chekd
walks your app · stays on your disk

your verification tool.

chekd gives you a second set of eyes: it walks your app, surfaces the broken or ugly parts, and tries to learn your taste with every decision.

Run it yourself
$ npx chekd <url>
or
Hand it to your AI
Setup prompt for your agent

Paste into Claude, Codex, Cursor, Antigravity. full setup →

see how it works ↓
localhost:5173 / checkout
walking
brokenbtn.checkout

“Pay now” does nothing on click — handler never bound.

findingh2.pricing

“undefined” is showing in visible text — leftover from a missed variable?

quietedtooltip nit

You dismissed this last chek — staying quiet.

learning your taste
sharper every run
The correction flywheel

Generic design audits tell everyone the same thing. chekd learns what bugs you.

Every time you correct a finding, chekd cements your taste. The longer you run it, the more irreplaceably yours it becomes — a loop no off-the-shelf audit can copy.

ChekCheck

It walks your app

chekd opens the cockpit and walks your shipped app, surfacing dead, broken, and ugly findings as cards.

Dispatch

You correct it

Fix it, snooze it, dismiss it, or move to the next. Each call is you teaching chekd what you actually care about.

Learn

It cements your taste

Every correction is remembered. The noise you ignore quiets down; the things you care about get sharper.

Chek againCheck again

The next chekcheck is yours

Run it again and it’s quiet where you said to be quiet — and louder only where you’ve cared before. Sharper every chekcheck.

Every chekcheck, it nags a little more like you.

chekd gets more yours the longer you use it — until skipping it feels like shipping with your eyes closed.

The autonomy slider

One slider, four stops.

Choose how much is you and how much is chekd — from quietly logging while you drive, to fixing and re-chekingre-checking for you. Land anywhere; every stop hands the calls back to you.

Stop 1

Share live

You drive and talk; chekd rides along in the browser, logging what it sees — every click observed, every error caught, your narration pinned to the moment you said it. It changes nothing. No fix, ever, at this stop.

Stop 2

Scan live

chekd drives; you review. It walks your app and surfaces findings as cards — you snooze or dismiss, and there’s no fix affordance at all. The first hand-off on the trust ramp.

Stop 3

Fix live

chekd drives; you correct and dispatch. Fix it writes a real fix on a qa/ branch in your own repo — main never moves; you review and merge it yourself.

Stop 4

chekd ON

Correct, then unleash: you approve the fixes; chekd applies them, re-walks, and verifies — bounded, never “until perfect.” Anything new it finds comes back to you.

More chekd never means less you.

Every stop up the ramp is something chekd does for you, never instead of you: it only fixes what you corrected, it re-cheksre-checks what it touched, and whatever it finds next lands back in your queue.

How it works

Paste one line. ChekCheck your app or PR. Kill the slop you shipped.

chekd is the verification cockpit for apps your AI wrote — it walks the running app, maps changed files into a test plan when a repo or PR is available, and keeps you in the chair for every fix.

Step 01

Paste one line

Drop npx chekd into your AI setup or your terminal. No chekd account, no upload to a chekd service, no config to wade through.

Step 02

ChekCheck the surface

chekd opens the cockpit and walks your shipped app. Add a repo, branch diff, or GitHub PR when you want changed-file context too.

Step 03

Dispatch a disposition

Each finding is a card: fix it, snooze it, dismiss it, or move to next. You stay in the chair the whole time.

Principles

Built to be yours — and to stay yours.

Three commitments shape every part of chekd, from the first run to the thousandth.

01

Accessible to all

Real value at $0, forever — no required API key, no expertise gate. Plain language, never gatekeeping.

02

Make it your own

Dark + light and three themes on day one. Choose Claude, ChatGPT, local Ollama, or no model. Set the autonomy yourself — from watch-only to fix-it-for-me.

03

Compounding

Every use makes it more yours and better. It learns your taste, your corrections compound, and the next chekcheck is already sharper — the flywheel.

Your model · your machine

Choose the model path. See the boundary.

Nothing cloud-backed is ever selected for you.

Claude via Claude Codeyour Anthropic plan and terms apply
ChatGPT via Codexyour OpenAI plan and terms apply
A local model (Ollama)runs on your device — no per-use bill Local · no cloud
Your own provider keyuse once or native vault — a ceiling for every paid job
Google / AntigravityExperimental — text-only, explicit opt-in
Local records · explicit cloud use

Your CLI, your account, your call — chekd never picks a cloud for you.

chekd keeps findings, corrections, screenshots, and history local — never uploaded to a chekd service.

A model-backed action sends only what it needs to your selected provider (Antigravity is text-only). Provider terms apply to every cloud route.

Run a local model and nothing leaves at all. That’s the genuinely airgapped path — zero bytes off your machine.

Pricing

Free forever. Pay for the model-backed DESIGN pass and bounded fixes.

Everything objective is free. Free remembers and reuses your decisions; Plus adds new DESIGN observations and safely applies and re-checks fixes.

Free forever
$0

unlimited local chexchecks — no account, no card, no upload

  • Unlimited local chexchecks — no quota, $0 every time
  • Share & Scan autonomy — you drive, or chekd walks and flags
  • FUNCTION · CONTENT · LAYOUT lenses — dead clicks, “undefined” text, broken layout
  • PR/diff-aware review planning + baseline runtime security
  • Policy profiles & local evidence exports
  • Claude, ChatGPT, or local Ollama — cloud provider plan terms apply
  • Corrections, priors & taste controls — your decisions stick and reorder future findings

Free is the floor, not a trial.

Cancel anytime — your taste stays yours. Your artifacts and corpus are plain local files; you always keep read and export access. 14-day money-back on the first charge. Refund policy →
How it talks

A trusted copilot, never a critic.

Subjective findings are suggestions; only deterministic ones get stated flatly. Honest about limits — and it never judges the work you made.

What it says

  • The “Pay now” button doesn’t fire — want me to wire it up?Concrete, names the spot, hands the decision back.
  • This label is hard to read on that surface. Bump it one step?Plain language a non-technical builder gets.
  • This PR drops a defensive header on checkout. Want the evidence pack?Specific to the changed surface; no scare language.
  • You dismissed this kind of thing before — want it quieted for good?Shows it’s learning you, chek over chekcheck over check.
  • I might be wrong here — ignore if this is intentional.Honest about limits; never authoritative.

What it never says

  • Critical accessibility violation detected. Fix immediately.Alarmist, gatekeeping, treats a finding as a verdict.
  • Your design is bad. Here’s what a real designer would do.Judges the creative work. Never do this.
  • 17 issues found across 4 severity tiers (see dashboard).Noisy, jargon-heavy, dumps work on the user.
  • SOC 2 compliant after one scan.False compliance claim. chekd can produce evidence; it cannot certify you.
  • AI-powered QA automatically fixed your app.Overrides the builder; hype, not honesty.
Copied to clipboard