Graphify for Claude Code: how much usage e go save you?
Graphify dey build your code graph free for your machine, but docs and PDF fit chop your Claude Pro usage. Install am pinned, then measure the real saving yourself with /usage.
Graphify for Claude Code: e go stretch your Pro limit?
Graphify for Claude Code fit reduce the tokens wey Claude dey use to read your code, but nobody fit tell you the saving before you measure am for your own repo. The code part of the graph dey build for your machine, and e no dey touch your Claude usage. The docs, PDF and pictures part dey use model, and on Claude Pro that one dey chop from the same limit wey you dey use for chat.
So if you dey pay for Pro and you see post wey talk "71x fewer tokens", no just install am because of that number. Read wetin e dey do first, install the correct package, then measure with /usage. This guide show you all the steps, from install go reach how to comot am.
Wetin Graphify dey do inside your repo
Graphify na tool wey dey read your project and build a graph of am. A graph here mean list of things (functions, files, classes) and the lines wey join them, like "this function dey call that one". When Claude need answer question about your code, e fit ask the graph first instead of to open file after file.
For code, Graphify dey use tree-sitter. Tree-sitter na parser wey dey run for your own machine. E no need LLM (large language model) at all. The project docs talk am straight: "Code costs nothing to extract. Docs, PDFs and images do."
The other files na different matter. Docs, PDF and images dey pass through wetin Graphify call "semantic pass", and that one dey use model. The README for version 0.9.108 talk say e dey use "your assistant's model, or a configured API key". The docs also warn say "A README and a folder of screenshots are enough to trigger it." So almost every real repo get something wey go enter this semantic pass.
One more thing dey use model: to give name to the groups (Graphify call them communities) inside the graph. The Ollama cookbook for the docs talk say "Only docs, papers and community names use the model." If you add --no-label, Graphify no go name the communities with LLM.
When e finish, Graphify go write everything inside one folder, graphify-out/. Inside am you go see graph.html (graph wey you fit click and search for browser), GRAPH_REPORT.md (report of the most connected nodes and questions wey you fit ask) and graph.json (the graph itself, wey you fit query weeks later).
Which part dey chop your Claude Pro usage
Claude Pro carry Claude Code inside the subscription. The Claude help center talk say "all activity in both tools counts against the same limits", meaning chat for claude.ai and Claude Code dey share one five-hour session limit and one weekly limit. So if the Graphify semantic pass run with your assistant model, that one na Claude usage. If you wan understand how those limits dey work first, read how many tokens Claude Pro really include.
You get three ways to control am.
graphify extract . --code-onlydey index only the code. The docs talk say "No model is called and no key is needed." Graphify go print how many files e skip.--backend ollama(orGRAPHIFY_BACKEND=ollama) dey send the semantic pass to Ollama for your own machine, so e no touch Claude at all.GRAPHIFY_BACKEND=none(oroff, ordisabled) dey turn off every model call. The local code pass still dey run.
If you no set backend, the default na auto-detect. That mean Graphify go choose by itself. If you wan dey sure say nothing go call model, set GRAPHIFY_BACKEND=none or use --code-only.
The Ollama route need you to install the extra and pull a model first. The default model na qwen2.5-coder:7b, and you fit change am with OLLAMA_MODEL or --model. Remember say Ollama dey run on your own hardware, so na your machine go do the work. The cookbook advise --max-concurrency 1 if you get only one GPU (graphics processing unit).
uv tool install "graphifyy[ollama]==0.9.108"
ollama pull qwen2.5-coder:7b
graphify extract . --backend ollama --max-concurrency 1Install the correct package: graphifyy get two y
This part important pass. The package for PyPI (Python Package Index) na graphifyy, with two y. The command wey you go type after install na still graphify, with one y. The PyPI page talk am plainly: "Other graphify* packages on PyPI are not affiliated."
Wetin that one mean for you: if you type pip install graphify by mistake, you dey install package wey somebody else upload. Package like that fit carry any code, and that code go run for your machine. That na how typo-squat attack dey work, so check the spelling before you press Enter.
The newest release na 0.9.108, wey come out on 12 November 2026. Before am, 0.9.104 come out on 8 November, so five releases land inside five days. Because e dey change almost every day, pin the version. That way, wetin you test today na wetin go run tomorrow. Graphify need Python 3.10 or newer, but below 3.15.
uv tool install graphifyy==0.9.108
graphify --helpuv tool install dey put the tool inside its own environment, so e no go change your system Python. If you no sabi the difference between uv, pipx and venv, this guide on venv vs pipx vs uv explain am. pipx install graphifyy also dey work if na pipx you dey use.
