AI on your Macno cloud*, no per-run fees
Click through your web app. A vision AI model on your Mac names what you click. Replay any time and see exactly what broke. No code to write.
curl -fsSL https://breakpatch.dev/install | shBetaFree and open source · Apple Silicon · macOS 14 or later · 16 GB memory
* No cloud? Mostly sunny, with a light chance of cloud. Your screens and the AI stay on your Mac, and Breakpatch goes online only to download the AI model and the test browser once, for updates, and for anonymous usage counts you can turn off. Team workspaces (coming later) sync through the cloud, because that's how a team shares tests.
tests run with Breakpatch
The AI
The AI looks at your app's screen the way a person does, and it runs on your Mac. It never needs to see your code, and nobody has to write test scripts.
For developers: Qwen3-VL 4B, an open-weights vision-language model, run on Apple Silicon through MLX. No selectors.
Named steps
Click something and the AI says what it is, so the test reads “Click New project”, not a pair of coordinates. Each step also keeps a plain What to look for, like “Done button, bottom right of the Create project dialog”. Edit it if it's wrong.
Editing stepsPrivate by design
The model runs on your Mac, so your screens never leave it. There's no cloud service, no API key to manage and no fee per run. Once it's downloaded, the AI works offline.
For developers: Qwen3-VL 4B, open weights, run on Apple Silicon through MLX.
About the AI assistantSelf-healing tests Team, coming later
When a button has moved, the AI finds it again and the run carries on. The report marks the step Fixed automatically and shows the old and new positions. Nothing changes in the test until someone accepts it.
How fixing worksWhy on-device AI
AI testing usually happens in someone else's cloud. Breakpatch runs the test browser and the model side by side on your Mac, which changes what testing costs and what it can reach.
localhost. No tunnel, and no vendor to let through your firewall.| Question | Typical cloud AI tool | Breakpatch |
|---|---|---|
| Where the model runs | The vendor's servers | Your Mac |
| Your app's screens | Sent to the vendor | Stay on your Mac |
| Running more tests | Metered by run or credit | No per-run fees |
| Apps behind a VPN | Tunnel or allow-list | Work as they are |
| The model | Updated by the vendor | Open, the same every run |
| API keys | Yes | None |
If you can use your app, you can test it.
Open your app in Breakpatch and use it the way a person would: click, hover, drag, type, upload. Each action becomes a step, and the AI names what you clicked.
Press Run, or run a whole suite. Breakpatch does each step again and checks the screen before and after every one. In Team, coming later, the AI also finds buttons that moved.
The first thing that's wrong stops the run. The report shows what was expected next to what was on screen, why, and what to try.
Screen checks and the report
After every step, Breakpatch checks the screen changed the way it did when you recorded it. There's nothing to draw or approve. When a run fails, the report puts what was expected next to what was on screen, and says why in plain words.
Visual checking, not AI. Screen checks compare how the screen looks with how it looked when you recorded, leaving out the parts of a page that change by themselves, like clocks and carousels. The same screen gets the same answer every time.
Plain files
In Community, every test is a readable JSON file in a folder you choose. Keep it next to your code and Git is your history: a change in the app shows up as a small, readable diff.
breakpatch.json apps/web-app/ app.json tests/create-a-project.json tests/log-in-and-out.json shared/log-in.json suites/smoke.json
Local runner and CI Team, coming later
Pick one Mac to be the local runner. It runs suites on a schedule and whenever anyone presses Run on runner, with one queue for the whole team. Start suites from CI, a script or an automation tool, and get the result wherever you want it.
breakpatch-ci
$ breakpatch-ci run --test tests/log-in.json { "name": "Log in", "result": "pass", "source": "ci", "durationMs": 21840, … } $ echo $? 0
Your data
The test browser and the AI model both run on your Mac, so your screens, steps and saved secrets stay there. In Community there's no Breakpatch server holding your tests.
Community is free and open source, for one person on one Mac, and it's what you can install today. Solo adds the automation for one person. Team, which shares everything and runs it for you, is coming later. Its price isn't decided yet.
