Graphify-Labs/graphify

A CLI and AI-assistant skill that maps a codebase and its docs into a queryable knowledge graph using local AST parsing, with no vector store.

What it is

Graphify turns a folder of code, docs, SQL schemas, configs, PDFs, and other media into a knowledge graph you traverse instead of grepping through files. Code is parsed deterministically and locally with tree-sitter, so nothing leaves the machine for the code pass; only the optional semantic pass over docs and media calls a model backend. It installs as a /graphify skill for Claude Code, Cursor, Codex, Gemini CLI, and 20+ other assistants, and is aimed at developers who want their coding agent to answer architecture and cross-file questions from a real graph rather than a similarity search.

Key features

Local, deterministic code parsing via tree-sitter AST across roughly 40 languages, producing calls / imports / inherits / mixes_in edges resolved across files — with no LLM and no data leaving the machine.
Every edge is tagged EXTRACTED (explicit in the source) or INFERRED (resolved by graphify), so you can tell what was read directly from what was derived.
A real graph you traverse rather than an embedding index: graphify query returns a scoped subgraph for a plain-language question, graphify path A B traces how two things connect, and graphify explain describes one node.
Three output artifacts per run: graph.html (interactive), GRAPH_REPORT.md (god nodes, surprising connections, suggested questions), and graph.json (the full graph).
Leiden community detection with LLM-free labels; design rationale (# NOTE: / # WHY: comments, ADR/RFC citations) is extracted as first-class nodes linked to the code.
Optional inputs beyond code — docs, PDFs, images, and video/audio — plus an MCP server (stdio or HTTP) exposing query_graph, get_node, get_neighbors, shortest_path, and PR tools.
Git hook for auto-rebuild on commit, a graph merge driver so graph.json is never left with conflict markers, and a .graphifyignore that merges with .gitignore.

Tech stack

Python, requires 3.10+; distributed on PyPI as graphifyy (double-y) with the CLI command graphify; MIT licensed; version 0.9.22 in pyproject.toml.
Core dependencies: networkx >=3.4, numpy >=1.21, rapidfuzz >=3.0, and a large set of pinned tree-sitter grammar packages (python, javascript, typescript, go, rust, java, c, cpp, ruby, c-sharp, kotlin, scala, php, swift, lua, zig, powershell, elixir, objc, julia, verilog, fortran, bash, json, and more).
Optional extras for PDF, office (.docx/.xlsx), Google Workspace, video/audio (faster-whisper + yt-dlp), MCP (with starlette floored above CVE fixes), Neo4j/FalkorDB export, SQL/Postgres introspection, Leiden (graspologic, Python < 3.13), and several model backends (OpenAI, Gemini, Anthropic, Bedrock, Azure, Ollama).
setuptools build backend; dev tooling includes pytest, ruff, pyright, bandit, and pip-audit.

When to reach for it

You want an AI assistant to answer codebase questions from a graph, preferring scoped queries over reading files or grepping.
You need cross-file relationships (call/import/inherit) for a large, multi-language repo mapped deterministically.
You want code analysis that stays fully local with zero LLM cost for the code graph itself.
You want a graph you can commit to git so a whole team starts from the same map, refreshed automatically on commit.

When not to reach for it

You want semantic similarity retrieval over prose; graphify is explicitly not a vector index and traverses a graph instead.
Your questions are answered mainly over docs, PDFs, or media rather than code — that path requires a configured model backend and is not the local, free part.
You installed the package with plain pip on macOS/Windows, which the README warns can lead to ModuleNotFoundError because the skill resolves Python at runtime; uv tool install or pipx install are recommended to avoid it.
You need a language not covered by the bundled tree-sitter grammars or the optional extractors.

Maturity signal

The project is young but developing quickly: created in April 2026, pushed within days of this page's facts (July 2026), MIT licensed, with a high star count and a pre-1.0 version (0.9.22). The default branch is v8, the extractor and skill layout is broad, and the dependency set is carefully pinned with security-motivated floors, all of which read as active, fast-moving development rather than maintenance mode. A Y Combinator S26 badge and a hosted "always-on" product on the roadmap indicate the open-source CLI sits alongside a commercial effort.

Alternatives

Microsoft GraphRAG — use it when you want an LLM to build the graph and are comfortable with per-token extraction cost; graphify deliberately parses code without an LLM.
mem0 — use it when you want a general memory layer for an application rather than a code-and-docs knowledge graph; graphify benchmarks against it on LOCOMO recall.
supermemory — another memory system graphify compares against; reach for it when your need is conversational/agent memory rather than repository structure.
A conventional embeddings/vector RAG store — use it when semantic nearest-neighbor search fits better than deterministic graph traversal.

Notes

The official PyPI package is graphifyy with two y's while the command stays graphify; the README warns that other graphify* packages are unaffiliated and that uvx graphify fails because the package, not the command, must be named (uvx --from graphifyy graphify install).
The graph build reportedly uses zero LLM credits, and benchmark scoring used a judge blind-validated against a second judge (reported 90.6% agreement, Cohen's kappa 0.81).
A strict mode on Claude Code can block the first raw source read of a session and redirect it to the graph, then revert to a soft nudge for the rest of the session.

Wiki: Graphify-Labs/graphify