The Problem

AI coding assistants lose context between sessions. They re-read files, re-learn your codebase, and burn tokens on the same questions. GraphHub fixes this.

Traditional: grep → read 5 files → 9,216 tokens
GraphHub:    get_context("functionName") → 507 tokens (94% savings)

Quick Start

# Clone and install
git clone https://github.com/slnquangtran/graphhub.git
cd graphhub && npm install

# Index your project
npm run index -- /path/to/your/project

# Configure Claude Code (one command)
npm run install-claude

That's it. Claude now has persistent memory of your codebase.

Features

Core Intelligence

Developer Experience

Integrations

Token Savings

Real measurements on GraphHub's own codebase:

Bottom line: 5x more tasks in the same context window.

MCP Tools

When connected to Claude Code or other MCP agents:

// Find code by description
semantic_search({ query: "authentication validation" })

// See callers and callees
get_context({ name: "validateToken" })

// Check blast radius before editing
impact_analysis({ name: "handleRequest" })

// Save learnings for future sessions
remember({ content: "Auth uses JWT middleware", type: "learning" })

// Retrieve past learnings
recall({ query: "how does auth work?" })

// Run custom Cypher queries
query_graph({ cypher: "MATCH (s:Symbol)-[:CALLS]->(t:Symbol) RETURN s, t LIMIT 10" })

Architecture

┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
│ Source Code │ ──▶ │  Tree-sitter │ ──▶ │ KuzuDB Graph DB │
└─────────────┘     │    Parser    │     └────────┬────────┘
                    └──────────────┘              │
                           │                      │
                    ┌──────▼──────┐      ┌───────▼───────┐
                    │  Embeddings │      │   MCP Server  │
                    │ (MiniLM-L6) │      │  REST API     │
                    └─────────────┘      │  Dashboard    │
                                         └───────────────┘

Graph Schema:

File ──CONTAINS──▶ Symbol ◀──DESCRIBES── Chunk
File ──IMPORTS───▶ File
Symbol ──CALLS──▶ Symbol

Language Support

Configuration

Claude Code (Recommended)

npm run install-claude

This automatically:

• Configures MCP server in .claude/settings.json
• Installs PreToolUse hook (graph context before reads)
• Installs PostToolUse hook (auto-reindex after commits)
• Updates CLAUDE.md with usage instructions

Manual Setup

{
  "mcpServers": {
    "graphhub": {
      "command": "npx",
      "args": ["tsx", "src/index.ts", "serve"],
      "cwd": "/path/to/graphhub"
    }
  }
}

Commands

Privacy

• 100% Local — All data stays in .graphhub/ in your project
• No External APIs — Embeddings generated locally with Xenova/transformers
• No Telemetry — Zero network calls during indexing or querying

Tech Stack

Comparison

Roadmap

• Session memory (remember/recall/forget)
• Always-on PreToolUse hooks
• PostToolUse auto-reindex
• Graph report generation
• Verified 85% token reduction
• Worker thread indexing for large repos
• .gitignore support
• Class hierarchy edges (INHERITS, IMPLEMENTS)
• Native Python/Go Tree-sitter grammars
• Community detection (Leiden algorithm)

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

# Development setup
git clone https://github.com/slnquangtran/graphhub.git
cd graphhub && npm install

# Run tests
npm test

# Index GraphHub itself for testing
npm run index -- ./src

License

ISC © 2024