graphify turns any repo into a persistent knowledge graph: every file, symbol and relationship, clustered into the systems they actually form, queryable in plain English. You build it once and every session after that reads it.
Loading a whole repo into context for every question is slow and costly, so out of the box Claude reads a few files, greps for a string, and infers the rest. grep finds text, not structure. "What reaches the Mac daemon?" is a relationship across a dozen files, and a string search can't see it.
The architecture it pieced together last time is gone at /clear. It re-derives the same map, burns the same tokens, and sometimes derives it differently. That is exactly how confident, wrong answers happen.
graphify builds the map once, on disk, and lets every future session read it instead of guessing.
Two layers. The structural layer is free: tree-sitter parses every function, import and call. The semantic layer is an LLM pass that adds the relationships AST can't see. The result is a graph.json you keep and commit.
21 of the 92 subsystems shown. It separated the brain (Persistent Warm Engine, Second-brain MCP Server, Metacognition Spine) from the hands (Mac Daemon Core, OCR & Click-by-text) from the HUD. graphify generated these names itself, and they match how the system is actually built.
The structural layer is free (tree-sitter). The semantic layer ran on one Claude subagent riding the Max plan. No API key touched.
Brain, hands, HUD, voice: the clusters and their names came out of the graph structure, and they match how the system is actually built.
When graphify-out/ exists, the skill answers architecture questions from the graph first, without re-reading the repo. Node names above are real communities from the live build; exact output depends on your question.
The honest caveat: on a tiny repo, or a known string in a known file, plain grep wins and a graph is overkill. graphify earns its keep when "how does this actually work?" spans more files than fit in your head.
Code, markdown, PDFs, images, even video transcripts land in one graph. Community detection then surfaces cross-document links you'd never think to search for.
graphify update re-extracts only changed files, no LLM needed. graphify watch rebuilds on save. The map stays current as the code moves.
merge-graphs joins multiple repos into one view. Exports: Neo4j, FalkorDB, GraphML, SVG, Obsidian, or an agent-crawlable wiki.
A companion graphify-mcp binary serves the graph over MCP (stdio or HTTP), so any agent queries structure directly. It installs into 19 agent platforms, not just Claude.
Every edge is tagged EXTRACTED, INFERRED or AMBIGUOUS. Claude knows which is which, and so do you.
The first rule of good engineering is to read the code fully before you change it. graphify makes that the default, even on a codebase too big to read.
The architecture lives on disk, not in a context window. New sessions start oriented, not blind, and never burn tokens re-deriving the map.
The structural layer is free. The semantic layer rides your existing subscription. A 185-file map cost $0 of metered API.
Grounded traversal with an EXTRACTED/INFERRED audit trail beats "Claude skimmed three files and extrapolated."
Found in the wild, audited by scout, adopted on merit. On a real codebase, a map beats a guess.
Part of the LevelUp kit · numbers from the live build, June 2026 · the keynote ↗ · the handbook ↗