Docs MCP Server (Grounded Docs) - Local Documentation Index for AI Assistants

Docs MCP Server (Grounded Docs) - Local Documentation Index for AI Assistants

Research Date: January 26, 2026 Source URL: https://grounded.tools GitHub Repository: https://github.com/arabold/docs-mcp-server npm Package: https://www.npmjs.com/package/@arabold/docs-mcp-server Version at Research: v1.36.0 License: MIT

Overview

Docs MCP Server (branded as "Grounded Docs") is an open-source MCP server that solves AI hallucinations and outdated knowledge by providing a personal, always-current documentation index. It fetches official docs from websites, GitHub, npm, PyPI, and local files, allowing AI assistants to query the exact library versions being used in a project.

Core Value Proposition: Open-source alternative to Context7, Nia, and Ref.Tools that runs entirely on your machine with full privacy.

Problem Addressed

Key Statistics (as of January 26, 2026)

Key Features

1. Multi-Source Documentation Indexing

• Web Scraping: Index any documentation website using Playwright
• GitHub Repositories: Index README, docs folders, source code
• Package Registries: Fetch from npm and PyPI
• Local Files: Index local folders and zip archives
• llms.txt Support: Automatic detection of LLM-ready documentation files

2. Rich File Format Support

• HTML and Markdown
• PDF documents
• Microsoft Office (Word .docx, Excel, PowerPoint)
• Source code files
• Password-protected PDFs

3. Semantic Vector Search

• Multiple Embedding Providers: OpenAI, Ollama, Google Gemini, Azure, AWS
• Hybrid Search: Combines vector similarity and full-text search
• Reciprocal Rank Fusion (RRF): Configurable weights for optimal ranking
• Dual-Mode FTS: Exact phrase and keyword matching for improved recall

4. Version-Specific Queries

• Index multiple versions of the same library
• Query targets exact version in use
• Version tracking with re-indexing capability
• Change detection for documentation updates

5. Privacy-First Architecture

• Runs entirely on local machine
• No code or queries sent to external servers
• SQLite database with schema migrations
• Optional telemetry (privacy-first design)

6. Multiple Access Interfaces

• Web UI: Add/manage documentation at http://localhost:6280
• CLI: Command-line management
• MCP Protocol: Integration with Claude, Cline, Roo, etc.
• SSE Endpoint: Server-Sent Events for real-time updates

Technical Architecture

Documentation Sources
         |
         v
+---------------------------+
|    Scraper Strategies     |
|  - Web (Playwright)       |
|  - Local filesystem       |
|  - Package registries     |
+---------------------------+
         |
         v
+---------------------------+
|    Content Fetchers       |
|  - HTTP                   |
|  - Filesystem             |
|  - Registry APIs          |
+---------------------------+
         |
         v
+---------------------------+
|  Processing Pipelines     |
|  - Middleware chains      |
|  - Content-type handling  |
+---------------------------+
         |
         v
+---------------------------+
|   Document Splitters      |
|  - SemanticMarkdown       |
|  - JsonDocument           |
|  - TextDocument           |
|  - GreedySplitter (size)  |
+---------------------------+
         |
         v
+---------------------------+
|   Embedding Generation    |
|  - OpenAI, Gemini         |
|  - Ollama (local)         |
|  - Azure, AWS             |
+---------------------------+
         |
         v
+---------------------------+
|   SQLite Storage          |
|  - libraries table        |
|  - versions table         |
|  - documents table        |
+---------------------------+
         |
         v
+---------------------------+
|   Hybrid Search           |
|  - Vector similarity      |
|  - Full-text search       |
|  - RRF ranking            |
+---------------------------+

Deployment Modes

Unified Mode (Default):

• Single process with MCP server, web interface, embedded worker
• Suitable for development and simple deployments
• Direct method calls between components
• Local event propagation via EventBus

Distributed Mode (Docker Compose):

• Separate coordinator and worker processes
• Hub (shared worker) + Spokes (Web, MCP, CLI)
• tRPC for inter-process communication
• WebSocket for real-time events
• Scalable processing across containers

Protocol Auto-Detection

• No TTY: stdio transport for direct MCP communication
• Has TTY: HTTP transport with Server-Sent Events
• Manual override: --protocol stdio|http

Installation and Usage

Quick Start

# Start server (requires Node.js 20+)
npx @arabold/docs-mcp-server@latest

# Open Web UI
# http://localhost:6280

# Add to MCP client config
{
  "mcpServers": {
    "docs-mcp-server": {
      "type": "sse",
      "url": "http://localhost:6280/sse"
    }
  }
}

Docker Deployment

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

Enable Embeddings (Recommended)

# OpenAI
OPENAI_API_KEY="sk-proj-..." npx @arabold/docs-mcp-server@latest

# Or Ollama (local, free)
OLLAMA_BASE_URL="http://localhost:11434" npx @arabold/docs-mcp-server@latest

MCP Tools Provided

The server exposes tools via MCP protocol:

Configuration System

Configuration resolves via Zod schema with four layers: Defaults < Config File < Environment Variables < CLI Arguments

Key Settings:

# config.yaml
embedding:
  provider: openai  # or ollama, gemini, azure, aws
  model: text-embedding-3-small

storage:
  path: ~/.local/share/docs-mcp-server

server:
  port: 6280
  protocol: auto  # stdio, http, or auto

Supported Embedding Providers

Comparison with Alternatives

Relevance to Claude Code Development

Direct Applications

1. Documentation Grounding: Eliminate hallucinations when coding with unfamiliar libraries
2. Version-Specific Context: Query docs for exact versions in project dependencies
3. Local Privacy: Keep proprietary documentation indexed locally
4. Multi-Format Support: Index internal docs in various formats (PDF, Word, etc.)

Patterns Worth Adopting

1. Semantic Chunking: Two-phase splitting (semantic structure + size optimization)
2. Hybrid Search: Combining vector similarity with full-text for better recall
3. Protocol Auto-Detection: Automatic transport selection based on environment
4. Write-Through Architecture: Immediate persistence for recovery capability
5. Distributed Mode Design: Hub-spoke pattern for scalable processing

Integration Opportunities

1. Claude Code Plugin: Create plugin that auto-indexes project dependencies
2. Skill Enhancement: Ground skills in current documentation automatically
3. RAG Pattern: Apply hybrid search approach to existing context systems
4. Multi-Source Truth: Pattern for combining documentation from multiple sources

Complementary to Existing MCP Ecosystem

This fills a gap in the research directory's MCP ecosystem coverage:

• Narsil MCP: Code intelligence and security scanning
• OctoCode MCP: Research Driven Development with GitHub search
• Docs MCP Server: Documentation grounding and version-specific retrieval

References

1. GitHub Repository: https://github.com/arabold/docs-mcp-server (accessed 2026-01-26)
2. Official Website: https://grounded.tools (accessed 2026-01-26)
3. npm Package: https://www.npmjs.com/package/@arabold/docs-mcp-server (accessed 2026-01-26)
4. Architecture Documentation: https://github.com/arabold/docs-mcp-server/blob/main/ARCHITECTURE.md (accessed 2026-01-26)
5. Installation Guide: https://github.com/arabold/docs-mcp-server/blob/main/docs/setup/installation.md
6. MCP Protocol Specification: https://modelcontextprotocol.io/

Freshness Tracking

Change Detection Indicators:

• Monitor GitHub releases for version changes
• Check npm for new versions
• Review changelog for breaking changes
• Track MCP protocol compatibility updates
• Monitor for new embedding provider support