ego lite asks you to share a browser with an agent
ego lite's first-launch migration can give an agent your existing Chrome logins, cookies, extensions and bookmarks. That is the consequential choice in this product. If an assistant needs authenticated websites while you continue browsing, the shared-browser design is worth evaluating. Decide which accounts it should reach before accepting the migration prompt.
The README describes agents working in separate Spaces while the person keeps their own tabs. It advertises faster automation and fewer tokens. Our evidence supports neither comparative claim: the recorded run exercised the Node helper package, with no timed browser workload. The useful question is whether the browser's access model fits the tasks you intend to delegate.
What happened when we ran it
Our ego lite helper install succeeded in 4 seconds, the build in 8 seconds and tests in 29 seconds at commit bacaece on November 8, 2026. The run targeted ./package/ego-browser/ and recorded 49 dependencies with 68 MB on disk. No passing-test count accompanies the successful test command, so there is no justified count to print.
The npm audit reported zero known vulnerabilities in that run. That result concerns the audited dependencies. It is not a browser permission review, a check of downloaded application artifacts or proof that model-generated scripts cannot access sensitive information. No authenticated task, account-isolation exercise or long-running browser session appears in these measurements.
The configured sandbox had 3 CPUs, 8 GB RAM and image label lab-node:22. The actual runtime and operating-system versions need the raw toolchain logs. The checkout contained 254 files, about 57,505 source lines and 5.2 MB. The scan found five CI workflow files and a tests directory, with no Dockerfile. Those observations describe the checkout rather than the installed desktop browser.
The helper package does not contain the whole browser runtime
The helper README assigns tabs, task spaces, CDP transport, snapshots and event delivery to the browser runtime. The package supplies agent-facing helper functions. Its output is an ESM bundle used by that runtime, so a successful package build cannot establish that the complete browser application works.
This split also explains the licensing boundary. The top-level README says the repository contents are MIT licensed and the browser is a separate, free download. Do not infer that this checkout contains everything needed to reproduce that application. Teams requiring a fully reproducible browser distribution need a separate answer before adopting it.
For an individual, the documented route is to download the app or add its ego-browser skill and follow the installation flow. The README names Claude Code among compatible agent tools. It says macOS and Windows are supported while Linux remains on the roadmap. A successful helper command in a sandbox does not turn that roadmap entry into Linux desktop support.
Spaces are insufficient evidence of account isolation
Issue 319 reports cookie access across browser profiles through Storage.getCookies and Storage.setCookies on macOS application version 0.4.7.1. The reporter distinguishes page-level cookies, which the report says were isolated correctly, from the browser-global CDP operations. We did not reproduce the report, and the application version is different from the repository commit in our run.
In an August 31 reply, a maintainer says spaces do not isolate cookies and describes profile-based space creation as something planned for a later application release. The issue remains open in the inspected record. Neither that plan nor the open label establishes the behavior of today's download. The evidence does establish that treating a Space as a verified security boundary would be unwarranted.
I would decline ego lite for mutually untrusted agents sharing sensitive accounts unless the deployed version's boundaries had been checked independently. Even a secondary profile is not verified containment by this review, because the report concerns access across profiles. For a first trial, choose a supervised task with no sensitive session. Browser convenience is a reason to evaluate access carefully, especially when the selling point is reuse of existing logins.
Browser actions still need application-level checks
The helper documentation says fill verifies that editing took effect, without guaranteeing that application formatting preserves the requested string byte for byte. It leaves business postconditions to explicit reads or waits. That is a useful limitation: filling a form field does not establish that the final submitted value matches the intended value.
Use that boundary to design the first task. Let the agent populate a disposable form, inspect the rendered values and stop before an irreversible submission. If a price, date or account identifier is reformatted, the workflow must check the result. This is a proposed acceptance exercise, not a browser task completed in our lab. Our passing helper checks leave that application behavior unanswered.
Choose a different control model when the desktop browser adds little
Browser Use documents an open-source agent that writes Python to act in a real browser, with local and cloud execution. Consider it when you want the agent loop in your own application. Its current README also offers a CLI, so a simple library-versus-CLI comparison would miss part of its scope.
agent-browser is a native Rust CLI that returns compact page snapshots and lets an agent act through shell commands. Choose it when that command interface fits your existing tooling. Neither alternative was measured alongside ego lite, and this review establishes no speed or isolation winner.
GitHub records an ego lite push on November 10, 2026. Release v2.2.0, published November 5, lists screenshot-path and page-ledger fixes alongside actionability work. Those changes and current repository activity show maintenance, while issue 319 leaves a separate access question to verify. Keep the helper's successful checks in their proper scope and make the account-access decision before importing your everyday sessions.