Graft — Code Intelligence for AI Coding Agents

Introduction

Graft is a TypeScript-based code intelligence engine that builds a persistent knowledge graph of your codebase. It parses source files with tree-sitter, extracts symbols, relationships, and call chains, then serves this structured context to AI coding agents. The result: agents understand your code faster, use fewer tokens, and produce more accurate edits.

What Graft Does

• Indexes codebases into a queryable knowledge graph of symbols and their relationships
• Provides an MCP server that AI coding agents can query for contextual code understanding
• Reduces token consumption by serving precise, relevant context instead of raw file dumps
• Supports incremental re-indexing so the graph stays current as code changes
• Works with Claude Code, Cursor, Codex, Gemini CLI, and any MCP-compatible agent

Architecture Overview

Graft uses tree-sitter parsers to produce ASTs for each source file, then extracts a graph of definitions, references, imports, and call sites. This graph is stored in a local SQLite database for fast querying. The MCP server exposes graph queries as tools that agents can call — for example, asking for all callers of a function, or the dependency chain of a module. The indexer runs as a background process that watches for file changes and updates the graph incrementally.

Self-Hosting & Configuration

• Initialize with npx graft init in your project root to generate a config file
• Run npx graft index to build the initial code graph
• Start the MCP server with npx graft serve and point your agent to it
• Configure language support and excluded directories in graft.config.json
• Set the graph database path and cache size for large monorepos

Key Features

• Tree-sitter-based parsing with support for 50+ programming languages
• Incremental indexing keeps the graph fresh without full rebuilds
• MCP server integrates with all major AI coding agents out of the box
• Sub-millisecond query responses for symbol lookups and relationship traversals
• Open source with no external API dependencies — everything runs locally

Comparison with Similar Tools

• Graphify — builds knowledge graphs from code; Graft focuses on MCP-native agent integration
• Codebase Memory MCP — persistent code index; Graft uses tree-sitter for precise AST-level graphs
• Sourcegraph — enterprise code search; Graft is local-first and agent-oriented
• Codegraph — similar graph approach; Graft emphasizes incremental updates and MCP serving

FAQ

Q: Does Graft work with monorepos? A: Yes. Configure workspace roots in the config file and Graft indexes them as a unified graph.

Q: Which languages are supported? A: Any language with a tree-sitter grammar — TypeScript, Python, Go, Rust, Java, C/C++, and more.

Q: How much disk space does the graph use? A: Typically 1-5% of the source code size, stored in a local SQLite database.

Q: Can I use Graft without an AI agent? A: Yes. The CLI includes query commands for exploring the code graph directly.

Sources

• https://github.com/trailhq/Graft