Browser pane: a real Chromium browser in your workspace
A full Chromium tile with DevTools, subtabs, native H.264 video, a page annotator, and a CLI that lets agents click, fill, and screenshot pages. macOS and Windows.
The browser pane is a full Chromium-based browser tile inside atrium. It is a real browser, not a headless automation surface — pages that refuse WebKit (Google sign-in, Gmail, Figma sign-in, Zoom meetings) work normally, and H.264/AAC video plays natively, so YouTube, streaming sites, and most embedded web video work out of the box.
Available on macOS and Windows. Shortcuts below are written with macOS keys; on Windows, Ctrl generally takes the place of Cmd, and Settings → Keyboard shows the exact binding for each action.
Opening a browser
Cmd+Shift+Bopens a browser pane in the current room.- Click the Browser tile in the toolbar and drop it into any split zone.
- In an existing browser pane, any
window.open()call creates a grouped subtab on the same pane, so OAuth popups and multi-window flows keep working. - A browser pane can dock as a chromeless picture-in-picture resident — a compact URL bar, visible across every room, with per-pane keyboard zoom (
Cmd+=/Cmd+-) to fit it into the slot.
The UI
Every browser pane has a chrome strip at the top:
- Back, Forward, Reload buttons.
- Security indicator in the URL bar lead icon — green lock (valid cert, no mixed content), amber lock (cert errors or insecure sub-resources), warning triangle (plain HTTP), globe (internal pages like
about:andchrome://). - URL bar with an autocomplete dropdown. Typing blends visit history, open-tab titles, and live Google search suggestions. Arrow keys select,
Enternavigates. When the bar is not focused,https://and trailing slashes on bare domains are elided (Chrome-style); focus or click to edit the full URL. Non-http(s)schemes (file://,chrome://,about:) navigate as-is — nohttps://prefix is auto-injected. - Color-scheme button next to DevTools cycles System / Light / Dark. The choice is global across all browser panes, persists across restarts, and swaps what the page sees for
prefers-color-schemewithout changing atrium's own theme. - DevTools button opens Chrome DevTools in its own window (Elements, Console, Network, Sources, Performance). Right-click anywhere in the page and choose Inspect to do the same.
Keyboard
All standard browser shortcuts work inside the pane:
Cmd+R— reload.Cmd+Shift+R(or shift/cmd-click the reload button) hard-reloads, bypassing cache.Cmd+T— open a new room. For a subtab on the focused pane, useCmd+Ctrl+T(Ctrl+Alt+Ton Windows), then choose Browser from the launcher. Links and popups can still open browser subtabs directly.Cmd+[/Cmd+]— back / forward.Cmd+Left/Cmd+Rightare equivalents.Cmd+L— focus URL bar.Cmd+=/Cmd+-/Cmd+0— zoom in / out / reset. Pinch-to-zoom works on trackpads. Per-site zoom is remembered.Cmd+F— find in page.
Page annotator
Turn on comment mode and the browser pane becomes a review surface for the page itself: click an element to pin a comment to it, exactly as you would leave a per-file comment in the diff-review surface. Comments live in a closed shadow root in your theme's colors, so they never collide with the page's own styles.
- Comments are stored per URL and persist after you navigate away or close the pane.
- The Send dialog dispatches them to an agent or director. It reads the whole per-URL store and renders an origin → page checkbox tree, so one send can cover several pages at once, each as its own section, with the captured page title as the label. Each comment goes with its id and the command to reply to it.
- Each comment is a thread. Click its pin to read it: agents' replies and yours, who wrote each and when, a reply box, and Resolve / Reopen. A pin with a red dot has a reply you haven't opened yet; so does its row in the comments list.
- Resolving hides the pin and leaves the comment out of the next Send. Resolved comments stay in the list, and Show resolved on the page brings their pins back so you can reopen one.
- Comment mode is disabled on docked browsers.
atrium browser annotations list [--url <page url>]
atrium browser annotations reply <id> "<text>"
atrium browser annotations resolve <id> [--reopen]
atrium browser annotations delete <id>
Agents answer comments with reply; the reply shows on the page signed with the agent's name. Replying never resolves a comment, and it keeps working while you drive the browser. Comments made before threads existed have numbered ids that can repeat across pages; add --url <page url> when a command says so.
An open pane reconciles live when the store changes.
History and persistence
- Per-pane navigation history — back/forward arrays persist across restarts.
- Global visit history — the history drop-down in the URL bar is retained across restarts and used for autocomplete.
