前段时间我在 GitHub 上翻到一个项目：ego-lite，一款面向 AI Agent 人机协同的浏览器，专门用来解决 Claude Code / Codex / Cursor 这类 AI 智能体做网页自动化的痛点，让人和 Agent 共享同一套登录会话，互不抢占标签页， 我自己用了一段时间，觉得很实用，分享给大家。
这个项目，记住几个关键信息就行：
- 项目：ego-lite，Citro Labs 出的AI Agent浏览器，Agent 在隔离工作区里跑 Web 自动化，直接复用你 Chrome 里的真实登录态
- 技术栈：JavaScript，仓库 MIT 协议，装完是一个桌面浏览器 + ego-browser 命令行
- 热度：18373 star，写稿这天实测
- 迭代：今年 4 月建仓，v2.1.0 是 10 月 16 号发的，写稿当天仓库还有提交
- 平台：macOS（Apple Silicon 和 Intel）、Windows 都有安装包，Linux 还在路线图上
- 商业：浏览器免费、无订阅；官方拿 31 项真实网站任务跟 playwright-cli 等工具对比，自称快 47% 到 63%，这类口径没法核验，听听就行
- 地址：github.com/citrolabs/ego-lite
它干的事很简单，用一句话就能概括：
ego-lite 是一个给人和AI Agent共用的 Chromium 浏览器，Agent 在自己独立的 Space 里干浏览器的活，登录态、Cookie、扩展直接继承你的 Chrome，但互不抢占标签页，开箱即用、无需复杂调试远程调试协议。
核心能力，一个一个说
1、Space 隔离工作区（最核心设计）
内置独立隔离的 Space 空间，每个浏览器任务开一个独立 Space，你自己的浏览标签在主空间，AI Agent 在专属 Space 里面跑自动化任务；多个 Agent 可以并行跑多个网页任务，互不干扰，不会抢占你正在使用的页面、不会乱切你的窗口。你随时可以切入 Agent 的 Space 查看进度、接管操作，处理验证码、滑块这类人工校验场景。
2、直接复用你的登录态、Cookie
支持一键导入 Chrome 全部配置：账号密码、Cookie、登录会话、扩展、书签。Agent 执行网页操作时，直接继承你已经登录好的状态，不用每次自动化从头登录、绕验证码、处理 SSO，这也是对比 Playwright/Browser-Use 最大优势。
3、原生 Agent 调用能力 ego-browser skill
不需要写复杂 CDP、webdriver 脚本，Agent 通过/ego-browser指令，用自然语言下发任务（填表、抓取页面、点击、截图）；底层是 Chromium 内核原生快照，而不是 JS 注入垫片，可以稳定识别 Shadow DOM、跨域 iframe、第三方组件，很多传统自动化工具抓不到的元素它可以识别；并且批量执行页面动作，减少 LLM 工具调用次数、节省 Token、提升速度。
4、数据本地留存
浏览 Cookie、登录信息、页面内容全部保存在本机，不上传到云端，不需要注册账号，只需要你自己配置模型 API Key。
5、可以同时当日常浏览器使用
内核是 Chromium，Chrome 扩展、标签组、无痕模式、书签全部兼容，日常上网和 Agent 自动化可以在同一个浏览器里完成，不用同时开 Chrome + 无头浏览器两套环境。
ego-lite 和传统 Chrome 的核心区别
很多人看完上面，可能会产生疑问，貌似ego-lite好像就是一款浏览器啊，那它和传统Chrome浏览器有啥区别呢？别急，用一表张格，带你弄明白。
补充一句话核心差异： Chrome 是给人用，自动化是外挂能力；ego-lite 是从底层设计成：人和 AI Agent 一起共用的浏览器。
做测试的人为什么会需要它
Web 自动化做了这么多年，脚本本身早就不是难点，卡人的经常是第一步，登录。公司内部系统清一色 SSO，跳转七八层，外加短信验证码、扫码、风控滑块。Playwright 的 storageState 方案能把登录态存下来，可服务端 token 一轮换就失效，测试账号三天两头被风控锁掉。功能验证五分钟的事，登录环境维护两小时。
所以大量内部系统的验证至今靠人肉点。每天早上打开十几个后台，看服务状态、对数据、查配置，想做自动化，拦路的全是登录。
另一头，现在也有人直接把这类活派给AI Agent，给 Claude Code 挂个 playwright-mcp 或者 chrome-devtools-mcp，让它去驱动浏览器。毛病集中在三处。起出来的通常是全新空白实例，登录态没有，还是得先伺候登录；每点一下都是一次工具调用，上下文每轮重发一遍，token 烧得快；要么 headless 看不见过程，要么Agent和你挤同一个窗口，它一动你就得停手。
ego-lite 恰好把这两头的堵点各解掉一半。登录态用继承的，Agent拿真实会话干活；执行放进 Space，看得见也管得住。
ego-lite VS Playwright / Browser-Use（传统自动化框架）
Playwright、Browser-Use 本质是自动化 SDK 库，需要代码调用，拉起独立无头浏览器进程；ego-lite 是浏览器底座，自带可被 Agent 调用的运行时。
- Playwright：每次启动全新干净浏览器上下文，没有登录态，很容易触发风控、验证码；依赖 CSS 选择器，前端样式一改，脚本直接挂掉；适合稳定 CI 流水线、标准化 E2E 回归。
- ego-lite：复用真实浏览器指纹和登录态，适合探索式测试、后台页面校验、AI Agent 网页测评、人工 + Agent 协同测试；不用维护大量稳定选择器，遇到复杂登录系统成本极低；不替代 Playwright，而是互补。
