Questions, answered honestly.

Running tests on your own machine or CI is free and unlimited. In the cloud, 1 credit is 1 cent: a run of up to 20 website tests is 2 credits (2¢), and writing a test with our AI is 10 credits. The Free plan has 100 credits a month with no card, and paid plans start at $5 a month with your own AI key ($8 a month with ours too). Writing a test with our AI is 10 credits (5 more per further 15 steps over 15); with your own key it is 0. Every feature is on every plan, except cloud CLI runs, which are coming soon to paid plans, and there is no seat pricing. Pricing.

You might not. Playwright's planner, generator and healer are good, free and open, and if your team is happy owning Playwright code they may be all you need. We are different in what you keep: the test is an English file anyone can read, each Expect is a checked-by-code assertion, heals are proposals you review, and the same file runs on Android. We also host the browsers and emulators, PR checks and inboxes, so there is nothing to set up. And because every website test exports as a Playwright spec, choosing us doesn't close the other door. Playwright test agents vs us.

Yes. Stay on Free and top up from $7 when you need credits, at 100 credits per dollar, usable for 12 months. Compared to a plan you get fewer credits per dollar, one cloud run at a time, results kept 7 days, and no safety buffer when your credits run out. Pay as you go.

No. The first run works out the steps and records them. Later runs replay the recording with no AI, so a cloud run costs a couple of credits and the same result every time. AI comes back only for a new or changed step, or when your app changed and a step needs healing, and then it proposes a fix for you to review. How a run works.

No. Tests are plain English, one step per line, or you describe the test in a paragraph and let the app draft it. Engineers can add exact steps or Playwright code where they want to. The test format.

Yes. Connect GitHub and every pull request is checked in our cloud with one comment that updates in place and a status check; a test that couldn't run is neutral, never a false pass or fail. Or run the open GitHub Action in your own runners with your own key. GitLab CI, CircleCI and Bitbucket use the same CLI. GitHub Action.

Yes: an MCP server, a CLI with JSON output, and instructions for agents that include the rule never to edit an Expect line to make a test pass. Coding agents.

No. Each Expect line becomes a typed check that plain code evaluates, and the verdict is computed from the checks. The AI works out how to do a step, never whether it passed. A heal can change how a step is done, never what is checked, and it is shown to you for review. Checks and verdicts.

It depends on how you write. Step-by-step tests are the reliable path. On Bench's tidy, step-written shop tests the model authored 11 of 11 tests correctly, with 0 false passes and 1 wrong fail in 73 replays that should have passed. Messy prose and speech are worse: in the same run, a terse one-line description authored only 6 of 11. A wrong fail is annoying; a false pass is dangerous, so "zero false passes" is a claim we make only for step-written tests, not for prose or speech. Those runs are single runs, so read them as measurements, not guarantees. The Bench numbers and method.

Writing a test uses AI; replaying one does not. When Optestra's own AI writes your test, the request goes through OpenRouter to DeepSeek-V4.1-Flash, pinned to DeepInfra (US, fp8) with zero data retention, and with no fallback to any other provider. It receives the step text, a text description of the page and sometimes a screenshot. Secrets are typed by the test runner and never sent to the AI. If you use your own AI key instead, the request goes straight from the runner to your provider and not through us. Security and data.

Nothing breaks. Your tests are Markdown files in your own repository, the engine that runs them is open source (MIT) and runs on your machine or in your CI, and every recorded website test is also a plain Playwright spec that runs without Optestra. The cloud, the app and hosted AI are the paid parts; your tests don't depend on them. If Optestra disappears.

The agent can only reach the domains you allow, enforced below the browser. It has a closed set of actions, no shell or file access, types secrets it never sees, and treats page content as untrusted data. Production environments block destructive actions unless a test declares them. The safety model.

Results and evidence are kept for your plan's retention period (7 to 90 days) and then deleted. You can delete a project or your account and its data at any time, or ask us to. The sub-processors we use are listed with their purpose and location. Subprocessors.

Cloud runs reach public sites and preview deploys, including protected previews. For localhost or a private network, run the open-source CLI or the GitHub Action on your own machine or CI runner, which is free. The Optestra desktop app, which also runs local tests, comes after launch.

Yes, in our cloud or on your own machine: your APK on a fresh emulator for every run, Android 13 to 17, with the same English test format. It costs 2 credits per test (at least 10 per run). iOS isn't supported. Android.

Email codes and links work with a hosted test inbox per project. SMS numbers aren't offered at launch. When a flow sends a code to your own phone, a run you are watching asks you for it in a six-box prompt and types it for you; when nobody is watching (CI, PR checks) the test is marked blocked, never failed and never charged. Codes from a person.

Web, Android, CLI tools and Electron apps, in our cloud or on your own machine. Native desktop apps are coming, on your own machine. You can also compare versions: run the same tests on two versions and get the regressions in one report. Native Windows, macOS and Linux apps are coming, on your own machine. Compare versions.