graphify

// Overview

graphify builds a knowledge graph out of whatever you point it at — source code, Markdown notes, PDFs, screenshots and diagrams — and then lets you and your coding agent query that graph instead of re-reading the raw files. It runs locally: code is parsed deterministically with tree-sitter into an AST plus a call-graph pass, while docs, papers and images are turned into concepts and relationships by an LLM. There is no vector store and no Neo4j server required; the stack is NetworkX, Leiden community detection via graspologic, tree-sitter and vis.js.

A run writes a `graphify-out/` directory containing an interactive `graph.html` you can click through, an Obsidian-compatible vault, a persistent `graph.json` you can query weeks later, a `GRAPH_REPORT.md` that surfaces 'god nodes' (the highest-degree concepts everything routes through) and surprising cross-document connections, and a SHA256 cache so re-runs only reprocess changed files. Every edge is labelled `EXTRACTED`, `INFERRED` or `AMBIGUOUS`, so it stays clear which relationships were read out of the source and which were guessed.

It is designed to be driven by an agent. It installs as a `/graphify` skill for Claude Code and can also expose the graph over MCP with `--mcp`, so an agent can ask for a path between two symbols or an explanation of one. A `--wiki` mode emits Wikipedia-style Markdown articles with an `index.md` entry point, which lets any agent navigate the knowledge base by reading files. A `--watch` mode and a post-commit git hook keep the graph in sync as code changes.

// What it does

• Deterministic AST extraction with tree-sitter for Python, TypeScript, JavaScript, Go, Rust, Java, C/C++, Ruby, C#, Kotlin, Scala and PHP, plus a call-graph pass
• Multimodal ingestion — Markdown and text, PDFs with citation mining, and images (screenshots, diagrams, whiteboard photos) via model vision
• Runs fully locally on NetworkX + Leiden communities; no vector store and no Neo4j server needed
• Every edge tagged EXTRACTED, INFERRED or AMBIGUOUS so you can tell findings from guesses
• Outputs an interactive graph.html, an Obsidian vault, a persistent graph.json, and a GRAPH_REPORT.md of god nodes and surprising connections
• Agent-facing surfaces: a /graphify skill, an MCP stdio server (--mcp), and a crawlable Markdown wiki (--wiki)
• Incremental by design — SHA256 cache, --update for changed files, --watch for live sync, and a post-commit git hook
• Exports to SVG, GraphML (Gephi, yEd) and Neo4j cypher

// Getting started

graphify installs as a Python package that registers a Claude Code skill. It requires Claude Code and Python 3.10 or newer.

pip install graphifyy && graphify install

/graphify .
/graphify ./raw --mode deep      # more aggressive INFERRED edge extraction
/graphify ./raw --update         # re-extract only changed files

/graphify query "what connects attention to the optimizer?"
/graphify path "DigestAuth" "Response"
/graphify explain "SwinTransformer"

/graphify ./raw --watch
graphify hook install
/graphify ./raw --mcp

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Reach for it when an agent needs the shape of a large codebase and reading the raw files would burn the context window
• Reach for it to connect scattered research material — papers, notes, screenshots — into one navigable graph
• Reach for it when you want graph-based retrieval with auditable edges rather than opaque vector similarity
• Reach for it to keep a live map of a repository while several agents write code in parallel

// How graphify compares

graphify alongside other open-source rag frameworks & platforms tools AI/TLDR tracks, ranked by GitHub stars.