一句话总结： Chrome 是人用的浏览器，自动化是附加功能；Playwright 是写脚本驱动无头浏览器；ego-lite 是一个专门让 AI Agent 和你一起共用的浏览器。
安装与快速上手
方式一，直接下 app。
macOS 有 Apple Silicon 和 Intel 两个 DMG，Windows 有 exe，官网首页（https://lite.ego.app/）就有。
第一次启动问你是否迁移 Chrome 数据，点同意，onboarding 走完后 ego-browser 命令就注册进 PATH，一般在 ~/.local/bin。
方式二，只装 skill。
npx skills add citrolabs/ego-lite
第一次跑浏览器任务时它会引导你把 app 装上。
方式三，让Agent自己装。 把这段直接贴进 Claude Code，让它照 README 干。
Set up ego lite for me: https://github.com/citrolabs/ego-lite
Read `skills/ego-browser/references/install.md` and follow the steps to install ego lite.
装完验证一下。
command -v ego-browser
ego-browser nodejs -e "console.log('ego-browser ready')"
打出 ego-browser ready 就是通了。Windows 有个文档里专门提醒的坑，PATH 要开新终端才生效。
接 Claude Code 有两个细节。要用 Code 模式，Cowork 模式下 /ego-browser 不会出现；权限建议开 bypass permissions，桌面端在输入框旁边的权限选择器里切，终端用 claude --dangerously-skip-permissions 启动。理由很简单，Agent得能自己启动浏览器进程。
然后就可以派活了。
/ego-browser 打开 github.com/anthropics/claude-code 的 issues，把最近 10 条的标题、标签、回复数整理成表格给我
Agent自己开 Space、自己写上面那段代码、自己把结果整理回来。过程中你可以点开 Space 看它操作，碰上验证码直接上手接管，弄完告诉它继续。
或者输入以下指令，调用 ego-browser：
/ego-browser
打开我们后台用户管理页面，查询id=1001的用户，检查用户状态是否正常，截图保存页面，提取页面上的手机号、创建时间信息。遇到验证码停下来交给我处理。
执行之后，ego-lite 会自动新开一个独立 Space，在里面打开页面执行操作，不会碰你当前正在浏览的标签页。 执行完成之后，Agent 返回提取结果和截图；如果中途出现滑块验证码，任务自动暂停，通知你切到对应 Space 人工完成校验，之后继续执行。
你也可以用来做表单自动化、页面巡检、数据导出等 Web 任务。
从零到跑通，下 app、迁数据、一句话派活，三步。
Agent 在里面到底怎么干活
整套流程由一个 skill 串起来。装完 ego-lite，ego-browser 这个 skill 就进了 ~/.claude/skills，其他 Agent 的技能目录同样会写。你在 Claude Code 里敲 /ego-browser 加一句任务描述，Agent读 skill，先跑一遍 ego-browser skill 拿使用指南，然后开工。
底层长这样，一段 heredoc 交给 Node.js 跑，浏览器能力以函数形式预先注入，不用 import。
ego-browser nodejs <<'EOF'
const task = await useOrCreateTaskSpace('search github issues')
await openOrReuseTab('https://github.com/issues', { wait: true, timeout: 20 })
cliLog(await snapshotText())
EOF
useOrCreateTaskSpace 开任务空间，openOrReuseTab 开页面，snapshotText 把页面拍成结构化文本打出来。Agent看完快照，接着写下一步。
await click('@21', { label: 'check the login state' })
await fillInput('@2', 'user@test.com')
@21 这种编号只对最近一次快照有效，页面一变就得重拍，要跨步骤引用就用 loc= 选择器或者直接写 CSS。
什么时候选 ego-lite，什么时候用 Playwright？
✅ 选 ego-lite：
- 做 AI Agent 网页交互测评、Harness 任务验证
- 内部后台、企业 SaaS，登录复杂、有 2FA、风控校验
- 需要人工和 Agent 交替接管页面，边看边调
- 不想花大量精力维护登录脚本、账号池、复杂选择器
✅ 继续用 Playwright：
- 流水线 CI 大规模稳定 E2E 回归
- 无登录、公开页面、标准化业务流程
- 需要容器化、集群批量执行测试用例
写在最后
写到最后，很多做 AI 测试、Agent 测评的同学都卡在同一个卡点：我们的 Agent 有能力理解自然语言，但是很难稳定、低成本地和真实网页系统交互。
ego-lite 这个项目的巧妙之处，不是造了一套新的自动化语法，而是换了思路：不再让自动化进程变成一个孤立、干净、容易被风控拦截的无头实例，而是让 Agent 作为你的协作伙伴，复用你真实、可信的浏览器会话，在隔离的空间里完成测试、巡检、页面校验。
工具可用可不用，看团队手里的活，但有两个视角我觉得比工具本身值钱。