- Per-site zoom — your last zoom level for each domain is restored on return.
Subtabs
When a page opens a new tab (via target="_blank", window.open(), or middle-click), atrium creates a subtab on the current browser pane rather than a floating window. Subtabs sit in the pane's header strip:
- Click a subtab to switch to it.
- Drag a subtab out onto a split zone to promote it to its own pane.
- Close with the
×on the subtab.
Sign-in flows
Passkeys (FedCM/WebAuthn), Google Sign-In, and Bluetooth-pairing prompts work inside the pane the same way they do in Chrome. Bluetooth in particular has a one-time native permission prompt — atrium handles the macOS TCC handshake on the main thread so the prompt actually surfaces instead of crashing the engine. Cookies and sign-ins from the earlier WKWebView-based browser do not carry over to the current Chromium engine; sign-ins made before the upgrade must be redone once.
Background behavior
Browser panes throttle their CPU usage while their room is inactive, dropping host CPU when web content isn't visible. When the room is in any state other than focused, the underlying CEF NSView is hidden inside its wrapper view by a watchdog so it never intercepts pointer events meant for atrium chrome. Switching back to the room restores the pane to its previous URL, scroll, and zoom — and hit-testing IPCs from atrium overlays batch across visible panes only, so context menus and dropdowns stay smooth even with many browser panes open.
On macOS, Settings → General → Browser panes → Unload hidden pages can release a page's memory after it has been out of sight for 30 minutes. This is off by default. Returning reloads the page where you left it. Pages with typed input or detected audio are protected except under critical memory pressure; downloads, camera or microphone use, DevTools, and active agent use keep a page loaded. Some sites' audio cannot be detected, so turn the setting off if background audio stops.
Crash recovery
If Chromium's render process crashes, the pane shows a recovery screen with a Reload button. Closing a browser pane while it is loading does not exit the app. Rapid reloads and URL-bar enters are coalesced so navigation work does not pile up.
On Windows, the same recovery surface reports This page stopped drawing if the page stops producing frames. Reload stays available; the notice clears when drawing resumes.
Driving the browser from scripts and agents
Every browser pane is scriptable from the atrium CLI. This is how agents drive pages alongside editor and terminal panes, and how you can script browser work from your shell. The automation surface is comprehensive — DOM snapshots, click/fill/type/press/select, scroll, eval, wait conditions, screenshots, cookie management.
# Snapshot the accessibility tree (interactive elements get refs like e1, e2)
"$ATRIUM_CLI_PATH" browser snapshot <pane> [--interactive true] [--selector CSS]
# Click, fill, type, press, select
"$ATRIUM_CLI_PATH" browser click <pane> e3 [--snapshot-after]
"$ATRIUM_CLI_PATH" browser fill <pane> e5 "hello"
"$ATRIUM_CLI_PATH" browser type <pane> e5 "hello"
"$ATRIUM_CLI_PATH" browser press <pane> Enter
"$ATRIUM_CLI_PATH" browser select <pane> e7 "value"
# Scroll, evaluate, wait, screenshot
"$ATRIUM_CLI_PATH" browser scroll <pane> --dy 400
"$ATRIUM_CLI_PATH" browser eval <pane> "document.title"
"$ATRIUM_CLI_PATH" browser wait <pane> --selector ".done" --timeout 10000
"$ATRIUM_CLI_PATH" browser screenshot <pane> --path out.png --full-page
# Read DOM and element state
"$ATRIUM_CLI_PATH" browser get <pane> text e5
"$ATRIUM_CLI_PATH" browser is <pane> e5 visible
# Cookies
"$ATRIUM_CLI_PATH" browser cookies <pane> [--domain example.com] [--name session]
"$ATRIUM_CLI_PATH" browser cookies <pane> --delete session
# Raw Chrome DevTools Protocol — what eval can't reach
"$ATRIUM_CLI_PATH" browser cdp <pane> Accessibility.getFullAXTree
browser cdp is the escape hatch for what eval can't do — network inspection, device/media emulation, the full accessibility tree — as a one-shot request→response (no event streaming). See CLI → browser for every flag.
Constraints
- macOS and Windows only. Other platforms return an error stub today.
- DevTools keyboard events (Enter, Tab, arrow keys inside a DevTools input) are routed through the DevTools process, not atrium — they affect DevTools, not the page.
atrium browser consolereads the buffered DevTools console messages (500 most recent per pane). Object arguments are stringified by the renderer, so structured payloads land as[object Object]— usebrowser evalif you need to round-trip structured data.