Brain MCP Server

Semantic knowledge base access for markdown notes via Model Context Protocol

Brain adds large directories of notes (such as Obsidian vaults), emails, or individual files (such as PDF's) into an intelligent, searchable knowledge base that integrates with Claude Code and other agentic frameworks via MCP.

🧠 How Brain Works

flowchart TD
    A[📄 Your Second Brain] --> B[🔧 Chunk & Embed]
    B --> C[💾 Vector Store]
    
    D[❓ Your Query] --> E[🔍 Find Similar Chunks]
    C --> E
    
    E --> F{Similar?}
    F -->|Yes| G[📝 Add to LLM Context]
    F -->|No| H[Skip]
    
    style A fill:#e1f5fe
    style D fill:#f3e5f5
    style G fill:#e8f5e8

Brain reads your markdown files (such as Obsidian vaults), PDF's, etc and creates semantic embeddings, and provides intelligent search through MCP tools. Ask Claude Code naturally: "Using my meeting notes from the REACH collaboration and the hexagonal antenna array PDFs, help me understand the chromatic distortion mitigation strategy for the 50-200 MHz band" and it automatically searches your notes and returns adds only relevant notes to the context window. Brain automatically determines which notes are relevant and if no notes are relevant nothing is added to the context.

🚀 Install

1. Install Brain

sudo npm install -g brain-mcp

2. Setup Configuration

brain setup

This will ask for your OpenAI API key.

Note: If you have the OPENAI_API_KEY environment variable set, Brain will automatically use it. You can skip storing the API key in the config file if you prefer to use environment variables.

3. Add to Claude Code

After running brain setup, add Brain to Claude Code:

claude mcp add brain node "$(npm root -g)/brain-mcp/dist/mcp/server.js" -s user

This installs Brain globally for all your Claude Code projects. Brain stores all data in ~/.brain/ directory, making it accessible across all your projects. Learn more about Claude MCP configuration.

4. Add to Other Agentic Frameworks

Brain is a standard MCP server. For other MCP-compatible systems, use:

• Command: brain-mcp
• Args: ["server"]
• Protocol: Model Context Protocol v1.0

The server runs on stdio and is compatible with any MCP client or agentic framework.

🔧 Manual Commands

# Interactive setup (notes path, API key, MCP config)
brain setup

# Add files to knowledge base (PDFs, markdown, etc.)
brain add <path>              # Add specific file or directory
brain add .                   # Add all supported files in current directory

# Remove files from knowledge base
brain remove <path>

# Update search index for changed files
brain update

# Rebuild entire search index  
brain clear

# Show configuration and status
brain status

Note: Brain stores all data globally in ~/.brain/ directory, allowing you to access your knowledge base from any project location.

🎯 Usage

Once installed, Brain provides these MCP tools:

• brain_search - Semantic search across your notes
• brain_read - Read specific notes with context
• brain_overview - Knowledge base summary
• brain_related - Find connected notes
• brain_list - Browse notes by directory

Simply ask Claude Code questions about your notes and it will automatically use these tools.

5. Example Interaction

Here's what a typical interaction with Claude Code looks like once Brain is installed:

You: Using my notes from the REACH team meeting and the bayesian analysis PDF, can you explain how we're handling the foreground contamination in our 21-cm signal extraction?

Claude: I'll search your notes for information about REACH's foreground and Bayesian analysis approach.

<function_calls>
<invoke name="mcp__brain__brain_search">
<parameter name="query">REACH foreground bayesian 21-cm signal</parameter>
</invoke>
</function_calls>

<function_results>
Found 4 relevant chunks:
- meetings/2024-reach-collaboration.md (similarity: 0.93)
- pdfs/reach-bayesian-pipeline.pdf (similarity: 0.89)
- reach-antenna-design.md (similarity: 0.76)
- cosmic-dawn-theory.md (similarity: 0.72)
</function_results>

Let me read the most relevant meeting notes:

<function_calls>
<invoke name="mcp__brain__brain_read">
<parameter name="notePath">meetings/2024-reach-collaboration.md</parameter>
</invoke>
</function_calls>

Based on your REACH collaboration notes and the Bayesian pipeline PDF, you're using a Bayesian framework to separate the 21-cm signal from foregrounds. The meeting notes highlight that the hexagonal array's chromatic response is key - by having well-characterised frequency-dependent beam patterns, you can distinguish between spectrally smooth foregrounds to the and the structured 21-cm signal.

The PDF details how the pipeline uses physically motivated foreground models with 5-7 log-polynomial terms, while simultaneously fitting for the 21-cm signal using turning points at z~17 (absorption trough) and z~6 (emission peak). Your notes mention achieving 100-200 mK sensitivity in the 50-200 MHz band after 1000 hours of integration.

Note: While Brain automatically searches when relevant, at this stage we recommend occasionally hinting to Claude that it should "look in your notes" or "search my knowledge base" for optimal results, as full automatic detection is still being optimized.

🚧 Roadmap

• Parallel semantic search
• Status CLI commands
• Auto initialization
• Multiple vaults support
• Accept other sources (email, todo lists, etc.)
• Include other file types (images, PDFs)
• Add open source embedding models
• Add to brain?

📋 Requirements

• Node.js 16+
• OpenAI API key (for embeddings)
• Markdown notes directory (e.g., Obsidian vault)
• MCP-compatible client (Claude Code, etc.)