mcp-qdrant-docs MCP Server

A Model Context Protocol server

This is a TypeScript-based MCP server that scrapes website content, indexes it into a Qdrant vector database, and provides a tool to answer questions about the indexed content.

使用例

ask_hono_docs: Hono.devのドキュメント内容について質問できます
ask_reactrouter_docs: ReactRouter.comのドキュメント内容について質問できます
ask_gradio_docs: Gradio.appのLLMsドキュメントについて質問できます

Features

Tool: ask_<doc name>_docs

• Name: Dynamically generated based on the DOCS_URL or --start-url (e.g., ask_reactrouter_docs).
• Functionality: Allows users to ask natural language questions about the content scraped from the specified website.
• Process:
  1. On startup (or if force-reindex is specified), the server scrapes the target website.
  2. The scraped content is processed, chunked, and embedded using a sentence transformer model.
  3. These embeddings and content chunks are stored in a Qdrant collection specific to the website.
  4. When the tool is called with a query, the server embeds the query, searches Qdrant for relevant chunks, and returns the found content as the answer.
• Input: A natural language query (string).
• Output: Text containing the relevant content chunks found in the Qdrant index.

Installation

To use with Claude Desktop, add the server config:

On MacOS: ~/Library/Application Support/Claude/claude_desktop_config.json On Windows: %APPDATA%/Claude/claude_desktop_config.json

npm i -g @kazuph/mcp-qdrant-docs

Using npx

The recommended way to run this server is using npx within your MCP client configuration (e.g., Claude Desktop's claude_desktop_config.json). This avoids the need for global installation.

Example claude_desktop_config.json using npx:

{
  "mcpServers": {
    "react-router-docs": { // A unique name for this server instance
      "command": "npx",
      "args": [
        "@kazuph/mcp-qdrant-docs", // The command registered in package.json bin
        // Optional: Add command-line arguments here if needed
        // "--start-url", "https://some-default-url.com/",
        // "--debug"
      ],
      // Optional: Set environment variables for configuration
      "env": {
        "DOCS_URL": "https://reactrouter.com/",
        "QDRANT_URL": "http://your-qdrant-instance:6333",
        "COLLECTION_NAME": "react-router-docs", // Base name for the collection
        "EMBEDDING_MODEL": "sentence-transformers/all-MiniLM-L6-v2"
        // "DEBUG": "true" // Alternative way to enable debug logging
      }
    }
    // You can add more server instances for different documentation sites here
  }
}

Command-Line Options

When running the server directly (e.g., using npx mcp-qdrant-docs or npm run dev --), you can use the following command-line options. These options override corresponding environment variables if both are set.

• --start-url <url> or -s <url>:
  • Required (if DOCS_URL env var is not set).
  • The starting URL of the website to scrape.
  • Overrides the DOCS_URL environment variable.
• --limit <number> or -l <number>:
  • Maximum number of pages to scrape.
  • Default: 300.
• --match <pattern> or -m <pattern>:
  • URL path patterns (prefix match) to limit scraping. Can be specified multiple times.
  • Example: --match /docs/ --match /api/
  • Default: Scrapes all pages under the start-url domain.
• --force-reindex:
  • Force re-scraping and re-indexing even if the Qdrant collection already exists.
  • Default: false.
• --collection-name <name> or -c <name>:
  • Base name for the Qdrant collection. The final collection name will be <base_name>-<sanitized_hostname>.
  • Overrides the COLLECTION_NAME environment variable.
  • Default: docs-collection.
• --qdrant-url <url>:
  • URL of the Qdrant instance.
  • Overrides the QDRANT_URL environment variable.
  • Default: http://localhost:6333.
• --embedding-model <model_name>:
  • Name of the sentence transformer model to use for embeddings (from Hugging Face or local).
  • Overrides the EMBEDDING_MODEL environment variable.
  • Default: Xenova/all-MiniLM-L6-v2.
• --debug:
  • Enable detailed debug logging.
  • Overrides the DEBUG environment variable (if set to true).
  • Default: false.
• --help or -h:
  • Show the help message listing all options.

Example using command-line options:

npx @kazuph/mcp-qdrant-docs --start-url https://example-docs.com/ --collection-name my-docs --limit 50 --debug

Configuration Priority:

The server uses the following priority for settings:

1. Command-line arguments: (e.g., --start-url, --collection-name) - Highest priority.
2. Environment variables: (e.g., DOCS_URL, COLLECTION_NAME) - Used if command-line arguments are not provided.
3. Default values: (Defined within the code) - Lowest priority.

Example: Adding React Router Documentation

To add a server instance specifically for querying React Router documentation, add the following entry to your mcpServers configuration (e.g., in claude_desktop_config.json):

{
  "mcpServers": {
    // ... other servers ...
    "react-router-docs": {
      "command": "npx", // Or the direct path if not installed globally
      "args": [
        "@kazuph/mcp-qdrant-docs"
        // No need to specify --start-url etc. if using env vars
      ],
      "env": {
        "DOCS_URL": "https://reactrouter.com/",
        "QDRANT_URL": "http://your-qdrant-instance:6333", // Replace with your Qdrant URL
        "COLLECTION_NAME": "react-router-docs", // Base name, will become 'react-router-docs-reactrouter_com'
        "EMBEDDING_MODEL": "sentence-transformers/all-MiniLM-L6-v2" // Or your preferred model
        // "DEBUG": "true" // Enable debug logs if needed
      }
    }
    // ... other servers ...
  }
}

Resulting Tool:

Once this server instance is running and connected to your MCP client, it will provide a tool named similar to ask_reactrouter_docs (not ask_reactrouter_com_docs).

• Tool Name: ask_<doc name>_docs (e.g., ask_reactrouter_docs)
• Description: Ask a question about the content of the site specified by DOCS_URL (or --start-url).
• Input: A natural language query about the documentation.

The server will automatically scrape the site (if the collection doesn't exist or --force-reindex is used), index the content into the specified Qdrant collection (react-router-docs-reactrouter_com in this example), and then use the index to answer your queries via the provided tool (ask_reactrouter_docs).