Graphify
Graphify is an open-source CLI + AI-assistant skill that parses a repo into a knowledge graph you query instead of grepping. Type /graphify in your AI coding assistant and it maps your project (code, docs, PDFs, images, videos) into a knowledge graph. Currently ~99k stars, Apache-2.0/MIT dual-licensed, by Safi Shamsi / Graphify Labs (YC S26).
The core idea Code is parsed with tree-sitter AST — deterministic, no LLM, nothing leaves your machine. Docs, PDFs, images and video use your assistant’s model for a semantic pass. Each connection is tagged EXTRACTED (explicit in the source) or INFERRED (resolved by graphify). No embeddings, no vector store: a real graph you traverse. That’s the pitch versus RAG — traversal with file:line provenance instead of fuzzy top-k chunks.
Install and use
uv tool install graphifyy # note the double-y; graphify* on PyPI is not affiliated
graphify install # registers the skill (Claude Code by default)
/graphify .Output is three files in graphify-out/: graph.html (interactive), GRAPH_REPORT.md (key concepts, surprising connections, suggested questions), and graph.json (the full graph).
Query surface:
graphify query "what connects auth to the database?"
graphify path "UserService" "DatabasePool"
graphify explain "APIRouter"What’s in the box - God nodes (most-connected concepts), Leiden community detection with LLM-free labels, cross-file calls/imports/inherits edges across ~40 languages - # NOTE: / # WHY: comments and ADR/RFC citations become first-class nodes linked to the code - MCP server (python -m graphify.serve) exposing query_graph, get_node, get_neighbors, shortest_path; stdio or shared HTTP transport with --api-key - Exports to Neo4j/FalkorDB cypher, GraphML, SVG, Obsidian vault, markdown wiki - Git hooks for auto-rebuild on commit, plus a merge driver that union-merges graph.json so parallel commits don’t conflict - --postgres DSN for live PostgreSQL schema introspection - 20+ assistant platforms; graphify claude install writes a PreToolUse hook that nudges the assistant to query the graph before grepping, --strict blocks the first raw source read of a session
Relevant to your stack: the Postgres introspection plus .py/.tsx/.sql extraction means a Django + React monorepo maps reasonably well, and graphify-out/ is meant to be committed so the whole team starts with a map.
My honest read - The AST layer is the real product and it’s sound — tree-sitter, deterministic, free, offline. That part I’d trust. - The benchmarks (LOCOMO recall@10 0.497 vs mem0 0.048; LongMemEval-S 76% tied with dense RAG) are self-published by the maintainer on his own harness. Treat as directional, not validated. - 99k stars with ~1,300 commits and a release nearly every day, plus a heavily marketed enterprise waitlist, is a pattern worth being skeptical of. The star count is not evidence of engineering quality. - Practical gotchas: HTML viz falls over above ~5000 nodes, and the graphifyy-vs-graphify package/command split causes a lot of the reported install failures.
If you want to evaluate it, run it on Recovery Metrics code-only (graphify extract . --code-only) — zero API cost, fully local, and you’ll know within ten minutes whether the graph actually reflects your architecture.