variables:

{
  "mcpServers": {
    "web-search": {
      "command": "node",
      "args": ["/path/to/web-search-mcp/dist/index.js"],
      "env": {
        "MAX_CONTENT_LENGTH": "10000",
        "BROWSER_HEADLESS": "true",
        "MAX_BROWSERS": "3",
        "BROWSER_FALLBACK_THRESHOLD": "3"
      }
    }
  }
}

Environment Variables

The server supports several environment variables for configuration:

• MAX_CONTENT_LENGTH: Maximum content length in characters (default: 500000)
• DEFAULT_TIMEOUT: Default timeout for requests in milliseconds (default: 6000)
• MAX_BROWSERS: Maximum number of browser instances to maintain (default: 3)
• BROWSER_TYPES: Comma-separated list of browser types to use (default: 'chromium,firefox', options: chromium, firefox, webkit)
• BROWSER_FALLBACK_THRESHOLD: Number of axios failures before using browser fallback (default: 3)

Cache

Search results and extracted page content are cached to disk to avoid redundant network calls.

• CACHE_TTL_HOURS: Cache entry lifetime in hours (default: 72)
• CACHE_DIR: Directory where cache files are stored (default: ~/.cache/web-search-mcp/)
• CACHE_DISABLED: Set to 1 or true to disable the cache entirely (default: unset)
• CACHE_DEBUG: Set to true to log cache hits and misses to stderr (default: false)

Cache files are JSON stored under <CACHE_DIR>/search/ and <CACHE_DIR>/content/. Use the clear-cache MCP tool to flush all entries on demand.

Search Quality and Engine Selection

• ENABLE_RELEVANCE_CHECKING: Enable/disable search result quality validation (default: true)
• RELEVANCE_THRESHOLD: Minimum quality score for search results (0.0-1.0, default: 0.3)
• FORCE_MULTI_ENGINE_SEARCH: Try all search engines and return best results (default: false)
• DEBUG_BROWSER_LIFECYCLE: Enable detailed browser lifecycle logging for debugging (default: false)
• DEBUG_EXTRACTION: Enable verbose per-URL content extraction logging (default: false)

Troubleshooting

Slow Response Times

• Optimised timeouts: Default timeout reduced to 6 seconds with concurrent processing for faster results
• Concurrent extraction: Content is now extracted from multiple pages simultaneously
• Reduce timeouts further: Set DEFAULT_TIMEOUT=4000 for even faster responses (may reduce success rate)
• Use fewer browsers: Set MAX_BROWSERS=1 to reduce memory usage

Search Failures

• Check browser installation: Run npx playwright install to ensure browsers are available
• Try headless mode: Ensure BROWSER_HEADLESS=true (default) for server environments
• Network restrictions: Some networks block browser automation - try different network or VPN
• HTTP/2 issues: The server automatically handles HTTP/2 protocol errors with fallback to HTTP/1.1

Search Quality Issues

• Enable quality checking: Set ENABLE_RELEVANCE_CHECKING=true (enabled by default)
• Adjust quality threshold: Set RELEVANCE_THRESHOLD=0.5 for stricter quality requirements
• Force multi-engine search: Set FORCE_MULTI_ENGINE_SEARCH=true to try all engines and return the best results

Memory Usage

• Automatic cleanup: Browsers are automatically cleaned up after each operation to prevent memory leaks
• Limit browsers: Reduce MAX_BROWSERS (default: 3)
• EventEmitter warnings: Fixed - browsers are properly closed to prevent listener accumulation

For Development

git clone https://github.com/mrkrsl/web-search-mcp.git
cd web-search-mcp
npm install
npx playwright install
npm run build

Development

npm run dev    # Development with hot reload
npm run build  # Build TypeScript to JavaScript
npm run lint   # Run ESLint
npm run format # Run Prettier

Local Testing with MCP Inspector

The MCP Inspector is the official interactive UI for testing MCP servers locally without needing a full LLM client.

Prerequisites

Build the project first:

npm run build

Run the Inspector

npx @modelcontextprotocol/inspector node dist/index.js

This opens a web UI at http://localhost:5173 (or the port shown in the terminal). From there you can:

• Browse the four available tools (full-web-search, get-web-search-summaries, get-single-web-page-content, clear-cache)
• Fill in arguments and invoke them interactively
• Inspect the raw JSON responses

Pass Environment Variables

MAX_CONTENT_LENGTH=20000 BROWSER_HEADLESS=true \
  npx @modelcontextprotocol/inspector node dist/index.js

Quick Tool Examples

Search for something:

{ "query": "MCP protocol specification", "limit": 3 }

Extract a single page:

{ "url": "https://modelcontextprotocol.io", "maxContentLength": 5000 }

Using tsx (skips the build step)

If you want to test source changes without rebuilding:

npx @modelcontextprotocol/inspector npx tsx src/index.ts

MCP Tools

This server provides four specialised tools for different web search needs:

1. full-web-search (Main Tool)

The most comprehensive web search tool that:

1. Takes a search query and optional number of results (1-10, default 5)
2. Performs a web search (tries Bing, then Brave, then DuckDuckGo if needed)
3. Fetches full page content from each result URL with concurrent processing
4. Returns structured data with search results and extracted content
5. Enhanced reliability: HTTP/2 error recovery, reduced timeouts, and better error handling

Example Usage:

{
  "name": "full-web-search",
  "arguments": {
    "query": "TypeScript MCP server",
    "limit": 3,
    "includeContent": true
  }
}

2. get-web-search-summaries (Lightweight Alternative)

A lightweight alternative for quick search results:

1. Takes a search query and optional number of results (1-10, default 5)
2. Performs the same optimised multi-engine search as full-web-search
3. Returns only search result snippets/descriptions (no content extraction)
4. Faster and more efficient for quick research

Example Usage:

{
  "name": "get-web-search-summaries",
  "arguments": {
    "query": "TypeScript MCP server",
    "limit": 5
  }
}

3. get-single-web-page-content (Utility Tool)

A utility tool for extracting content from a specific webpage:

1. Takes a single URL as input
2. Follows the URL and extracts the main page content
3. Removes navigation, ads, and other non-content elements
4. Useful for getting detailed content from a known webpage

Example Usage:

{
  "name": "get-single-web-page-content",
  "arguments": {
    "url": "https://example.com/article",
    "maxContentLength": 5000
  }
}

4. clear-cache (Cache Management)

Flushes all cached data from disk. Takes no arguments.

Example Usage:

{
  "name": "clear-cache",
  "arguments": {}
}

Standalone Usage

You can also run the server directly:

# If running from source
npm start

Documentation

See API.md for complete technical details.

License

MIT License - see LICENSE for details.

Feedback

This is an open source project and we welcome feedback! If you encounter any issues or have suggestions for improvements, please:

• Open an issue on GitHub
• Submit a pull request