Graphify

What it does

• Multimodal input in one graph: code via tree-sitter AST + call-graph pass, markdown/rst docs via Claude, PDFs via citation mining + concept extraction, images via Claude vision
• Outputs: interactive graph.html (click, search, filter by community), an Obsidian vault, --wiki (Wikipedia-style articles per community with an index.md agents can navigate), GRAPH_REPORT.md with god nodes, surprising connections and suggested questions, plus persistent graph.json
• Stays current: --watch rebuilds on saves (code instantly, AST-only, no LLM), a post-commit git hook rebuilds per commit, and a SHA256 cache means re-runs process only changed files
• Every edge carries provenance: EXTRACTED, INFERRED or AMBIGUOUS — what was found vs guessed is always visible
• Extras: /graphify add <url> pulls arXiv papers or tweets in; --mcp starts an MCP stdio server; exports to SVG, GraphML (Gephi/yEd) and Neo4j cypher

When to use it

Giving an agent (or yourself) structural memory over a folder that keeps changing — a research corpus, an unfamiliar codebase you're onboarding onto. The wiki mode is the agent play: point any agent at index.md and it navigates the knowledge base by reading files instead of parsing JSON.

Getting started

pip install graphifyy && graphify install

Then open Claude Code in any directory and type:
/graphify .

Requires Claude Code and Python 3.10+. The PyPI package is temporarily named graphifyy while the graphify name is being reclaimed — the CLI and skill command stay graphify. Windows: if graphify is not recognized, add the Python Scripts folder to PATH (or use pipx install graphifyy); on macOS externally-managed environments use pipx install graphifyy.

Things to know

The README's 71.5x token reduction is the authors' own benchmark on a 52-file mixed corpus — their own worked examples show ~1x on small corpora, where the value is structural clarity rather than compression. Extraction calls Claude (cost scales with corpus size and --mode deep); code rebuilds are AST-only and free. Star counts and benchmarks aside, the honest habit is the repo's own: run it on your corpus and read the edge provenance tags.