Open source · MIT · Linux first

Give your AI agent eyes on a real browser, while you watch.

open-eyes is a command line tool that lets an agent like Claude read and drive a visible Chromium window. You see every click as it happens, take over at any moment, and approve anything risky.

Install See the benchmarks

agent shell~370 tokens

$ open-eyes fill --name "Search Wikipedia" Chromium
filled combobox "Search Wikipedia"
changes:
      form
+       combobox Search Wikipedia value=Chromium [focused, expanded] @256
+       listbox Search results @258
+         option Chromium Chemical element with atomic number 24 (Cr) @259
+         option Chromium (web browser) Open-source web browser project @261
          …
$ open-eyes click --name "Chromium (web browser)" --role link
clicked link "Chromium (web browser) Open-source web browser project"
the page changed to https://en.wikipedia.org/wiki/Chromium_(web_browser)

Fewer tokens than a screenshot

An agent reads the page as compact text with references it can act on, and each action reports only what changed. Measured on six live sites against the previous release:

…tokens for what is on screen

…tokens per click that changes page

…tokens of instructions every session

Full results, including where 0.4.0 is not better →

How it works

Reads the page as text

snapshot shows what is on screen, one element per line like button Save @3. --find searches the whole page, iframes included.

Acts like a person

Real mouse and keyboard events: click, fill, select, hover, scroll, upload, by ref or by name.

Shows you everything

The agent works in its own window. An overlay draws each action, and one click on Take control hands the browser back to you.

Your browser or its own

launch opens a separate profile. connect uses your everyday Chromium, Chrome, Brave or Edge with your logins, after you approve once.

Guardrails against prompt injection

You approve risky actions

Buying, sending, deleting, publishing, uploading a file or typing a password waits for your Allow in the agent window. Only real clicks answer.

You choose the sites

open-eyes policy sets allow and block lists, also for content embedded in iframes.

Everything is logged

open-eyes log lists every action and its outcome, never the text typed or the content of uploaded files.

Install

You need Node.js 22 or newer and Chromium, Chrome, Brave or Edge.

npm install -g https://github.com/SchlaftGut0-1/open-eyes/releases/download/v0.5.0/open-eyes-0.5.0.tgz

open-eyes launch
open-eyes open news.ycombinator.com
open-eyes snapshot

Each release has a checksum and a signed build provenance you can verify first. To use it from an agent, copy the bundled skill into your agent's skills directory. The README has both steps.