登录态本来就是测试资产。你每天真实使用的会话，比任何测试账号和导出来的 storageState 都新鲜可靠。哪怕不装 ego-lite，自己搭框架时也值得想想怎么让自动化直接吃真实登录态，登录维护省下来的工时都是净赚。
执行方式这个视角也通用。让Agent一次写一段程序跑完，跟敲一步看一步比，成本差好几倍。以后评估任何Agent工具，先看它走哪条路。
不合适不必硬上，但学习一下这两个新视角，比多装一个浏览器值。
但值得一提的是，它和我们前面分享的 Keploy、Promptfoo、DeepEval、Langfuse 这些工具一样，都是 AI 测试工程链路里一块很有价值的拼图。后续在我们的企业级一站式智能测试平台项目中，也会把 ego-lite 作为 Web Agent 测评的可选底座纳入进来，打通 Agent 任务调度、页面执行、结果断言、报告产出的完整链路。
项目地址放这，感兴趣的自己翻。
GitHub: github.com/citrolabs/ego-lite
官网: lite.ego.app
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
I turned to a project on GitHub the previous time: ego-lite, a browser for AI Agent, dedicated to solving the pains of Claude Code / Codex / Cursor, such as AI AI agent, to automate the web page, to share the same log-in session with Agent, not to rob each other of the label pages, and I myself spent some time, practical, shared it.
This project, just remember a few key messages:
- Project: ego-lite, citro Labs' AI Agent Browser, Agent runs Web Automation in an Isolated Work Zone, replicating the real login in your Chome State
- Technical warehouse: JavaScript, repository MIT protocol, complete with a desktop browser +ego-browser command line
- Heat: 18373 star, written by the sky
This year, the warehouse was built in April.
- Platform: MacOS (Apple Silicon and Intel), Windows have installation packages, Linux is still on the road map
- Commerce: Browser free of charge, no subscription; official comparison of 31 real website tasks with tools like Playwright-cli, claiming to be between 47% and 63%, which cannot be verified.
- Address: Github.com/citrolabs/ego-lite
It's easy to do, and it can be summed up in one sentence:
ego-lite is a Shromium browser shared by people and AI Agent, who works as a browser in its own separate Space, login state, Cookie, extension, which directly inherits you, but does not take tabs from each other, and opens boxes with a remote debugging protocol, without complex debugging.
Core competencies, one at a time.
1 Space Segregated Work Area (Full Core Design)
Internalally isolated Space space, each browser mission has an independent Space, your own browsing label in the main space, AI Agent running automating tasks in a unique Space; multiple Agents can run multiple web missions in parallel, without interfering with each other, without taking away the pages you are using, without cutting through your windows. You can access the space in Agent at any time to see progress, take over operations, handle authentication codes, sliders, etc.
