arabold/docs-mcp-server

Grounded Docs: Your AI's Up-to-Date Documentation Expert

The open-source alternative to Context7, Nia, and Ref.Tools.

AI coding assistants often struggle with outdated documentation and hallucinations. The Docs MCP Server solves this by providing a personal, always-current knowledge base for your AI. It indexes 3rd party documentation from various sources (websites, GitHub, npm, PyPI, local files) and offers powerful, version-aware search tools via the Model Context Protocol (MCP).

This enables your AI agent to access the latest official documentation, dramatically improving the quality and reliability of generated code and integration details. It's free, open-source, runs locally for privacy, and integrates seamlessly into your development workflow.

Why Use the Docs MCP Server?

LLM-assisted coding promises speed and efficiency, but often falls short due to:

• 🌀 Stale Knowledge: LLMs train on snapshots of the internet and quickly fall behind new library releases and API changes.
• 👻 Code Hallucinations: AI can invent plausible-looking code that is syntactically correct but functionally wrong or uses non-existent APIs.
• ❓ Version Ambiguity: Generic answers rarely account for the specific version dependencies in your project, leading to subtle bugs.
• ⏳ Verification Overhead: Developers spend valuable time double-checking AI suggestions against official documentation.

Docs MCP Server solves these problems by:

• ✅ Providing Up-to-Date Context: Fetches and indexes documentation directly from official sources (websites, GitHub, npm, PyPI, local files) on demand.
• 🎯 Delivering Version-Specific Answers: Search queries can target exact library versions, ensuring information matches your project's dependencies.
• 💡 Reducing Hallucinations: Grounds the LLM in real documentation for accurate examples and integration details.
• ⚡ Boosting Productivity: Get trustworthy answers faster, integrated directly into your AI assistant workflow.

✨ Key Features

• Accurate & Version-Aware AI Responses: Provides up-to-date, version-specific documentation to reduce AI hallucinations and improve code accuracy.
• Broad Source Compatibility: Scrapes documentation from websites, GitHub repos, package manager sites (npm, PyPI), and local file directories.
• Advanced Search & Processing: Intelligently chunks documentation semantically, generates embeddings, and combines vector similarity with full-text search.
• Flexible Embedding Models: Supports various providers including OpenAI (and compatible APIs), Google Gemini/Vertex AI, Azure OpenAI, and AWS Bedrock. Vector search is optional.
• Enterprise Authentication: Optional OAuth2/OIDC authentication with dynamic client registration for secure deployments.
• Web Interface: Easy-to-use web interface for searching and managing documentation.
• Local & Private: Runs entirely on your machine, ensuring data and queries remain private.
• Free & Open Source: Community-driven and freely available.
• Simple Deployment: Easy setup via Docker or npx.
• Seamless Integration: Works with MCP-compatible clients (like Claude, Cline, Roo).

What is semantic chunking?

Semantic chunking splits documentation into meaningful sections based on structure—like headings, code blocks, and tables—rather than arbitrary text size. Docs MCP Server preserves logical boundaries, keeps code and tables intact, and removes navigation clutter from HTML docs. This ensures LLMs receive coherent, context-rich information for more accurate and relevant answers.

How to Run the Docs MCP Server

Choose your deployment method:

• Standalone Server (Recommended)
• Embedded Server
• Advanced: Docker Compose (Scaling)

Standalone Server (Recommended)

Run a standalone server that includes both MCP endpoints and web interface in a single process. This is the easiest way to get started.

Option 1: Docker

1. Install Docker.

2. Start the server:

   docker run --rm \
     -v docs-mcp-data:/data \
     -v docs-mcp-config:/config \
     -p 6280:6280 \
     ghcr.io/arabold/docs-mcp-server:latest \
     --protocol http --host 0.0.0.0 --port 6280
   

   Configuration: The server writes its configuration to /config/docs-mcp-server/config.yaml. Mounting the /config volume ensures your settings persist across restarts.

   Optional: Add -e OPENAI_API_KEY="your-openai-api-key" to enable vector search for improved results.

Option 2: npx

1. Install Node.js 20.x or later.

2. Start the server:

   npx @arabold/docs-mcp-server@latest
   

   This runs the server on port 6280 by default.

