Let an agent use the browser and terminal

An agent in an Admiral session can open a web page in that session's Browser pane, read it, click and fill it, and open a terminal tab and run commands in it. You see the same tabs while it works, and you can use them yourself.

What you need

• macOS. The Browser tools drive pages on macOS. Linux and Windows support is in progress; on those systems terminals work, and page tools refuse with a clear message.
• A Claude Code, Codex, or ACP session. These agents get Admiral's tools automatically. OpenCode and Cursor sessions do not have them yet.
• Full access for terminals. An Admiral terminal is a normal login shell with no sandbox, so only a session with full access can open or type in one. Browser tools work at every access level.
• The Admiral window open. The browser lives in the window. With no window, the agent is told to ask you to open Admiral.

Ask for it in plain words

Open http://localhost:3000, sign up with a test name and email, and tell me what the confirmation page says. Then open a terminal, run the test suite, and tell me if it passes.

The agent opens a tab in the Browser pane, reads the page, fills the fields, submits, and reads the result. Then it opens a terminal tab, runs the command, and reads the output. Each step shows in the transcript as a tool call.

If the session is the one on screen, the new tab appears and becomes active. Your keyboard focus stays where it was. If the session is in the background, the tab opens there and waits for you.

Sign in for the agent

The agent never types a password or a one-time code. When a page needs one, the agent asks you:

1. Open the tab (or let the agent open it).
2. Sign in by hand in that tab.
3. Tell the agent you are done. It waits for the signed-in page and continues.

Each project keeps its own website data. Sign in once per project, and the agent's tabs in that project stay signed in, also after a restart. Another project does not see that sign-in. To sign out everywhere in a project, use Clear site data for this project in the browser's Browser options menu.

What an agent can reach

• Only the tabs in its own session's Browser pane: the ones it opened and the ones you opened there. Tabs in other sessions are out of reach.
• Only terminals that an agent of its own session opened. Your own terminals stay private.
• Each browser tab has a robot button in its toolbar (Block agent access). Click it and the agent is refused on that tab until you click it again (Allow agent access).
• An agent can open at most eight tabs per session.

Page text and terminal output are data for the agent, not instructions. A page cannot give the agent orders.

Turn it off

Settings → Connections → Admiral tools has a switch for Browser and one for Terminal. It also sets whether the agent may run scripts in a page: never, on local pages only (the default), or on every page. With every page, an agent could read any field, including a password field.

Next steps

• Use terminals: the terminal pane itself.
• Drive mobile simulators: the same idea for phone screens.
• Run child agents: the rest of the tools agents get.