Two, go straight to your login, Cookie.
Supports the import of all Chrome configurations with one key: account password, Cookie, login session, extension, bookmark. Agent carries out web-based operations by directly inheriting the state you already login, without automating every time you login, bypassing the authentication code, and processing SSO, which is also the greatest advantage in comparison to Playwright/Browser-Use.
3 Native Age Callability ego-browser skill
There is no need to write complex CDP, webdiver scripts, Agent sends missions in natural languages through/ego-browser commands (filling forms, scratching pages, clicking, intercepting); the bottom is the Chromium nuclear primary snapshot, not JS injections, which can be used to identify Shadow Dom, cross-dorame, third-party components, many elements of traditional automation tools that can be identified; and volume execution of page actions to reduce LLM tool call, save Token, increase speed.
4. Local retention of data
Browse Cookie, login information, page contents are stored on this machine without being uploaded to the cloud. You do not need a registered account number, you only have to configure the model API Key.
5. Use as a daily browser at the same time
The kernel is the Chromium, Chromium, Chrome Extension, Label Group, no trace mode, full bookmark compatibility, day-to-day Internet access and Agent automation can be done in the same browser without opening two sets of Chrome+ headless browsers.
Core difference between ego-lite and traditional crome
A lot of people can be suspicious after looking at it. It looks like an ego-lite browser. What difference does it make to the traditional Chrome browser? Don't worry, use a table box to show you.
Complementing the core difference of the sentence: Chrome is for human use, automation is for external capability; ego-lite is a browser that is designed from the bottom: people share with AI Agent.
Why do people need it?
Web has been automated for so many years, the script itself has long been easy, and the carcass are often the first step to log in. The company's internal system is clean, SSO, jumps to the 78th floor, plus texting codes, sweeps, wind control slides. Playwright's statusstate scheme saves login, service end-to-ken is disabled as soon as rotation takes place, and test account number is locked in the wind for three days. Functional validation for five minutes, login for two hours.
So a lot of internal systems have so far been validated by people. More than a dozen backstages are opened every morning to see service status, data, configuration, automation, and access is all blocked.
On the other side, there are now people who send this kind of work directly to AI Agent and put a playwright-mcp on Claude Code or chrome-devtools-mcp to drive the browser. The problem is three. It usually comes up with a completely new blank example, with no log-in status, or do you have to wait for log-in; tool call, context, re-transmits every round, token burns fast; or headless can't see the process, or Agent and you squeeze in the same window, and you have to stop as soon as it moves.
ego-lite just happens to have half of each of the two blocks. Inheritance for login, Agent works with a real session; execution is placed in Space, visible and manageable.
ego-lite VS Playwright / Browner-Use (traditional automation framework)
Playwright, Browner-Use is an automated SDK library that requires code call to pull up an independent headless browser process; ego-lite is the browser's base and carries a running time that Agent can use.
- Playwright: Every time a completely new clean browser context is activated, without login, it is easy to trigger wind control, authentication codes; relying on the CSS selection, the front-end style is changed and scripts are suspended; suitable for stabilizing the CI flow line, standardized E2E returns.
- ego-lite: Reuse real browser fingerprints and login, suitable for exploratory testing, back-office page validation, AI Agent web assessment, manual + Agent co-test; do not maintain a large number of stabilizers and encounter a complex log-in system at very low cost; do not replace Playwright, but complement it.
In summary, Chrome is a human browser and automation is an additional feature; Playwright is a script-driven headless browser; and ego-lite is a browser that allows AI Agent to share with you.
Installation and quick start
Way one, straight down app.
MacOS has Apple Silicon and Intel, two DMGs, Windows have exe, and can be found on the web home page (https://lite.ego.app/).
The first time you start asking if you are moving Crome data, dot agree, the ego-browser command is registered after you leave, usually in ~.local/bin.
THE WAY TWO, JUST SKILL.
npx skills add citrolabs/ego-lite
The first running browser mission will guide you to load the app.
The way three, get Agent to load himself. Put this in Claude Code and let it dry with README.
Set up ego lite for me: https://github.com/citrolabs/ego-lite
Read `skills/ego-browser/references/install.md` and follow the steps to install ego lite.