   Optional: Prefix with OPENAI_API_KEY="your-openai-api-key" to enable vector search for improved results.

Configure Your MCP Client

Add this to your MCP settings (VS Code, Claude Desktop, etc.):

{
  "mcpServers": {
    "docs-mcp-server": {
      "type": "sse",
      "url": "http://localhost:6280/sse",
      "disabled": false,
      "autoApprove": []
    }
  }
}

Alternative connection types:

// SSE (Server-Sent Events)
"type": "sse", "url": "http://localhost:6280/sse"

// HTTP (Streamable)
"type": "http", "url": "http://localhost:6280/mcp"

Restart your AI assistant after updating the config.

Access the Web Interface

Open http://localhost:6280 in your browser to manage documentation and monitor jobs.

CLI Usage with Standalone Server

You can also use CLI commands to interact with the local database:

# List indexed libraries
OPENAI_API_KEY="your-key" npx @arabold/docs-mcp-server@latest list

# Search documentation
OPENAI_API_KEY="your-key" npx @arabold/docs-mcp-server@latest search react "useState hook"

# Scrape new documentation (connects to running server's worker)
npx @arabold/docs-mcp-server@latest scrape react https://react.dev/reference/react --server-url http://localhost:8080/api

Adding Library Documentation

1. Open the Web Interface at http://localhost:6280.
2. Use the "Add New Documentation" form.
3. Enter the documentation URL, library name, and (optionally) version.
4. Click "Start Indexing". Monitor progress in the Job Queue.
5. Repeat for each library you want indexed.

Once a job completes, the docs are searchable via your AI assistant or the Web UI.

Benefits:

• Single command setup with both web UI and MCP server
• Persistent data storage (Docker volume or local directory)
• No repository cloning required
• Full feature access including web interface

To stop the server, press Ctrl+C.

Configuration overrides

• Configuration Precedence: Configuration is loaded in the following order (last one wins):

  For a complete reference of all configuration options, see the .

  1. Defaults: Built-in default values.
  2. Config File: config.json or config.yaml in global store, project root, or current directory.
  3. Environment Variables: Specific DOCS_MCP_* variables override file settings.
  4. CLI Arguments: Command-line flags (e.g., --port) have the highest priority.

Configuration File

You can create a config.json or config.yaml file to persist your settings. The server searches for this file in:

1. The path specified by --config (Read-Only).
2. The path specified by DOCS_MCP_CONFIG environment variable (Read-Only).
3. The system default configuration directory (Read-Write):
   • macOS: ~/Library/Preferences/docs-mcp-server/config.yaml
   • Linux: ~/.config/docs-mcp-server/config.yaml (or defined by $XDG_CONFIG_HOME)
   • Windows: %APPDATA%\docs-mcp-server\config\config.yaml

Note: On startup, if no explicit configuration file is provided, the server will seek the system default config. If present, it loads it. If missing, it creates it with default values. It will also update it with any new setting keys. If you provide a custom config via --config or env var, the server treats it as Read-Only and will NOT modify it or write defaults back to it.

Example config.yaml:

server:
  host: "0.0.0.0"
  ports:
    mcp: 9000
    default: 8000
scraper:
  maxPages: 500
  pageTimeoutMs: 10000
splitter:
  maxChunkSize: 2000
embeddings:
  vectorDimension: 1536

Environment Variables

Specific configuration options can be set via environment variables. These override values from the configuration file.

Embedded Server

Run the MCP server directly embedded in your AI assistant without a separate process or web interface. This method provides MCP integration only.

Configure Your MCP Client

Add this to your MCP settings (VS Code, Claude Desktop, etc.):

{
  "mcpServers": {
    "docs-mcp-server": {
      "command": "npx",
      "args": ["@arabold/docs-mcp-server@latest"],
      "disabled": false,
      "autoApprove": []
    }
  }
}

Optional: To enable vector search for improved results, add an env section with your API key:

{
  "mcpServers": {
    "docs-mcp-server": {
      "command": "npx",
      "args": ["@arabold/docs-mcp-server@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-proj-..." // Your OpenAI API key
      },
      "disabled": false,
      "autoApprove": []
    }
  }
}

Restart your application after updating the config.

