Documentation — Benzo Agent

Goals

A goal is a plain-English sentence you type into the panel. Benzo parses it into one or more actions, then runs them.

Structure

Chain multiple clauses with then, and, or commas:

Go to wikipedia.org, then extract every heading
Click Sign in, then type admin into the email box
Scroll to bottom and extract all links

What local mode understands

Pattern	Example
Navigate	`Go to bbc.com`
Scroll	`Scroll to bottom`
Search	`Search for wireless headphones`
Fill form	`Type hello into the search box`
Click	`Click the Sign in link`
Extract	`Extract all links`
Press key	`Press Enter`
Wait	`Wait 2 seconds`

Actions

Every action the agent can perform.

click

Click an element by index.

type

Type text into an input.

select

Choose a <select> option.

scroll

top / bottom / up / down.

navigate

Go to a URL.

extract

Return structured rows.

press

Fire a keyboard key.

wait

Pause for N milliseconds.

Settings

Setting	Default	Notes
Planner mode	local	Local or LLM (proxy URL required)
Proxy URL	—	Only needed in LLM mode
Model	gpt-4o-mini	Sent to your proxy
Max steps	25	Hard cap on ReAct iterations
Approval mode	guarded	off, guarded, or paranoid
Auto-export	off	Download CSVs when done

Keyboard shortcuts

Ctrl+EnterRun the current goal

Ctrl+Shift+EOpen extensions page

Screenshots

Click Screenshot in the panel or popup. Benzo scrolls the page to trigger lazy images, freezes fixed elements, resizes the viewport to the whole page height, and captures in a single pass.

Known limits: Chrome caps single images at 16,384 px. If your page exceeds that, zoom out (Ctrl -) and retry.

Recording

Open the recorder from the popup or the panel's Record button. Configure quality, bitrate, and format, then click Start.

In Chrome's picker, click a tab first — the Share button stays grey until you do. Check Share tab audio if you want the tab's sound.

Use Pause and Resume to skip dead air. When you Stop, the file auto-downloads (if that switch is on) and appears in the session list.

Unlocking copy

Click Unlock on any page. Benzo strips user-select:none, removes inline oncontextmenu handlers, disables overlay traps, and installs capture-phase listeners that run before any page script.

The unlock is per page load. If the page reloads or navigates, click Unlock again.

FAQ

Do I need an API key?

No. Local mode uses a rule-based planner. You only need an API key if you want to enable LLM mode, and even then, it's your own proxy.

Does it send data anywhere?

In local mode — the default — Benzo never makes a network request. In LLM mode, only page structure (a text outline of the DOM) is sent to your proxy.

Why does Chrome show a debugger banner?

Full-page screenshots use Chrome's DevTools Protocol. Chrome shows the banner as a security indicator whenever an extension attaches to a tab. It disappears the moment the screenshot completes.

Which sites does it work on?

Any site you can visit. Cross-origin iframes and DRM-locked content (Netflix video, Spotify playback) are the only exceptions.

Can it log me into sites?

Not automatically. It can fill login forms, but 2FA prompts, CAPTCHAs, and bot checks block automated logins.

Everything, in one place.