Fill it up.
command -v ego-browser
ego-browser nodejs -e "console.log('ego-browser ready')"
Hit ego-browser ready to get through. Windows has a dedicated alarm pit in a document, and PATH needs a new terminal.
Take Claude Code with two details. To use Code mode, /ego-browser does not appear; permission suggests opening bypass missions, desktop end in permission Selector next to the input box, starting with claude-dangerously-skip-missions. The simple reason is that Agent has to be able to start the browser process himself.
Then we can send them alive.
/ego-browser opens the github.com/anthropics/claude-code issues, grouping the titles, labels, responses of the most recent 10 to me in a table
Agent opened his own space, wrote the code on it, and put it back together. In the process, you can light the Space to see how it works, touch the code and take over, and tell it to go on.
Or enter the following command to call ego-browser:
/ego-browser
Open our back-office user management page, query id=1001 users, check if the user is in normal condition, screenshot saves the page, extracts phone numbers on the page, creates time information. I'll take care of it.
After execution, ego-lite automatically opens an independent Space, opens the page to perform the operation without touching the tab you are currently browsing. After the execution is completed, Agent returns the extracts and intercepts; if the slider authentication code appears, the task is automatically suspended and you are advised to cut to the corresponding Space manual validation and continue thereafter.
You can also use it to automate forms, page checks, data export, etc.
From zero to run, down app, move data, move a sentence, three steps.
Agent, what the hell are you doing in there?
The whole process is a skill string. After loading ego-lite, ego-browser, this skill goes into ~/.claude/skills, and other Agent skills catalogues are also written. You banged /ego-browser in Claude Code with a description of the mission, Agent read skill, run the ego-browser skill to get the guide and start the work.
The bottom length is this: a section of thisdoc is given to Node.js to run, and the browser's ability is pre-injected in a function, without effect.
ego-browser nodejs <<'EOF'
const task = await useOrCreateTaskSpace('search github issues')
await openOrReuseTab('https://github.com/issues', { wait: true, timeout: 20 })
cliLog(await snapshotText())
EOF
Use OrCreateTaskSpace taskspace, openOrReuseTab page, snapshotText to shoot the page into a structured text. After reading the snapshot, Agent continued with next step.
await click('@21', { label: 'check the login state' })
await fillInput('@2', 'user@test.com')
@21 This number is valid only for the last snapshot, and the page changes are reshooted, using loc = selector or writing CSS as a step-by-step reference.
When's the ego-lite, playwright?
✅Ego-lite:
- Do AI Agent Web Interactive Evaluation, Harness Job Validation
- Internal back office, business, SaaS, complex access, 2FA, Wind Control. Check
- We need manual and Agent to take over the page and watch the tone.
- I don't want to spend a lot of effort on login scripts, account pools, complex choices.
Playwright:
Waterline C.I. Mass stabilization E2E returns
- No login, open pages, standardized business processes
- Need for containerization, test examples for cluster batch execution
At the end.
And finally, many of the students who did the AI test and the Agent test were stuck at the same card point: our Agent is capable of understanding the language of nature, but it's difficult to interact with the real web system in a stable, low-cost way.
The trick of the ego-lite project is not to create a new automated syntax, but to change the thinking: instead of turning the automation process into an isolated, clean, wind-controlled no-head example, it is to have Agent, as your partner, return to your real, credible browser session to complete testing, inspection, page verification in isolated spaces.
Tools can be used to see what the team is doing, but I think two perspectives are worth more than the tool itself.
The login is a test asset. You're actually using a conversation every day that's more reliable than any test account number and sourced stockstate. Even without ego-lite, it is worth thinking about how to get automation to eat the real login, and the hours saved by login maintenance are net.
The implementation approach is also common. Having Agent run a piece of the program at a time is several times less than a knock. Assess any Agent tools at a later stage and see which way it goes.
It would not be appropriate to be hard, but to learn from these two new perspectives, which are more than a browser value.
But it's worth mentioning that it's a valuable puzzle in the AI test engineering chain, like the Keploy, Promptfoo, DeepEval, Langfuse tools that we shared. Following up on our enterprise-level all-in-one Smart Test Platform project, ego-lite will also be included as an optional base for WebAgent testing, which will provide access to the Agent task schedule, page execution, outcome assertion, and full reporting of output.
The project address is here, interested in flipping.
GitHub: github.com/citrolabs/ego-lite
Network: place.ego.app
Auto-fetched and translated by AI Atlas. All rights belong to the original author.