Published: 2026-04-15

Claude Code + Graphify: Build Instant Persistent Knowledge Graphs for Free

Chapters / key moments (click to jump — plays here on the page)

FuturMinds demonstrates Graphify, a free tool that creates a persistent knowledge graph of your codebase inside Claude Code. Instead of Claude re-reading the same files from scratch every new session — burning tokens and time — Graphify builds a one-time graph that loads instantly on session start.

Source video

"Claude Code + Graphify = Instant Knowledge Graph (Free)" by FuturMinds — Watch on YouTube →

Key Takeaways

Graphify solves the cold-start problem: without it, every new Claude Code session reads your entire codebase file by file before answering a single question — on a 50+ file project that's a meaningful token cost before any work begins.
The knowledge graph stores component relationships, not just file contents — giving Claude a richer mental model (what calls what, what depends on what) from fewer reads than a raw file scan.
Setup is a one-time operation: run Graphify against your project directory, it generates a graph file, then add the graph to your CLAUDE.md or session start instruction to load it automatically.
Largest gains on large codebases: projects with 50+ files see the most dramatic token savings since the baseline re-read cost is highest.
Graphify is free and works with any project Claude Code can access — no external API key required, no cloud upload of your code.

Why Codebase Cold-Start Is Expensive

FuturMinds walks through a live example: asking Claude "What does browser-use do? Give me a one-line summary of each major component." Before answering, Claude reads the README, then multiple source files, then configuration files — building its mental model from scratch. Every file read costs tokens. On a mid-size project this preamble can cost 10–30K tokens before Claude even starts the actual task.

Close and reopen the session tomorrow: Claude starts from zero again. The Graphify knowledge graph breaks this cycle — load the graph once per session and Claude has structural understanding of your codebase from token one.

Related on OpenClawDatabase

Claude Cowork Setup — project organization and memory management in the Cowork environment
Session Commands Guide — managing context rot and token costs during long sessions

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Commands & Code Shown

pip install graphifyy

Purpose: Installs the tool. Note the doubled 'y' in the package name.

When to use: First. <strong>On Windows the documented one-liner fails</strong> because the double ampersand is not recognised — run each command separately instead.

graphify install

Purpose: Completes the installation after the pip step.

When to use: Immediately after the pip install. For other harnesses (Codex, OpenCode and so on) the project lists an equivalent command to substitute here.

graphify claude install

Purpose: Wires the graph into Claude Code so every new session reads it automatically, without you asking.

When to use: Once per machine. This is the step that makes it passive rather than something you remember to invoke.

/graphify

Purpose: Builds the graph for the current project. Reports the corpus it found and, above a 200-file threshold, asks you to select subdirectories rather than doing everything at once.

When to use: Once per project. In the demo: 357 code files, 53 docs and 7 images found; two subdirectories selected; about 12 minutes to complete.

graphify update

Purpose: Detects what changed since the last run and updates the graph incrementally.

When to use: Whenever the project has moved on and you want the graph current.

graphify hook install

Purpose: Wires up git hooks so the graph rebuilds automatically on a new commit or a branch switch.

When to use: Instead of remembering to run update. This is the set-and-forget option.

Reproducible steps

The problem: every session is a new hire
Claude reads files, builds an understanding, and then you close the session. Tomorrow it starts from zero — <strong>same tokens, same searches, same files re-read</strong>. The framing: a graph turns the new hire into the senior colleague who already has the map.
Three passes, only one of which costs anything
<strong>Pass 1:</strong> a code parser across every file (Python, TypeScript, Go, Rust and anything with real syntax) extracting classes, functions, imports and calls — <em>entirely on your machine, no API calls, no tokens</em>. Facts, not guesses. <strong>Pass 2:</strong> audio and video transcribed locally with Faster-Whisper, also free. <strong>Pass 3:</strong> everything else — markdown, PDFs, images, readmes — with Claude subagents running in parallel to extract concepts and relationships. <strong>Only pass 3 touches the API, and only once.</strong> Every session afterwards reads the cached graph for free.
What you get out
A <code>graphify-out</code> directory with a <code>GRAPH_REPORT.md</code> and an interactive HTML graph. In the demo: <strong>4,041 nodes, 20,900 edges, 185 communities</strong>. Each dot is a concept, class, function or documented idea; each line a relationship; colours are neighbourhoods and dot size is how much connects to it. Communities can be toggled off individually.
How the saving actually happens
A session-start hook fires before Claude reads anything, and it reads a summary of the whole graph. Claude then makes <strong>two or three targeted reads instead of fifteen</strong>.

Gotchas

<strong>The honest measurement is the most valuable thing here.</strong> The reviewer ran ten identical questions in two sessions, with and without the tool: <strong>120,000 tokens against 113,000 — under 8% saved</strong>, against the repository's claimed 71.5× reduction. He says so plainly rather than repeating the headline figure.
<strong>But the response quality differed noticeably in the graph's favour.</strong> On a question both answered, the ungraphed session listed three phases as three function names; the graphed one explained each phase and added detail about loop detection and replanning. The saving to look for may be quality-per-token rather than raw tokens.
Windows users hit an immediate error on the documented install one-liner — the double ampersand is not recognised. Run the commands separately.
There is a 200-file threshold per graph build, so a large repository requires selecting subdirectories rather than indexing everything at once. The Django-scale demo took about 12 minutes for two subdirectories.
It works on more than code — research papers, meeting recordings, strategy documents, mixed content.

← Back to News digest · See also: Claude Cowork guide