In-app browser
A browser panel beside your terminals, dev servers that open themselves, and letting an agent drive the page.
⇧⌘B opens a browser panel next to your terminals. It is a WebKit view inside Agentty on macOS and a Microsoft Edge WebView2 (Chromium) view on Windows; on Linux links open in your default browser instead. WebView2 comes with Windows 10 and 11 — on a PC without it, opening the browser offers to download and install it from Microsoft first.
Opening things in it#
- ⌘-click a link in any terminal (Ctrl-click on Windows and Linux).
- Settings → Browser decides where links go. With In-app browser (local pages only) a dev server, a local page or a file opens here, and anything on the web opens in your default browser, where you are already signed in. With Default browser every link goes there.
- Type a URL, or a search term — the search engine is configurable.
With links set to Default browser, the in-app browser is off: the globe button is hidden, ⇧⌘B does nothing, file previews open in your default browser, agents get no browser tools, and Settings → Browser shows only the link choice. Plugins' own pages, and the sign-in windows they open, still open inside Agentty.
Dev servers#
A local server started in a tab appears as a :port chip in the workspace list and opens in the panel as soon as it answers with a page.
Closing the tab that started a server stops the server. Turn either behavior off in Settings → General and Settings → Browser.
Responsive mode#
The phone button beside the address bar lays the page out at a device's size instead of the panel's: iPhone SE, iPhone 15 Pro, iPhone 15 Pro Max, Pixel 8, Galaxy S24, iPad mini, iPad Air, iPad Pro 12.9", Laptop, Desktop — or any width and height you type. The page is centred and scaled to fit whatever room the panel has, so a 1920-wide desktop layout is legible in a narrow panel. Rotate swaps the two numbers, and dragging the right or bottom edge resizes by hand.
Turning it off returns the page to the panel's own width.
Sign in with Chrome#
On macOS, when a page asks you to sign in, a bar above it offers the sign-in you already have in Google Chrome. Choose Always to have it looked up by itself from then on, Just this once, or No to leave that site alone until Agentty restarts. The first time, macOS asks whether Agentty may use Chrome's storage key: choose Allow.
- If it works, the page reloads signed in, and the bar says which Chrome profile it came from.
- If Chrome has nothing that works, the page is covered and blurred until you finish signing in in Chrome, usually within a minute of doing so: it then carries over by itself. Sign in here brings the page back so you can sign in in it, and Open in Chrome opens the page there.
- If you sign out here, the sign-in is not loaded back on its own for 30 minutes: the bar asks first.
Change it in Settings → Browser → Sign-ins: Ask, Always or Off, and which Chrome profile to use when you have several. Otherwise the profile last used comes first, and the next one is tried if it doesn't sign in.
Only the cookies a browser would send to that site are copied, into the in-app browser (or that terminal's own cookies when they are kept separate). Nothing is read for a plugin's page or in private mode, and an agent can't answer the bar or ask for a site: it only sees how things stand, in signIn of browser_status. With Always, though, a sign-in page is enough, whoever opened it: an agent that opens one gets in if Chrome is signed in to that site, even in a tab you aren't looking at, and then acts in that session. Use Ask if you want to decide each time.
Some sites tie a sign-in to the browser it was made in and refuse the copy, and then you sign in here once. On Windows this isn't available: Chrome there ties its key to Chrome itself.
Letting an agent drive it#
Claude Code and Codex can control this browser through MCP: click, type, read the console, take a screenshot. It is how an agent checks its own work on a running page.
The tools an agent gets are:
| Tool | What it does |
|---|---|
browser_open | Open this terminal's tab (made the first time), optionally loading a URL |
browser_navigate | Load a URL, host:port, or search words |
browser_status | Current URL, title, and whether the page is still loading |
browser_wait_load | Wait until loading finishes |
browser_click | Click an element |
browser_console | Read the page's console output |
browser_screenshot | Capture the page |
browser_viewport | Lay the page out at a device size, WxH, or off |
browser_close | Close this terminal's tab |
They are registered with the agent when Agentty starts it, so there is nothing to install or configure. A typical loop is: the agent starts the dev server, opens the page, clicks through the change it just made, reads the console for errors, and fixes what it finds.
One browser per terminal#
Every terminal has a browser of its own. Open the browser in one terminal and go to a site, open it in another and go to another: selecting a terminal shows its browser, as it left it, and a terminal that never opened one shows none. Pages of the terminals out of sight keep running — loading, scripts, screenshots — so work in one never waits for you to look at it.
- Links and servers go to their terminal. A link clicked in a terminal, the port chip of a server it started, and a server that opens by itself all open in that terminal's browser. A port chip on another workspace's card selects that terminal.
- An agent works in a tab of its own in its terminal's browser, made on its first
browser_openand labelled with the terminal's folder and number (app #3). Everything that agent does goes to that tab only: it never changes another terminal's page, a page you opened, or a plugin's page. Its page is laid out at 1280 px (or the size set withbrowser_viewport) whether it is on screen or not, and scaled to fit the panel when it is, so what it checks does not depend on how wide your panel is. - Limits keep it light. At most 6 terminals' browsers run out of sight, and all in-app browser pages together use at most 30% of the computer's memory (both in Settings → Browser). Beyond either, the browsers out of sight are unloaded, the one seen longest ago first — never the one on screen, nor one whose agent sent a command in the last two minutes. An unloaded page keeps its address and loads again when its terminal is selected or its agent sends a command (which then waits for the page to load).
- Closing the panel hides the terminal's browser; it comes back as it was when you open it again. Closing a terminal closes its browser.
- Cookies and sign-ins are shared, so you sign in to a site once. To give each agent's tab cookies of its own, turn on Settings → Browser → Separate cookies and sign-ins for each terminal (macOS 14 or later; on Windows, WebView2 Runtime 101 or later); they are removed when the terminal closes.
The browser keeps the sessions you are signed into. An agent driving it acts inside those sessions. Enable it when you want that, and be aware of which tabs are open.