Browser and terminal

AgenC Desktop has a web browser and a terminal that open in a column beside the chat. You can use them yourself. On macOS and Linux the agent can also drive the same browser and terminal, so you watch it click through a page or run a command instead of reading about it afterwards.

Open the browser or the terminal

The buttons at the top right of a chat open and close them: Open the browser and Open a terminal. They stack in the side column, next to the agents panel when it is open. Each has Widen the panel and Narrow the panel, and a close button.

The browser

The browser works like a small web browser:

• tabs, with New tab and a close button on each;
• Back, Forward, Reload, and the address box ("Type a URL");
• Draw on the page and send it to the chat: mark up the page, and the app attaches a screenshot of your drawing to your next message;
• Pick an element and send it to the chat: click something on the page, and the app adds its address, selector, size, text and HTML to your message;
• Viewport: Responsive, "Mobile 375 x 812" or "Tablet 768 x 1024";
• Browser options, described below.

When an agent tool is driving the page, the toolbar shows an "agent" badge.

Browser options

• Open file…: open a local file in the browser.
• Save screenshot: saves the visible page as a PNG and shows it in its folder.
• Manage sites…: every site the panel has loaded, yours and the agent's. Block stops a site loading there, including redirects into it. Clear browsing data is in the same dialog.
• Open chat links here: open links you click in the chat in this panel instead of your default browser. Off by default.
• Keep cookies between runs: keep your sign-ins after the app restarts. On by default.
• Clear browsing data.

Downloads and screenshots go to a folder named .agenc/downloads inside the session's project folder.

When a page tries to open another application, for example a mailto: link, the app asks first: "Open (application)?", with Cancel and Open.

Tabs the agent opens

Tabs the agent opens are kept apart from yours. They accept web pages only (no local files or custom links), keep their cookies separate from other conversations, get no access to your camera, microphone or other devices, and cancel automatic downloads. When the agent needs to download a file, it uses Core's own file tools, which follow your permission mode.

The terminal

The terminal runs your own shell in the session's folder: the shell set in your account (zsh by default on macOS, bash on Linux), or on Windows the shell that COMSPEC names, usually Command Prompt. It keeps a dark background whatever the app theme.

• The first tab is named "Terminal", the next ones "Terminal 2", "Terminal 3" and so on. A tab that a Run button opens is named after its command, starting with "Run:". New terminal adds a tab and Close this terminal ends one.
• Closing the terminal panel does not end its shells; they are there when you open it again. Closing a tab, or the window, ends its shell.
• When the agent types in a shell, the tab says "The agent is typing in this shell".
• The Run button on shell commands in the chat runs them in this terminal; see Writing messages.

When the agent reads the terminal, the app hides values that look like secrets: variables named like keys, tokens, secrets and passwords, and private key blocks. What you see in the terminal is not changed.

What the agent can control in the app

On macOS and Linux every local session gets a private set of tools for the app itself. With them the agent can:

• read the app's state: supported settings, your local sessions and your projects;
• change presentation settings (theme, glass, palette, sizes, wrapping, line numbers, notifications) and the goal limits, and open a Settings section;
• list, create, change, run, cancel and delete routines, and read their runs;
• open, rename, pin, archive and unarchive sessions, and pick a project for the next session;
• maximize, minimize, restore, resize or close the window;
• drive the browser: open and close tabs, navigate, read the page, click, type, scroll, take screenshots, read downloads and the console;
• list, open, read, run commands in, type in and close terminals.

The agent only changes app state when you ask. Normal approvals apply to these tools, and plan mode refuses the ones that change something. Running or typing in the visible terminal needs a session in Bypass permissions, because that shell is not inside Core's sandbox; the agent never changes your permission mode to get there.

These tools cannot touch keys or credentials, grant permissions, change configuration, delete projects, end agents or change the active model. A routine that runs on a schedule can read routines but not create or change them.

The tools belong to one window and one conversation. A second window or another conversation cannot use them, and closing the window revokes them.

On Windows

You can use the browser and the terminal yourself on Windows. The agent cannot drive them there: "Desktop controls for agents are unavailable on Windows." The agent keeps Core's own tools.