DevDocs-MCP: Documentation Authority for AI Agents

Eliminate AI hallucinations with local, version-aware, and authoritative documentation.

DevDocs-MCP is a Model Context Protocol (MCP) server that provides version-pinned, deterministic documentation sourced from DevDocs.io to AI assistants (Claude, RooCode, Cline, Copilot etc.). It acts as a local Documentation Intelligence Layer, ensuring your agent always has the correct API context without network latency or training data drift.

This server follows the proposed MCP server standard defined by the MCP , which is being adopted by AI tools and agents across the ecosystem, including those from providers such as Anthropic, OpenAI, and Google.

⚠️ Project Status Notice ⚠️ This project is under active and heavy development. Still need to test existing features and implement more new features like optimal caching , update latest doc automatically and also on request basis , optimal querying and more... Design decisions, APIs, and structure are expected to evolve based on strong opinions, reviews, and feedback from the developer community. I warmly welcome discussions, critiques, and contributions to help shape this project in the right direction.

📑 Table of Contents

1. Overview
2. Key Capabilities
3. Quickstart
4. Installation & Setup
5. AI Agent Configuration
6. Further Reading

🎯 Overview

graph TD
    Agent["AI Agents"] -->|MCP Search/Explain| Server["DevDocs-MCP Server"]
    Server -->|Metadata| DB[(SQLite - sql.js)]
    Server -->|Content| FS[Local Disk - JSON Cache]
    Server -.->|Lazy Fetch| Web[DevDocs.io API]

🚀 Key Capabilities

• Offline-First: Documentation is cached locally; no internet is needed after ingestion.
• Ranked Fuzzy Search: Instantly find relevant entries (e.g., "intersection observer").
• Version Awareness: Automatically maps to specific library versions in your project.
• Node-Only Architecture: Uses sql.js for zero-native dependencies (no Python/C++ build steps).
• Structured Outputs: Returns clean, LLM-optimized content.

⚡ Quickstart

1. Clone & Install

git clone https://github.com/madhan-g-p/DevDocs-MCP.git
cd DevDocs-MCP
pnpm install

2. Configure Environment

cp .env.example .env
# Edit .env: Set DEVDOCS_DATA_PATH to where you want to store docs.

3. Build & Run

pnpm build
pnpm start:prod

Note: Strictly use pnpm for development and contributions.

🛠️ Installation & Setup

DevDocs-MCP is optimized for zero-fuss setup with pnpm. It is a Node-only project.

Prerequisites

• Node.js 18+ (tested on 18.x and 20.x).

🐳 Docker Usage

DevDocs-MCP is designed to be lightweight and persistent. Since it downloads and caches large documentation datasets, managing your data volumes correctly is essential.

Method A: Docker Compose (Dynamic & Persistent)

This is the recommended way. It supports 3 storage modes via the DEVDOCS_VOLUME_SOURCE variable in your .env file.

1. Configure your preferred storage in .env:

   # Mode 1: Named volume (Default - Docker manages it)
   DEVDOCS_VOLUME_SOURCE=devdocs_data
   
   # Mode 2: Host Path (Outside Docker - Local folder)
   # DEVDOCS_VOLUME_SOURCE=./my-docs-storage
   

2. Start the server:

   docker compose up -d
   

Flexibility: Every time you start the container, you can point to a new volume (to start fresh) or an existing volume (to keep your data) just by changing the DEVDOCS_VOLUME_SOURCE in your .env file. Both the mcp.db (database library) and the data/ (documentation downloads) are stored inside this volume.

Method B: Docker Run (Manual)

If you just want to run the public image directly from Docker Hub:

docker run -d -p 3000:3000 \
  -v "devdocs_data:/app/data" \
  --name devdocs-mcp \
  madhandock1/devdocs-mcp:latest

💾 Data Strategy

The project uses a single mount point at /app/data which contains:

• /app/data/mcp.db: The SQLite registry.
• /app/data/docs/: The cached documentation files. This unified structure makes it easy to backup or move your entire documentation authority by simply moving one folder or volume.

Configuration (.env)

Ingesting Documentation

Once the server is connected to your agent, use the ingest tool to download documentation for your specific stack: Agent will take care of it once the server is up and running , it is supposed to list the available to tools and take decision accordingly.

{
  "dependencies": {
    "react": "18.2.0",
    "typescript": "5.0.0"
  }
}

Note: The Devdocs fetch url required to be used for ingestion is defined in src/config/constants.ts. This URL currently points to the correct and officially supported path. If the devdocs community changes this path in future, the configuration may need to be updated accordingly.
This approach does not involve web scraping. It follows the same supported mechanism used by DevDocs to provide documentation for offline use, similar to how users download documentation within DevDocs itself.

🤖 AI Agent Configuration

RooCode / Cline / Claude Desktop (Recommended: STDIO)

Add this to your mcp_settings.json. By setting the cwd (Current Working Directory), the server will automatically load configuration from your .env file, allowing you to use relative paths for data storage.

{
  "mcpServers": {
    "devdocs": {
      "command": "node",
      "args": ["dist/main.js"],
      "cwd": "file://path/to/your/DevDocs-MCP"
    }
  }
}

Pro Tip: Use the absolute path to the project in cwd. The server will then find your .env and use the DEVDOCS_DATA_PATH and MCP_DB_PATH you've defined there (e.g., ./data or mcp.db), making the setup portable!

Remote / Web Clients (HTTP SSE)

If you need to expose the server via a port (for production/remote setups):

1. Start the server with a port flag:

   node dist/main.js --port 3000
   

2. Configure your client:
   • SSE URL: http://your-server:3000/mcp/sse
   • Type: SSE (Supported by most MCP clients, atleast I guess so)

📖 Further Reading

• System Architecture: Deep dive into the data flow, schema, and sql.js internals.
• Contributor Guidelines: Workflow, coding standards, and PR requirements.
• LICENSE: MIT License.

Inspired from this project devdocs-local , devdocs-local-demo by @Nathaniel Whiteinge