Prompt & Done · GRAPHIFY pack

Map a 3,000-file codebase in 43 seconds, with zero AI tokens.

Graphify (Apache-2.0, Graphify-Labs/graphify) parses your code with tree-sitter and turns every class, call and import into one graph you can query. Every edge is tagged EXTRACTED (read from the code) or INFERRED (resolved by Graphify).

1 Install

uv tool install graphifyy      # note the double y (official package); or: pipx install graphifyy
graphify install               # registers the skill with Claude Code (other assistants: graphify install --platform <name>, graphify cursor install)

Needs Python 3.10+ and uv or pipx. The PyPI name is graphifyy (two y); other graphify packages are not official.

2 Build the map

cd your-repo
graphify update .              # code only: tree-sitter, no LLM, no API key, nothing leaves your machine
open graphify-out/graph.html   # the clickable map; GRAPH_REPORT.md lists the "god nodes"

Or type /graphify . in Claude Code (Codex: $graphify .). Run graphify claude install in a project and Claude Code checks the graph before it greps (CLAUDE.md rule + a PreToolUse hook).

3 Ask how two things connect

graphify path "ModelAdmin" "HttpResponse"
Shortest path (2 hops):
  ModelAdmin --uses [INFERRED]--> SimpleTemplateResponse --inherits [EXTRACTED]--> HttpResponse

graphify explain "HttpResponse"      # one concept and its neighbours
graphify query "how does a request reach a view?"   # a scoped subgraph for a plain-language question

4 What we measured (24 Oct 2026, M4 Pro, graphifyy 0.9.95)

Django (3,140 files in the corpus): 55,135 nodes, 131,306 edges, 2,095 communities in 42.6 to 42.9 s; report line "Token cost: 0 input, 0 output". 91% of edges EXTRACTED, 9% INFERRED.
Flask (106 files): 2,342 nodes, 3,749 edges in 2.2 s.
Query: graphify path "ModelAdmin" "HttpResponse" on Django took 1.79 to 1.81 s (mostly loading the 59 MB graph.json).

Fair to say: this is the free, local, code-only part (tree-sitter). Docs, PDFs, images and video need a model backend. The paid Graphify Cloud is a separate product. We tested Claude Code with and without the graph on Flask and Django (Haiku, 8 runs): no consistent saving in tokens or cost, so we do not claim one. We tested the Claude Code install; Cursor and the other assistants are per the README.