🧠 Agentic Knowledge

Search any documentation as if you had written it yourself

An MCP server that guides AI assistants to navigate documentation using their built-in tools (grep, file reading) instead of traditional RAG. Leverages ever growing capabilities of large language models, better tool-calling and interpretation and agentic search patterns for precise, intelligent documentation discovery.

---

🎯 What Is This For?

Give your AI assistant access to any documentation—yours or third-party—so it can find answers as naturally as you would. No embeddings, no vector databases, no complex infrastructure.

Perfect for:

• 📚 Project documentation - Your team's internal docs, APIs, guides
• 🔧 Framework references - React, TypeScript, MCP SDK, any library
• 🏢 Enterprise knowledge - Company wikis, architecture docs, runbooks
• 🌐 Open source projects - Clone any repo's docs for instant access

🚀 Quick Start

1. Configure an MCP Client

Add to your coding agent config something along the lines of

{
  "mcpServers": {
    "agentic-knowledge": {
      "command": "npx",
      "args": ["-y", "@codemcp/knowledge@latest"]
    }
  }
}

2. Set Up Your First Docset

Option A: Use the CLI (Recommended)

# For a Git repository
npx agentic-knowledge-mcp create \
  --preset git-repo \
  --id react-docs \
  --name "React Documentation" \
  --url https://github.com/facebook/react.git

# Initialize (downloads the docs)
npx agentic-knowledge-mcp init react-docs

# The MCP server starts automatically when Claude Desktop launches

Option B: Manual Configuration

Create .knowledge/config.yaml:

version: "1.0"
docsets:
  - id: my-docs
    name: My Project Documentation
    sources:
      - type: local_folder
        paths: ["./docs"]

3. Use It

Your AI assistant now has access to search_docs and list_docsets tools. Ask questions naturally:

"How do I implement a cleanup function in React useEffect?"
"Show me the authentication setup in our docs"
"Find examples of rate limiting in the API docs"

The assistant will receive intelligent navigation instructions and use grep/file reading to find the exact information.

📖 Documentation

• User Guide - Detailed CLI commands, lifecycle, configuration
• Examples - Configuration examples and integration guides
• Testing Guide - Comprehensive testing documentation

💡 How and Why It Works

The Paradigm Shift

Traditional RAG (Retrieval-Augmented Generation) was built for the context-poor era when models had 8K token limits. It:

• Chunks documents (losing relationships)
• Computes embeddings (missing precise terminology)
• Retrieves fragments (losing context)
• Requires massive infrastructure (vector DBs, rerankers)

Agentic Knowledge leverages modern AI capabilities:

• ✅ 200K+ token context windows - Can read entire documentation sets
• ✅ Powerful filesystem tools - grep, ripgrep, file reading built-in
• ✅ Intelligent navigation - Provides search strategies, not fragments
• ✅ Zero infrastructure - Just a config file and your docs

From Retrieval to Navigation

Traditional RAG says: "Here are 50 fragments that mention your keywords"

Agentic Knowledge says: "Search for 'useState' in ./docs/react-18.2/hooks/. If that doesn't help, try 'state management' in ./docs/patterns/. Follow any 'See also' references you find."

The difference? Guidance over fragments. Investigation over retrieval.

How It Actually Works

1. Configure docsets - Point to local folders or Git repositories
2. Initialize - Downloads/symlinks documentation to .knowledge/docsets/
3. MCP server - Exposes search_docs and list_docsets tools
4. AI searches - Gets navigation instructions, uses grep/file tools
5. Finds answers - Reads complete documents with full context

Performance:

• Setup: Seconds (vs hours for RAG indexing)
• Response: <10ms (vs 300-2000ms for RAG)
• Infrastructure: None (vs Elasticsearch + Vector DB)
• Accuracy: Complete context (vs fragment-based)

Inspired By

This approach is inspired by The RAG Obituary by Nicolas Bustamante and how Claude Code revolutionized code analysis by ditching RAG for direct filesystem exploration.

🚀 Local Development

# Install dependencies
pnpm install

# Start development mode
pnpm dev

# Run tests
pnpm test

# Build all packages
pnpm build

See User Guide for installation from source.

🤝 Contributing

This project follows a structured development workflow. See our development documentation for contribution guidelines.

📄 License

Distributed under the MIT License. See LICENSE file for details.

---

🎯 Moving beyond RAG into the agentic era of knowledge systems

Inspired by The RAG Obituary by Nicolas Bustamante