If your shell talk say graphify no dey found after install, the docs talk say make you run uv tool update-shell (or pipx ensurepath), then open new terminal. The reason be say the folder where uv put the command never enter your PATH.
Connect Graphify to Claude Code
graphify install --platform claudeThat command go register Graphify with Claude Code. The docs talk say "the installer reports the files it writes", so read that list when e print am. If you want am only for one repo, run graphify install --project --platform claude inside the repo. The README talk say project install dey write files like .claude/skills/graphify/SKILL.md.
Then open Claude Code inside your project and run:
/graphify .This one go build the graph. Remember wetin we talk before: through this skill route, docs, PDF and images go through your Claude model. If you only want code, build am from the terminal with graphify extract . --code-only instead.
Wetin the hooks dey change for your session
The install add PreToolUse hooks along with the skill. A PreToolUse hook na small script wey Claude Code dey run before a tool call happen. If you never use hooks before, Claude Code hooks explained show how dem dey work.
The README talk say the Graphify hook "fires automatically before search-style tool calls". For Claude Code e also fire before Claude read source files one by one with Read or Glob, and before Bash readers like cat, head, tail and sed -n. By default the hook only dey advise. E dey tell Claude make e check the graph first, but e no dey block anything. The README also talk say these soft nudges get cap per session: GRAPHIFY_HOOK_NUDGE_CAP, wey default na 5.
Strict mode: wetin happen when Graphify deny a Read
Strict mode no dey on by default. You go turn am on with graphify install --project --strict, or with GRAPHIFY_HOOK_STRICT=1 without reinstall. GRAPHIFY_HOOK_STRICT=0 dey turn am off even if you install with --strict.
When strict mode dey on, the README talk say e "makes Read and dedicated Grep share one denial per session unless a recent graph query has oriented the agent (default 30 minutes)". Make we break am down:
- Na only one denial for the whole session. Read and Grep dey share that one.
- If Claude don query the graph inside the last 30 minutes (the default), no denial go happen.
- Bash searches and Glob still dey advisory. Dem no dey ever deny.
- After context compaction, a
SessionStarthook reset the budget, because the compacted context no longer hold the old nudges.
So strict mode na one-time denial per session. The first time Claude try to read file without to check the graph, e go collect denial, and that denial go push am to run graphify query. After that one denial, Claude fit read normally. Think am well: one denial mean Claude go make another tool call, and that extra turn sef dey inside your usage. The Claude Code cost docs explain say every request dey resend the conversation, so "a one-line question in a session that has been open all day still draws usage for the whole conversation." Long session plus plenty nudges no be free.
Keep the graph fresh after git pull
Graph wey old go give Claude wrong answer. The README talk say make you run this after git pull:
graphify update .E dey re-extract only the code files wey change, and e no use model. If you no wan dey remember, install git hooks:
graphify hook installThis one add post-commit, post-checkout, post-merge and post-rewrite hooks, so the graph rebuild by itself. The docs page on keeping the graph fresh talk say with the hooks you no need run graphify update . after pull. The README still tell you to run am. The two no agree fully, so the safe rule na: run graphify update . after pull, unless you don install the git hooks. Also note say the hooks dey run for background, so query wey you do immediately after pull fit still see the old graph for small time.
Ask the graph with graphify query
You fit query the graph yourself from terminal:
graphify query "payment webhook" --budget 1000The --budget flag dey limit how many tokens the answer go carry. The CLI reference talk say "Matching uses label words, not full sentences." So use short keywords wey dey for your function or file names. No write long question.
How to measure the saving yourself with /usage
Make we clear the numbers first. The "80%" wey you see for the top of the README na promo line for Graphify Cloud, the hosted product. Cloud get Free plan for up to 3 repositories with no card required, and the paid plans start from Pro at $15 a month. The README no show how dem measure am. The "71.5x" come from older copy of the README, as one vendor measurement on one mixed corpus, and that same copy talk say small corpus no go see much compression. Graphify own docs talk say make you "measure savings on your own repo instead of trusting a headline number." Na their own advice, so follow am.
Graphify get graphify benchmark command. E dey compare tokens for your whole corpus with tokens for the answers Graphify return for five typical questions. Graphify own blog talk say "It measures size, not quality." E no tell you if the answer correct, and e no tell you how much you go pay.
