> For AI agents: the complete documentation index is available at /en/llms.txt, the full documentation bundle is available at /en/llms-full.txt.

# Browser Automation and Terminal AI Access in Octop

Octop's **Workbench** puts a remote browser and an AI-assisted terminal side by side in the web UI, so you can watch and take over what an agent is doing in real time. Remote desktop and remote phone have their own pages.

***

## Browser

The browser tab drives a real Chromium instance over CDP with persistent profiles, so agents can navigate, fill forms, and capture screenshots — and you can step in at any point.

### Enable the browser

The browser requires the `browser` extra. Install it using the method that matches how you installed Octop:


```bash One-line installer (macOS / Linux)
curl -fsSL https://download.octop.cloud/octop/scripts/install.sh | bash -s -- --extras browser
```
```bash PyPI
pip install "octop[browser]"
```


After installation, restart `octop run`, then open:

```text
Workbench → Browser
```

Click **Check** to confirm the environment is ready, then **Launch browser** to start a session. Type a URL in the address bar to browse; bookmarks and the AI assistant are available alongside.

Each Octop user gets an isolated Chrome profile. Click **Shut down browser** when you want to stop the local Chrome process. Octop also reaps an idle local Chrome after 30 minutes (`OCTOP_BROWSER_IDLE_TIMEOUT_MINUTES`). Login cookies stay on disk, so the next launch keeps the session.

The Docker image does not ship Playwright Chromium. Install the `browser` extra if you need this tab.


![Workbench browser](/images/en/ui-workbench.png "Workbench browser")

### Use the browser in chat

Ask your agent to browse in plain language. For example:

- _"Go to news.ycombinator.com and give me the top 5 headlines."_
- _"Take a screenshot of [https://example.com](https://example.com)."_
- _"Fill in the sign-up form at \[URL] with these details…"_

When an agent is driving the browser, you can take over the same session. Your clicks and typing take over until you hand control back.

***

## Terminal

The **Terminal** tab in the Workbench embeds a full interactive shell in the browser. You can open several session tabs, run commands directly, and ask the agent to suggest commands, explain output, or diagnose errors — without leaving Octop.

Workbench **Terminal** and the chat Dock terminal share the same session, so commands and output stay aligned whichever surface you use.

```text
Workbench → Terminal
```


![Workbench terminal](/images/en/ui-terminal.png "Workbench terminal")

***

## Related control pages

- **Server desktop** — [Remote desktop](/en/guide/guides/remote-desktop.md)
- **Phone** — [Remote phone](/en/guide/guides/remote-phone.md)
- **Cloud servers** — [Tencent Cloud CVM](/en/guide/guides/cloud-hosts.md) on the last workbench tab

***

## Guardrails

Because the browser and terminal give agents real reach into your machine, Octop centralises the safety controls under:

```text
Management → Security
```


![Security](/images/en/ui-security.png "Security")

The page is split into six tabs:

| Tab                       | What it controls                                                                                 |
| ------------------------- | ------------------------------------------------------------------------------------------------ |
| **Tool approval**         | Tools that pause before executing and wait for you to approve or reject them in the conversation |
| **File access**           | Which paths agents may read and write                                                            |
| **Sensitive information** | Redaction of secrets and personal data                                                           |
| **Command protection**    | Allow and deny rules applied to shell commands                                                   |
| **Skill scanning**        | Checks run against skills before they are installed                                              |
| **Audit log**             | A record of guarded actions                                                                      |

Tool approval ships enabled for the highest-risk tools — `bash`, `execute`, `write_file`, and `edit_file` — but the master **Enable manual tool approval** switch is off by default. Turn it on and any tool in that list pauses for your confirmation before it runs.

Saving applies every tab at once, and running agents reload automatically afterwards.


> **WARNING**
> 
> Review command protection and file access rules before giving agents shell access in a production environment. Overly permissive rules could allow an agent to run destructive commands. Start restrictive and expand only as needed.
> 



> **TIP**
> 
> To enable or disable skills from the CLI, see [`octop skills`](/en/cli/agent-commands.md#octop-skills).
> 