Adding Library Documentation

Option 1: Use MCP Tools

Your AI assistant can index new documentation using the built-in scrape_docs tool:

Please scrape the React documentation from https://react.dev/reference/react for library "react" version "18.x"

Option 2: Launch Web Interface

Start a temporary web interface that shares the same database:

OPENAI_API_KEY="your-key" npx @arabold/docs-mcp-server@latest web --port 6281

Then open http://localhost:6281 to manage documentation. Stop the web interface when done (Ctrl+C).

Option 3: CLI Commands

Use CLI commands directly (avoid running scrape jobs concurrently with embedded server):

# List libraries
OPENAI_API_KEY="your-key" npx @arabold/docs-mcp-server@latest list

# Search documentation
OPENAI_API_KEY="your-key" npx @arabold/docs-mcp-server@latest search react "useState hook"

Benefits:

• Direct integration with AI assistant
• No separate server process required
• Persistent data storage in user's home directory
• Shared database with standalone server and CLI

Limitations:

• No web interface (unless launched separately)
• Documentation indexing requires MCP tools or separate commands

Scraping Local Files and Folders

You can index documentation from your local filesystem by using a file:// URL as the source. This works in both the Web UI and CLI.

Examples:

• Web: https://react.dev/reference/react
• Local file: file:///Users/me/docs/index.html
• Local folder: file:///Users/me/docs/my-library

Requirements:

• All files with a MIME type of text/* are processed. This includes HTML, Markdown, plain text, and source code files such as .js, .ts, .tsx, .css, etc. Binary files, PDFs, images, and other non-text formats are ignored.
• You must use the file:// prefix for local files/folders.
• The path must be accessible to the server process.
• If running in Docker:
  • You must mount the local folder into the container and use the container path in your file:// URL.
  • Example Docker run:

    docker run --rm \
      -e OPENAI_API_KEY="your-key" \
      -v /absolute/path/to/docs:/docs:ro \
      -v docs-mcp-data:/data \
      -p 6280:6280 \
      ghcr.io/arabold/docs-mcp-server:latest \
      scrape mylib file:///docs/my-library
    

  • In the Web UI, enter the path as file:///docs/my-library (matching the container path).

See the tooltips in the Web UI and CLI help for more details.

Advanced: Docker Compose (Scaling)

For production deployments or when you need to scale processing, use Docker Compose to run separate services. The system selects either a local in-process worker or a remote worker client based on the configuration, ensuring consistent behavior across modes.

Start the services:

# Clone the repository (to get docker-compose.yml)
git clone https://github.com/arabold/docs-mcp-server.git
cd docs-mcp-server

# Set your environment variables
export OPENAI_API_KEY="your-key-here"

# Start all services
docker compose up -d

Service architecture:

• Worker (port 8080): Handles documentation processing jobs
• MCP Server (port 6280): Provides /sse endpoint for AI tools
• Web Interface (port 6281): Browser-based management interface

Configure your MCP client:

{
  "mcpServers": {
    "docs-mcp-server": {
      "type": "sse",
      "url": "http://localhost:6280/sse",
      "disabled": false,
      "autoApprove": []
    }
  }
}

Alternative connection types:

// SSE (Server-Sent Events)
"type": "sse", "url": "http://localhost:6280/sse"

// HTTP (Streamable)
"type": "http", "url": "http://localhost:6280/mcp"

Access interfaces:

• Web Interface: http://localhost:6281
• MCP Endpoint (HTTP): http://localhost:6280/mcp
• MCP Endpoint (SSE): http://localhost:6280/sse

This architecture allows independent scaling of processing (workers) and user interfaces.

Embeddings

Set the embedding model with YAML (embeddings.model), DOCS_MCP_EMBEDDING_MODEL, or --embedding-model. If you leave the model empty but provide OPENAI_API_KEY, the server defaults to text-embedding-3-small. Provider credentials use the provider-specific environment variables below.

See examples above for usage.

