Your AI software studio · runs on your machine · works 24/7

Describe the app.
Your AI team builds it.

The slowest part of every AI workflow is the human copy-pasting prompts in the middle. Crinkle deletes that job: agents plan, write real code, test, and review security — briefing each other around the clock. No coding. No babysitting.

Get started free

Free & local · npx crinkle · bring your own AI or use a local model

Crinkle — “Build me a recipe box app” team working
How it works

From one sentence to a working app

You describe what you want in plain language. The team takes it from there — and you watch it happen live.

01 — Describe

Tell Mira your idea

“A budgeting app with charts.” “A landing page for my bakery.” Mira, the manager, turns it into a plan with a clear definition of done.

02 — Build & test

The team runs itself

Forge writes real code. Verity clicks through the running app like a user. Failures route straight back to Forge — the agents prompt each other, so nobody has to sit there ferrying messages between AIs.

03 — Ship

Watch it live, then ship it

A live preview updates as they work. When every acceptance criterion passes — and a security review clears — you download it or deploy in one click.

Truly autonomous · 24/7

The bottleneck isn't the AI.
It's babysitting the AI.

Agents today are capable — but most tools still make you the manager: approve this, re-prompt that, shuttle output from one AI into another. Crinkle replaces the micromanager. Mira assigns work, reads every report, reroutes failures, and keeps the team moving day and night — while you sleep, work, or live your life — until the goal is genuinely met.

9h unattended 132 steps 0 times it needed you
Steer anytime, never required. Drop a message mid-run — the team folds it in at the next step without stopping.
Guardrails you set. Spend caps, instant Stop, per-change approvals if you want them, and restore points for every turn.
Honest by design. "Done" is only declared when the tester passes the real, running app — and you get a full run report of every decision made on your behalf.
Survives the real world. A model times out or rate-limits at 3 AM? The agent waits a minute, retries, then walks its fallback chain — one flaky provider never kills an overnight run.
Queue the next one before bed. Line up several app ideas; when one finishes, the next starts automatically as its own project. You get a desktop notification per finish and a run history — duration, steps, criteria met, cost — in the morning.
Overnight run — salon booking app unattended
Meet your team

Seven specialists, one goal

Each agent has a real job — and they hold each other accountable. Nothing ships until the tester and the security reviewer say so.

MI

Mira

Manager

Owns the plan. Assigns work, reads every report, and won’t call it done until it truly is.

FO

Forge

Engineer

Writes complete, runnable code — real files, real dependencies, precise edits.

VE

Verity

Tester

Runs your actual build and clicks through the app in a real browser. A fail is a fail.

PI

Pixel

Designer

Shapes the interface — layout, color, and polish, with an eye on the screenshots.

SC

Scout

Web operator

Drives a real browser to handle the web — research, sign-ups, and site tasks.

AD

Advisor

Strategist

Brainstorms fresh approaches when the team gets stuck, before anyone bothers you.

SE

Sentinel

Security

Reviews every build and vets every command before release. High-risk means no-go.

Proof, not vibes

Most AI builders generate.
Crinkle proves.

A chat window can't check whether the code it wrote actually works — Crinkle can, and does, on every cycle. Before your app is ever called done, it has to survive the gate:

1
The real toolchain runs. Type checks, lint, build, tests — executed for real on your machine. A red build can never be declared finished, no matter what the model claims.
2
Every criterion is proven in the running app. The tester agent clicks, types the invalid email, and watches for the error message. A criterion she saw fail cannot be graded as met.
3
Your app writes its own regression suite. Each proven criterion is recorded as a replayable test that re-runs on every later change — so a fix on Tuesday can't quietly break what worked on Monday.
4
It gets attacked before you see it. Empty submits, absurd numbers, unicode, double-clicks — hostile input is thrown at every build, and any uncaught error blocks completion.
5
"Live" data has to actually be live. A stock, weather, or news app that never touches the network is showing the model's stale memory — Crinkle detects it and fails the build until real data flows (or the app honestly labels its samples).
6
Security signs off last. Code review plus a dependency scan before anything is called finished — and every command an agent wants to run is vetted first.
✓ Only then does the team say done — with a full run report of the evidence.
Your models, your team

Assemble your best team from the AIs you already have

Crinkle doesn’t lock you into one model. Connect the subscriptions you already pay for, any API key, or a completely free local model — then put the right brain on the right job.

Subscriptions

Use the plans you already pay for

ChatGPT (via Codex), Claude (via Claude Code), or your Google account (via Antigravity) — no extra API bill. One click installs the tool, signs you in, and puts it on the roster. Even two accounts of the same service, side by side.

API keys

Any provider, one adapter

OpenAI, Claude, Gemini, OpenRouter, Groq, DeepSeek, Mistral, xAI, NVIDIA NIM, Together — paste a key and it’s on the roster. Every token is costed (live per-model pricing where available) with a breakdown per model and per agent, and a spend cap you control.

Free & local

Run on a local model

Point Crinkle at Ollama, LM Studio, or Open WebUI and the whole team runs on your hardware — $0, fully private, works offline. Connect any vision-capable model and the testers borrow its eyes automatically to judge how your app looks.

Every agent is tunable: its own model (Forge on Opus, Verity on something fast and cheap), its own temperature and parameters, and a fallback chain up to 25 deep with automatic retries. Or flip on smart routing and Mira weighs each model’s strength, price, and health to assign every task herself.

Why Crinkle

Built to never hand you something broken

The quality gate above is the safety net — these are the features that make the studio a joy to run day to day.

🧪

Tests like a real user

Verity launches your app in a real browser and proves every acceptance criterion by using it — typing the invalid email, watching for the error message. No criterion counts as met without observed evidence.

🛡️

Security review built in

Safe Mode reviews code and scans dependencies before anything is called finished, and vets every command an agent wants to run.

👁️

Live preview

Watch your app render and hot-reload beside the chat while the team edits — dev servers, full-stack backends, or plain HTML.

🔌

Smart model routing

Route each agent by hand, or let Mira assign every task — weighing each model’s strength, price, and current health, and routing around providers that start failing.

Restore points & history

Every turn is snapshotted and every change is git-committed. Roll anything back — even without Git installed.

🔒

Private by default

Crinkle runs on your machine. Your projects, keys, and credentials stay local and encrypted — agents never see your secrets.

💸

Costs you can actually see

One click breaks spending down per model and per agent — input tokens, output tokens, dollars. Local models and subscriptions honestly count as $0.

📈

Progress you can trust

The %-to-done is computed from real checkpoints — plan steps completed and criteria verified in the running app — never a made-up number. Reasoning models get a dedicated thinking channel, so work chat stays readable.

Get started

Up and running in one click

No signup wall, no cloud lock-in. Create an account when you want one — it works either way.

⬇  Download for Windows

Free · ~100 MB · installs in one click and opens as a desktop app

⚠️ Alpha build — Windows will warn you the first time

Crinkle is early software and isn’t code-signed yet, so Windows SmartScreen won’t recognize it. That’s expected — here’s how to install:

  1. Open the downloaded Crinkle-Setup.exe
  2. When “Windows protected your PC” appears, click More info
  3. Click Run anyway — Crinkle installs itself and opens as a desktop app
Or run it with Node.js — Windows, macOS, Linux
$ npx crinkle
Needs Node.js 18+ · opens in your browser at 127.0.0.1 · same app, no install