You describe the goal in natural language. Crystal plans the steps and carries them out with a virtual mouse and keyboard on a computer you connect. Because it operates through the UI, it can help even when an app has no API.

Examples include gathering data from web apps, updating spreadsheets, filing forms, triaging inboxes, downloading and renaming reports, reconciling records across systems, and other repeatable multi-step workflows.

Crystal reasons about what’s on screen and interacts like a human, so it’s not limited to predefined APIs or brittle selectors. If the UI changes (e.g., a button moves or a label updates), Crystal aims to adapt rather than break.

If a person can use it on the connected computer, Crystal can likely use it too. You can also combine actions across multiple apps in one workflow.

No. You decide when Crystal can take control. Sessions are permission-based, and you can pause, stop, or revoke access at any time. We recommend granting the minimum access needed for each task.

Crystal is designed with consent and transparency in mind: you approve sessions, and you can review what it did. It won’t attempt to bypass 2FA or CAPTCHAs, and we encourage using secret managers or browser password vaults rather than exposing plaintext credentials on screen.

By default, Crystal retains only what’s needed to execute your task and provide an audit trail. You can configure retention to meet your policies, including redaction of sensitive fields and automatic deletion after completion.

Highly dynamic pages, unexpected pop-ups, strict rate limits, or long-running jobs may require guardrails or human review. For regulated actions (payments, identity changes, production deploys), we recommend a human-in-the-loop step.

Yes. You can templatize common tasks, add pre-checks and post-checks, and require approvals for risky steps. Crystal can also log actions for auditing and hand off to a human when ambiguity or policy triggers occur.