From a sentence to a verdict you can forward.

Ryvetz turns a written flow into an executed test on a machine we operate, and hands back video, a trace of every action, and a judged verdict for every step you gave a criterion.

  1. 01 Describe

    One flow, in plain English, with a criterion on the steps that matter.

  2. 02 Run

    A hosted runner opens a supported application and drives it on a test account.

  3. 03 Judge

    A separate model grades each criterion from the screenshot alone.

  4. 04 Read

    Video, trace, and a cited verdict, in the dashboard or as a CI check.

It starts with a sentence

Write the test the way you would say it to a colleague. Every step is grounded on the live screen at run time, so nothing depends on selectors or coordinates.

Illustrative example

flows/invite-teammate.yaml
name: Invite a teammate and confirm the seat
app: acme-crm
account: staging

steps:
  - do: Open Settings › Members
  - do: Invite dana@example.com as an editor
  - do: Send the invite
    verify: The members list shows dana@example.com as Pending within 10 seconds
  - do: Revoke the invite
    verify: dana@example.com is no longer listed within 10 seconds
The Ryvetz Atlas map for the example product, Acme CRM: 16 screens across 6 paths, drawn as a tree of thumbnails from the sign-in screen. Illustrative example.
Every run feeds the Atlas: a map of the screens and controls Ryvetz has already seen, so the next flow grounds faster.

Plain-English steps

A step is what to do and, when it matters, what must be true afterwards. No page objects, no locators.

See it on a plan

Grounded on pixels

The executor looks at the real screen and finds the control the way a human would, so a moved button is not a broken test.

See it on a plan

Recovers from drift

When the interface changes, the agent sees the new screen and continues instead of failing on a missing selector.

See it on a plan

A verdict that cites pixels

The agent that clicks never grades itself. A separate strict judge sees only your criterion and the screenshot, and quotes what it used before anything passes.

Illustrative example

Executor sees
  • The live screen
  • The step to perform
  • The Atlas of known controls
Clicks, types, waits. Never grades.
Judge sees
  • Your verification criterion
  • The screenshot. Nothing else.
Issue found

“The avatar for dana@example.com failed to load: its URL reads ‘…acme-crm.example/avatarsundefined/’, missing the path separator.”

A pass it is not confident in is recorded as a fail.

Split roles

Executor and judge are different models with different instructions. One acts, one decides.

See it on a plan

Criterion and screen only

The judge never sees the executor’s reasoning, so it cannot be talked into a pass.

See it on a plan

Strict by default

A pass the judge is not confident in is recorded as a fail, with the evidence attached.

See it on a plan

“Yes, one broken element: the avatar for dana@example.com in Settings › Members failed to load. Its URL 'https://cdn.acme-crm.example/avatarsundefined/dana.png' is missing a path separator.”

1 issue found Recon health check · Acme CRM · illustrative example Answered from the screen, with 2 citations

Hosted machines we operate

Your laptop is not the test machine. Every supported platform ships as a curated image on a runner we run, with a session guard that refuses the wrong screens.

Illustrative example

web-runner-01 Ready
Image
chrome-stable · acme-crm-profile
Slot
1 run at a time, yours next
Accounts
Test and demo only
Session guard
Refuses log-out and real-money screens
dedicated-runner Enterprise
Tenancy
Single-tenant, never shared
Image
Your golden image
Accounts
Real-money and live accounts

Nothing to install

You author flows in the browser or from the CLI. The runner, the browser, and the platform profile are ours to keep working.

See it on a plan

Test accounts only

Credentials are encrypted at rest, injected only for the length of a run, and never shown back in the UI.

See it on a plan

Dedicated when it matters

Accounts with real money behind them, such as eval, funded, and live trading accounts, belong on a single-tenant runner under Enterprise. Never on a shared slot.

See it on a plan

Evidence your pipeline can gate on

Video, screenshots, and a typed trace for every run, plus JSON and exit codes. On Team, every pull request gets a check run.

An example recon in the Ryvetz dashboard: the recorded Acme CRM video with step, action, pass and fail markers on its timeline, and the recon panel reporting 1 issue found, a broken avatar with a malformed URL, boxed on the frame. Illustrative example.
An illustrative finding: the recon answers the health check from the screen, cites the frame, and boxes the broken element on it. Video, timeline, and transcript stay attached.

Video and trace

Every action is recorded and typed. Scrub to the step, read what the agent did, see what the judge cited.

See it on a plan

CI check runs

Install the GitHub App or call the webhook from your pipeline. Fail the release when the judge does.

See it on a plan

Share links

Time-limited capability URLs for a run, so a release manager can read the verdict without an account.

See it on a plan

Ship with evidence.

Join the waitlist. Describe one flow, watch it run on a hosted machine, and read the verdict.