What is Mcp Qdrant Docs

mcp-qdrant-docs is an MCP server built with TypeScript that scrapes content from websites, indexes it into a Qdrant vector database, and provides a tool for querying that indexed content.

Use cases

Use cases include querying documentation for frameworks like Hono.dev, ReactRouter.com, and Gradio.app, allowing users to easily access specific information from these resources.

How to use

To use mcp-qdrant-docs, install it via npm or npx, configure it in your MCP client (like Claude Desktop), and use the dynamically generated tools to ask questions about the scraped documentation.

Key features

Key features include dynamic tool generation based on the target website’s URL, natural language query support, content scraping, embedding using a sentence transformer model, and efficient retrieval of relevant content from the Qdrant index.

Where to use

mcp-qdrant-docs can be used in various fields such as software development documentation, educational resources, and any domain where structured information needs to be indexed and queried effectively.

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:

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):

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).