One person, one Mac.
$19 billed monthly · 1 person · 1 machine included
Just you, with the automation.
Pricing to be decided
Your whole team in one shared workspace.
Business, for 20 people or more, is coming later too: everything in Team, plus 5 machine licences, support within 4 business hours (Monday to Friday) and invoice billing. Tell me when it's ready.
| Feature | Community | Solo | Team (coming later) |
|---|---|---|---|
| Record by clicking | Yes | Yes | Yes |
| The AI names what you click | Yes | Yes | Yes |
| Screen checks and the report | Yes | Yes | Yes |
| Shared steps and suites | Yes | Yes | Yes |
| Where tests are kept | A folder | A folder or Git, or your own workspace | Hosted by Breakpatch, or your company's own Firebase project |
| History | Through Git | Through Git, or built in with a workspace | Built in |
| People | 1 | 1 | Your team |
| Copy a failure as Markdown | Yes | Yes | Yes |
| Self-healing tests (Fixed automatically) | No | Yes | Yes |
| Why did this fail? (the AI explains) | No | Yes | Yes |
| Create issue in GitHub, Linear or Jira | No | No | Yes |
| Local runner and schedules | No | Schedules on your Mac; the local runner with a workspace | Yes |
| CI command line and run requests | No | CI command line; run requests with a workspace | Yes |
| Result messages to Slack and Teams | No | Yes | Yes |
Community is free and open source under Apache 2.0, with no time limit and no account. Solo, for one person, is $19 a month, or $16 a month billed yearly. Team and Business, the paid plans for working together, are coming later, and their prices aren't decided yet. See pricing.
A vision AI model that runs on your Mac and reads the screen. It's about 3 GB, downloaded once during setup, and unloaded when it's idle.
For developers: Qwen3-VL 4B, an open-weights vision-language model, run on Apple Silicon through MLX.
The AI names the steps you record. In Team, coming later, it also finds buttons that moved and says why a step failed. Screen checks don't use it: they're a visual comparison. See The AI assistant.
It's the only model for now. Bringing your own local one, through Ollama or LM Studio, isn't available yet. Cloud AI models aren't planned: your screens stay on your Mac.
Only to download the model once, during setup. After that it runs offline, on your Mac. There are no API keys and no per-run fees, in any edition.
Your tests, screenshots and saved secrets don't. The test browser and the AI model both run on your Mac. In Community your tests are files in a folder you choose, and nothing of them reaches our servers. Team, coming later, keeps them in a shared workspace: in your company's own Google Firebase project (see Create a workspace), or hosted by Breakpatch.
Community sends anonymous daily totals of tests created and runs, with no names, addresses or IDs. You can turn that off in Settings → Privacy. The documentation lists exactly what's sent.
A Mac with Apple Silicon (M1 or later), 16 GB of memory and macOS 14 or later. Intel Macs aren't supported. The AI model takes about 3 GB of disk space.
Linux comes next, then Windows.
Paste one line into Terminal and Breakpatch is in your Applications folder a minute later. No password, no admin rights. After that it updates itself, so you're always on the latest version.
Want to see exactly what it does first? Read the script or the install guide.
Never your tests. Community sends only anonymous daily counts of tests created and runs, which you can turn off, and has no licence to check.
Team, when it comes, confirms its licence once a day and keeps working for 30 days without a connection. It sends the licence, who is using it (the person's work email, or the machine's name for a runner) and simple counts, such as how many tests ran and passed. Never test names, web addresses, steps or screenshots. The documentation lists every detail.
Not with Community. Team, coming later, adds two ways. Your pipeline can add a run request, and the local runner runs the suite and sends the result to Slack, Microsoft Teams or any web address. Or run breakpatch-ci run --test on the CI machine: it tells your pipeline whether the test passed, so a failing test stops the build. Both will work with Codemagic, GitHub Actions, Bitrise and Jenkins. See Run requests and From CI with breakpatch-ci.
Install Breakpatch and record your first test in about 10 minutes.
curl -fsSL https://breakpatch.dev/install | sh