agent-browser gives an agent a command surface for a disposable session. ego(lite) gives it a browser workspace. ego(lite) completed 93.5% of tasks perfectly in our published benchmark of 31 live-site browser automation jobs, the highest perfect-completion rate measured.
| Feature | ego(lite) | agent-browser |
|---|---|---|
| Default browser state | Your imported Chrome profile: logins, cookies, extensions | Fresh headless profile; reusing real Chrome needs a --profile snapshot copy or a manual CDP connection |
| How the agent acts on a page | Batches multiple actions in a few lines of JavaScript per turn | One command per action; batch mode can queue steps but can't branch on a snapshot mid-run |
| How the agent reads a page | Compressed Snapshot: accessibility tree with stable @N refs | Similar: compact accessibility snapshot with @e refs |
| Browser automation on live sites (31 tasks, same model) | 93.5% perfect, $1.75 per completed task, 6m 38s average | 62.9% perfect, $4.23 per completed task, 10m 05s average |
| Watch or take over a task | Yes. Headful windows by default; click in and finish it yourself | Headless by default; watching requires the separate --headed flag |
| Parallel tasks | Spaces run tasks side by side, never touching your tabs | Multiple isolated sessions via --session flags, managed and monitored from the CLI |
| Reusable skills (coming soon) | Distills successful runs into reusable skills for repeat tasks (limited beta) | No built-in equivalent |
| Setup | Install the app, run /ego-browser in your agent | npm, Homebrew, or Cargo install, then agent-browser install to download a Chrome for Testing build |
| Price | Free, no subscription; ego-browser shell is MIT-licensed | Free, open source |
Last updated Aug 30, 2026
This benchmark tests agent browsing on the live web instead of a sandbox: 31 multi-step jobs on real sites, many behind login walls. ego(lite) and agent-browser ran the identical set.
agent: pi · model: ChatGPT 5.6 Sol · thinking: max · 31 tasks
Every tool got the same 31 tasks: multi-step jobs on live websites, many behind login walls. A task counts only when every requirement is met. No partial credit. Tasks like: pull a week of post analytics from a logged-in X account, lock in a nonstop fare for next Friday, keep a 2048 board alive all the way to the 256 tile.
agent: pi · model: ChatGPT 5.6 Sol · thinking: max · 31 tasks
Failed runs still bill you. This divides total model spend by tasks actually completed, so it prices what a result really costs. Lower is better.
agent: pi · model: ChatGPT 5.6 Sol · thinking: max · 31 tasks
Rubric points earned per thousand model turns: how much finished work each model call buys. Higher is better.
agent: pi · model: ChatGPT 5.6 Sol · thinking: max · 31 tasks
Average wall-clock time per task across all 31. ego(lite) finished tasks 52% faster than agent-browser.
agent: pi · model: ChatGPT 5.6 Sol · thinking: max · 31 tasks
Real-World Bench is an open-source benchmark: 31 complex real-world browser automation tasks on live websites, every tool driven by the same pi agent with the same model (gpt-5.6-sol, thinking effort max) and graded by the same judge against per-task rubrics. Every score covers all 31 tasks, with no per-task cherry-picking.
The distinction is not what the agent can click, but what it can work with. ego(lite) keeps a full local Chromium session, your authorized state, and separate Spaces in view, so the same agent can search, research, extract, fill, test, or handle a longer web task without reducing the browser to one command at a time.
agent-browser's core loop is open, snapshot, click @e2, snapshot again: each step a command the model reads before deciding the next one, and every extra round trip is another chance to lose the thread. On those 31 jobs, more than one agent-browser task in three ends short of a perfect score.ego(lite) reaches the page as a compressed Snapshot with stable @N refs, and the agent acts on several of them in one JavaScript turn. Fewer round trips leave fewer places to stall: 93.5% of the same 31 tasks end perfect, 52% faster than agent-browser on average task time.
Time to finish a task, shorter is better
Data source - Task: finish an average live-site job
agent-browser's parallelism is a set of --session flags you manage yourself, headless by default, so a hung dialog leaves you guessing from CLI output alone. The --headed flag meant to fix that has open bug reports of being ignored, or of Chrome crash-looping instead of showing a window.In ego(lite) there's no cap on Spaces: each runs its own task at once, headful by default, on its own imported Chrome profile. None of them steal your tab, and you can watch or take over anytime.
A typical agent-browser flow is open, snapshot, click @e2, fill @e3, snapshot again: separate commands, each a round trip the model reads before deciding the next step. Its batch mode can queue steps, but the model can't act on a snapshot until the whole batch returns.ego(lite)'s agent writes a few lines of JavaScript that run several steps in one turn, reading the result as a compressed Snapshot. Stacked together, that keeps a whole task well below a command-per-action loop.
The same task, the same model. Left: one batched JavaScript call in ego(lite), 1.1k tokens total. Right: agent-browser's CLI, one command per action, 7 round trips, tokens climb to 13.6k.
Getting agent-browser onto your real Chrome takes setup: a profile name copies your profile into a read-only snapshot, or a persistent path is supposed to carry logins across restarts, though there's an open report of it staying empty on Linux. The alternative, a remote-debugging port, exposes full browser control to any local process on it. Skip the setup and you get a fresh headless profile that stalls at every captcha and 2FA prompt.ego(lite) imports your Chrome profile in one click, live.
If agent-browser convinced you that coding agents should drive a browser directly, ego(lite) is the same conviction with your real browser attached. Switching is minutes, and nothing has to be uninstalled.
Download ego(lite)
Download ego(lite) and import your Chrome profile in one click. Logins, cookies, and extensions come with it, with no --profile flag or profile-directory copy required.
Run your first task with /ego-browser
Paste one sentence into Claude Code, Codex, Cursor, or Gemini CLI and hit enter. No daemon to start, no CDP port to open, no session flag to manage.
Paste into your agent
/ego-browser Open ego.app and check that the pricing page loads
Watch it work
The task runs in its own Space, so your tabs stay untouched. The browser window is real: watch live or take over anytime. Results land back in your agent's CLI.
Keep agent-browser for headless CI checks if it's already wired into a pipeline there. ego(lite) covers browser work on your own machine, with your accounts, visible Spaces, and a full Chromium surface.
Side-by-side comparisons with Puppeteer, Playwright MCP, Claude Chrome extension, Browser Harness, and 7 more.
Free, runs on your computer, imports your Chrome profile in one click. Works with Claude Code, Codex, Cursor, and any CLI agent that writes code.