The better test na your real usage. Claude Code get /usage command. For Pro, e dey show plan usage bars, activity stats and a breakdown of wetin dey count against your limit. The breakdown dey attribute usage to skills, subagents, plugins and MCP servers. To understand wetin those numbers mean, read Claude Code token usage explained.
- Choose one question wey need many files to answer, like "which functions call the webhook handler and wetin dem do with the result".
- Without Graphify, open new Claude Code session, ask the question, then check
/usage. - Build the graph, run
/clear, ask the same question with the same words, then check/usageagain. - Add the cost of building the graph to the second run, if you build am through
/graphify .with docs inside. - Do am again with two or three different questions before you decide.
The docs talk say the /usage figures na approximate, and dem dey come from the session history wey dey that machine. Usage from claude.ai or another device no dey inside. Because chat for claude.ai dey share the same limit, no use am while you dey test, so the comparison go clean. Whatever result you get na your own result. E fit good, e fit no good, depending on your repo. How you manage context matter too, and managing context in Claude Code show other ways to reduce wetin enter each turn.
Graphify or Graft?
Graphify no be the only tool wey dey build codebase map for coding agents. We don write about another one for Graft, the codebase graph for agents, and the way to judge the two na the same. Ask which part dey run local, which part dey call model and spend your usage, how the map dey stay fresh after pull, and wetin e add to your Claude Code config. Then run the same /usage test for both, on the same question. No choose any of them because of headline number.
Obsidian and Windows: wetin the docs talk
For Obsidian: the README for 0.9.108 no mention Obsidian at all. The CLI reference list a separate export command, graphify export obsidian, wey e describe as "Obsidian vault notes and canvas." So Obsidian na optional export. Nothing wey we read talk say you need am for the Claude Code workflow. Because the flags fit change from release to release, check graphify export --help for the version wey you install.
For Windows: the README table talk say for Claude Code on Windows, graphify install go auto-detect am, or you fit run graphify install --platform windows. To get uv on Windows, the README show winget install astral-sh.uv, and the install docs show the uv PowerShell installer. We no see any Windows-specific limit for the docs.
How to comot Graphify
graphify uninstallThis one remove the registrations, but your graphify-out/ folder go still dey. If you wan delete the graph too, add --purge:
graphify uninstall --purgegraphify hook uninstall only remove the Graphify block from the four git hook files and the merge=graphify line from .gitattributes. We no see any page wey talk clearly say graphify uninstall remove the Claude Code PreToolUse hooks too. So after uninstall, open /hooks inside Claude Code and look .claude/settings.json to confirm say nothing remain. The package itself still dey for uv after all this. Check uv tool --help for how to remove am.
Last thing: the docs site no get version, and Graphify dey ship almost every day. Wetin we write here follow 0.9.108. If you install newer one, check graphify --help and the README for that version before you trust any flag.
FAQ
Graphify go really reduce my Claude Pro usage?
E fit, but nobody fit promise you number. Code indexing dey run local with tree-sitter and no cost Claude usage. Docs, PDF and images dey use your assistant model unless you use --code-only or GRAPHIFY_BACKEND=ollama. Graphify own docs talk say make you measure on your own repo. Ask the same cross-file question with and without the graph, run /clear between the two, and compare /usage.
Why I no fit just run pip install graphify?
The official package na graphifyy, with two y. The PyPI page talk say "Other graphify* packages on PyPI are not affiliated." If you install the wrong name, you dey run code from somebody wey no be Graphify. Pin the version too: uv tool install graphifyy==0.9.108.
I need Obsidian to use Graphify with Claude Code?
No. The README for 0.9.108 no mention Obsidian. The CLI reference list graphify export obsidian as optional export wey make Obsidian vault notes and canvas. The Claude Code workflow use graphify-out/ and the skill, not Obsidian.
Graphify dey work with Claude Code on Windows?
Yes, according to the README. graphify install go auto-detect Claude Code on Windows, or you fit run graphify install --platform windows. Install uv first with winget install astral-sh.uv or the uv PowerShell installer. If graphify no dey found after install, run uv tool update-shell and open new terminal.
Strict mode go block Claude from reading files forever?
No. Strict mode dey off by default. When you turn am on, Read and Grep share only one denial per session, and no denial happen if Claude query the graph inside the last 30 minutes (the default). Bash searches and Glob still dey advisory. Set GRAPHIFY_HOOK_STRICT=0 to turn am off without reinstall.