Hands-on guide · Claude Code + MCP

Your $20 plan — make it work like the $100 one

The tool I talked about in the reel: Graphify — free and open source, it turns your whole project into a local Knowledge Graph (the code is analyzed on your machine with zero tokens). Next session, Claude queries the graph and pulls only what it needs, instead of re-reading your files from scratch. Three commands to install, and every step comes straight from the official docs.

1

Why your subscription runs out mid-month

Every time you open a new session in Claude Code, it starts from zero: it re-reads the system prompt, the skills, your CLAUDE.md file, and the project files — all over again.

  • That means thousands of tokens burn before you type your first word
  • And whenever it needs to understand something, it opens whole files and reads them — even if it only needs two lines
  • The result: your subscription runs dry and you have no idea where it went

The fix I mentioned in the reel: a free tool that builds a Knowledge Graph for your project — a second brain. Instead of reading whole files, Claude queries the graph and takes only what it needs.

2

The tool: Graphify

Graphify (by Graphify Labs) turns your entire project — code, docs, SQL files, config, even PDFs — into a queryable knowledge graph. Free and open source (MIT / Apache-2.0).

  • Code is analyzed 100% locally — it reads the code structure with tree-sitter, no AI model involved: nothing leaves your machine, and building the code graph costs zero tokens
  • 37 programming languages supported: Python, TypeScript, JavaScript, Go, Rust, Java, PHP and more
  • Works with 15+ coding assistants: Claude Code (the primary one), Cursor, Codex, Gemini CLI…
  • No tracking: no telemetry, no vector store, no embeddings

After the build you get 3 files in the graphify-out/ folder:

FileWhat it is
graph.htmlThe interactive graph — see your project as a network of nodes and links (the shot from the reel)
GRAPH_REPORT.mdKey highlights + suggested questions to get you started
graph.jsonThe graph itself — this is what Claude queries instead of reading the files
Official repo — Graphify-Labs/graphify
3

Installation — 3 commands

Graphify needs Python 3.10+ and the uv package manager. If you don't have uv yet, install it first:

Windows paste this into PowerShell:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Mac / Linux paste this into Terminal:

curl -LsSf https://astral.sh/uv/install.sh | sh

1) Install Graphify (close your Terminal and reopen it after installing uv):

uv tool install graphifyy
Watch the name: the package is called graphifyy with two y's — not a typo.

2) Hook it up to Claude Code (it installs the skill on its own):

graphify install

3) Open claude in your project folder and build the graph:

/graphify .

Done — your first graph is built. Open graphify-out/graph.html in your browser and see your project as a network.

Using Cursor, Codex or Gemini CLI? Same thing with graphify install --platform cursor (or codex / gemini).

4

How it saves you tokens — daily use

The rule: query the graph instead of letting Claude read files. The answer comes back as a small, focused subgraph — far fewer tokens than opening the files themselves.

CommandWhat it does
/graphify .Build the graph for the first time (code is processed locally, zero tokens)
/graphify --updateUpdate the graph — re-reads only the files that changed
graphify query "..."Ask the graph: “what connects auth to the database?”
graphify path "A" "B"Shortest path between two components — how they connect to each other
graphify explain "X"Explain a component or concept from the graph
graphify hook installAuto-updates the graph on every git commit — memory that stays fresh on its own
One line in CLAUDE.md makes it automatic: add “Before reading project files, query the Graphify graph (graphify query) and prefer its scoped answers; read full files only when the graph lacks the detail.” — from then on, every session starts from the graph, not from scratch.
5

Ready-made prompts — copy and try

1) First session after building the graph:

Before reading any file: query the Graphify graph for everything related to [the task] and rely on its results. Open full files only if the graph isn't enough.

2) Understand a project that's new to you in minutes:

Run graphify explain on the main components of this project and show me: what each one does, and how they're connected to each other. Then suggest where I should start editing to add [the feature].

3) Before touching anything sensitive:

Run graphify query "what depends on [file/component name]?" and tell me what could break if I change it.

4) When you're done working:

Run /graphify --update so the graph catches up with the latest changes.
About that 70% figure: the savings depend on your project size — big projects where every session used to start by reading dozens of files feel the biggest difference (and building the code graph itself costs zero tokens). Try it for a week and compare your usage yourself.
6

Troubleshooting + FAQ

ProblemFix
graphify: command not foundClose your Terminal and reopen it after uv tool install graphifyy — the PATH needs to refresh
uv not installed / old PythonGraphify needs Python 3.10+. Install uv with the command from Step 3 and it takes care of Python on its own
/graphify not showing up in ClaudeRun graphify install and reopen your claude session
Graph is huge / includes files that don't belongAdd a .graphifyignore file (same format as .gitignore) and exclude build/, node_modules/ and the like
Graph is out of date/graphify --update, or turn on auto-updates: graphify hook install
Is it really free?
Yes — open source under a dual MIT / Apache-2.0 license. And building the code graph happens locally with no token usage at all.
Does my code leave my machine?
No — code is analyzed locally with tree-sitter, no API involved. The one exception: if you let it process docs/PDFs/images, those go through your assistant's model (Claude). There's also an Ollama option if you want even those to be 100% local.
How is this different from CLAUDE.md?
CLAUDE.md is static and gets read in full every session. The Graphify graph is built from the code itself and queried by search — Claude gets a small, focused subgraph instead of whole files. Together they're stronger: CLAUDE.md for the rules, the graph for the structure.
My project isn't just code — what about docs and PDFs?
Supported: Markdown, PDFs, images, even video and audio (transcribed locally with faster-whisper) and Google Workspace files — they all go into the graph.
Does it work with Cursor or anything other than Claude?
Yes — 15+ assistants: graphify install --platform cursor (or codex / gemini / agents).
Will the commands stay the same?
Verified against the official README as of 2026-08-24. The permanent reference: github.com/Graphify-Labs/graphify.
✓ Copied