Embedding Model Options

Set DOCS_MCP_EMBEDDING_MODEL to one of:

• text-embedding-3-small (default, OpenAI)
• openai:snowflake-arctic-embed2 (OpenAI-compatible, Ollama)
• vertex:text-embedding-004 (Google Vertex AI)
• gemini:embedding-001 (Google Gemini)
• aws:amazon.titan-embed-text-v1 (AWS Bedrock)
• microsoft:text-embedding-ada-002 (Azure OpenAI)
• Or any OpenAI-compatible model name

Provider-Specific Configuration Examples

Here are complete configuration examples for different embedding providers:

OpenAI (Default):

OPENAI_API_KEY="sk-proj-your-openai-api-key" \
DOCS_MCP_EMBEDDING_MODEL="text-embedding-3-small" \
npx @arabold/docs-mcp-server@latest

Ollama (Local):

OPENAI_API_KEY="ollama" \
OPENAI_API_BASE="http://localhost:11434/v1" \
DOCS_MCP_EMBEDDING_MODEL="nomic-embed-text" \
npx @arabold/docs-mcp-server@latest

LM Studio (Local):

OPENAI_API_KEY="lmstudio" \
OPENAI_API_BASE="http://localhost:1234/v1" \
DOCS_MCP_EMBEDDING_MODEL="text-embedding-qwen3-embedding-4b" \
npx @arabold/docs-mcp-server@latest

Google Gemini:

GOOGLE_API_KEY="your-google-api-key" \
DOCS_MCP_EMBEDDING_MODEL="gemini:embedding-001" \
npx @arabold/docs-mcp-server@latest

Google Vertex AI:

GOOGLE_APPLICATION_CREDENTIALS="/path/to/your/gcp-service-account.json" \
DOCS_MCP_EMBEDDING_MODEL="vertex:text-embedding-004" \
npx @arabold/docs-mcp-server@latest

AWS Bedrock:

AWS_ACCESS_KEY_ID="your-aws-access-key-id" \
AWS_SECRET_ACCESS_KEY="your-aws-secret-access-key" \
AWS_REGION="us-east-1" \
DOCS_MCP_EMBEDDING_MODEL="aws:amazon.titan-embed-text-v1" \
npx @arabold/docs-mcp-server@latest

Azure OpenAI:

AZURE_OPENAI_API_KEY="your-azure-openai-api-key" \
AZURE_OPENAI_API_INSTANCE_NAME="your-instance-name" \
AZURE_OPENAI_API_DEPLOYMENT_NAME="your-deployment-name" \
AZURE_OPENAI_API_VERSION="2024-02-01" \
DOCS_MCP_EMBEDDING_MODEL="microsoft:text-embedding-ada-002" \
npx @arabold/docs-mcp-server@latest

For more architectural details, see the .

For enterprise authentication and security features, see the .

Telemetry

The Docs MCP Server includes privacy-first telemetry to help improve the product. We collect anonymous usage data to understand how the tool is used and identify areas for improvement.

What We Collect

• Command usage patterns and success rates
• Tool execution metrics (counts, durations, error types)
• Pipeline job statistics (progress, completion rates)
• Service configuration patterns (auth enabled, read-only mode)
• Performance metrics (response times, processing efficiency)
• Protocol usage (stdio vs HTTP, transport modes)

What We DON'T Collect

• Search query content or user input
• URLs being scraped or accessed
• Document content or scraped data
• Authentication tokens or credentials
• Personal information or identifying data

Disabling Telemetry

You can disable telemetry collection entirely:

Option 1: CLI Flag

npx @arabold/docs-mcp-server@latest --no-telemetry

Option 2: Environment Variable

DOCS_MCP_TELEMETRY=false npx @arabold/docs-mcp-server@latest

Option 3: Docker

docker run \
  -e DOCS_MCP_TELEMETRY=false \
  -v docs-mcp-data:/data \
  -p 6280:6280 \
  ghcr.io/arabold/docs-mcp-server:latest

For more details about our telemetry practices, see the .

Development

To develop or contribute to the Docs MCP Server:

• Fork the repository and create a feature branch.
• Follow the code conventions in .
• Write clear commit messages (see Git guidelines above).
• Open a pull request with a clear description of your changes.

For questions or suggestions, open an issue.

Architecture

For details on the project's architecture and design principles, please see .

Notably, the vast majority of this project's code was generated by the AI assistant Cline, leveraging the capabilities of this very MCP server.

License

This project is licensed under the MIT License. See for details.