# The browser > Let the agent open, click through and check the app you build: a browser panel on desktop, a headless browser in the terminal. Local addresses only by default. ## Overview Empryo can open your running app in a browser, click through it, fill in forms, take screenshots and read what's on the page. Use it to check that a change actually works, reproduce a UI bug, or test a flow without doing it by hand. The agent drives the browser through one tool, `browser`, with an `action` argument (`open`, `click`, `snapshot` and so on). It's available on every surface: | Surface | Browser | What it can do | |---------|---------|----------------| | Desktop app | A visible, tabbed browser panel | Everything below, plus audits, network throttling, request recording and a dev-server runner | | Terminal UI and headless | A headless browser with no window | Open, navigate, snapshot, screenshot, click, type, keys, scroll, resize, inspect, run JavaScript, wait, highlight, manage tabs, clear storage | The same prompt works on both. ## Local addresses only, by default The browser opens `localhost` and private-network addresses only. Public websites are blocked. It's meant for the app you're building, not for browsing the web, and keeping it local limits what a malicious page could make the agent do. To allow public sites: - **Desktop:** Settings → Browser network, or `"browser": { "allowExternalNetwork": true }` in config. - **Terminal and headless:** `"terminalBrowserAllowExternal": true` in config.  *A screenshot is large and still leaves the agent guessing a selector. A snapshot lists every element with a short ref, so the click lands on the first try.* ## Snapshots before screenshots `browser(action:'snapshot')` returns every visible interactive element as a short text list, each with a stable `ref`: ``` button "Save changes" ref=e7 * textbox "Email" ref=e3 [value=""] [scrollable "Results" y=0/1840 ref=e12] ``` The agent then acts on the ref, for example `browser(action:'click', ref:'e7')`, instead of guessing a CSS selector from your source code. A `*` marks elements that are new since the last snapshot. That's how the agent checks what a click did without paying for a screenshot, which costs far more tokens than a text list. ## Desktop-only actions The desktop browser panel adds developer-tools features: - `audit`: Core Web Vitals, page weight, accessibility and SEO checks, and horizontal-overflow detection. - `throttle`: slow network (slow 3G, 4G, offline) and CPU slowdown. - `network_start` / `network_stop`: record every request with its status, size, timing and failures. - `run_server` / `stop_server` / `list_servers`: start your dev servers from `.empryo/launch.json` and open them in a tab. You share the panel with the agent and can watch or take over at any time. When several agents work at once, each tab is held by one agent at a time so two agents never type into the same page (see [tabs](/docs/agents/cross-tab)). ## The terminal browser In the terminal UI and in `--headless` runs, the browser has no window. It uses the system web view on macOS and Windows, and Chrome or Chromium on Linux. On Linux, install Chrome or Chromium, or set `BUN_CHROME_PATH` to its path. Without one, the browser tool isn't offered. Screenshots are sent to the model as images, so it can still see the page. To turn the terminal browser off, set `"terminalBrowser": false` in config. ## Config | Key | Default | Effect | |-----|---------|--------| | `terminalBrowser` | on where supported | The browser tool in the terminal UI and headless runs | | `terminalBrowserAllowExternal` | `false` | Allow public websites in the terminal browser | | `browser.allowExternalNetwork` | `false` | Allow public websites in the desktop browser panel | ## Browser vs. computer use This tool drives pages in Empryo's own browser. [Computer use](/docs/tools/computer-use) goes further: it can drive your own Chrome, other app windows or the whole screen, with separate permissions for each.