browser & computer use
Eyes and hands,
on a short leash
The agent can read the web, test your site and operate the apps on your Mac — in a browser and a session you can watch, with approvals at every step and a stop shortcut that always wins.
Photograph — Vasily Ledovsky via Unsplash
A browser built into the workbench
Orvlet's own browser sits next to your code and chat. You use it like any browser; the agent uses it only with your say-so, in a profile kept apart from yours.
- Browser UseThe agent navigates, reads, clicks and fills forms in the built-in browser, with a separate profile per task or per worker.
- Supervised sessionsA supervised session pauses sensitive actions when you leave its tab or window and needs a deliberate resume; temporary signed-out sessions keep no logins and discard their storage when they end.
- Site rulesAllow and deny lists for browsing, plus domain lists for web fetch and web search, decide where the agent may go at all.
- Point at the problemPick an element on a page, comment on it or tweak its CSS live, and send the element, its styles and component context to the agent.
- Developer tools for the agentPage snapshots, console output, screenshots, navigation and DevTools let it debug the app you are building.
- Background checksThe agent can verify a page in the background and report the scoped result, revealing the browser only when you want to look.
- RecordingsRecord a browser session as a GIF or capture a test with its logs, for a pull request or a bug report.
- Opt-in browser memoryKeeps selected browsing context for recall only when enabled, with per-site visibility and deletion.
- Clean web readingWeb fetch keeps a page's main content and drops navigation, banners and footers before the model reads it.
Computer use on macOS
Computer use lets the agent operate other apps on your Mac through a bundled helper. It is off by default, macOS-only, and asks for each app before touching it.
- Per-app approvals, a visible indicator while it works and a global shortcut that stops it instantly.
- Without taking over: approved apps can be operated in the background through accessibility actions or a separate display, so your cursor and focus stay yours.
- Live preview: the owning window shows what the agent sees, from its own screenshots and periodic refresh frames.
- Understands the screen: it reads the accessibility tree and, when that is thin, can ask a local vision model trained for user interfaces where to click.
- Careful with drags: a drag must start inside the target window and end on a connected display, or it is refused before anything moves.
- Multiple displays and full-screen apps are supported; a locked screen is a hard stop.
Your Mac, as context
Small ways to hand the agent exactly what you are looking at.
- AppshotsCapture the frontmost window straight into a chat.
- Apple appsTools for Apple apps under the macOS permissions you have already granted.
- Clipboard toolsLet the agent read or write the clipboard when a task needs it, with your approval.
- Selected text, anywhereA global shortcut or Services entry sends the selected text of any app to a saved prompt and can replace it with the result.
- Finder quick actionsStart a task on files right from Finder.
- Teach modeThe agent highlights the interface step by step so you learn to do it yourself, instead of having it done for you.
Web pages are untrusted input. Approvals show the exact URL or action before anything happens.
Photograph — Ahmed Atef via Unsplash
Prompt injection, taken seriously
Anything the agent reads on a page or screen is treated as untrusted content, never as an instruction. Tool definitions are scanned for injection, the embedded browser runs with default-deny permissions, and every risky action still comes to you as a plain approval.