前言
现在谈论 AI Agent 的能力边界时,大多数人都会关注模型自身的推理能力以及上下文的长度。
但有一个经常被忽略的点:Agent 在真实的网络环境中的执行能力。
目前我们所使用的 Agent 工具,在浏览网页时仍然存在一些门槛,比如说当面对需要登录、验证码或动态渲染页面时,往往都无力解决。
所以,今天给大家介绍一款专门针对浏览器自动化的 CLI 工具。在 GitHub 上非常火,超 2.8k Star 
BrowserAct
大部分 Agent 是利用模型的视觉能力,而 BrowserAct 则是直接通过命令行来执行。
从环境伪装、执行策略和人工协作这三个层面,来解决 Agent 执行浏览器相关任务时的成功率问题。
三层递进的执行策略
我们先来看看 BrowserAct 的设计思路,前文提到,它整体分为环境层、执行层和人工层三个层次。当执行一个需要操作浏览器的任务时,会依次经过这三层。
一、环境层
这一层相当于是给 Agent 换了身衣服,让它看起来更像真人。
BrowserAct 通过浏览器指纹伪装和动态网络身份,让 Agent 的浏览环境看起来和真人用户几乎是一致的。
另外每个账号还可以绑定独立的登录空间和网络路径,就算多个账号同时跑也不会互相干扰。
它能帮你做的是:
1、同一个账号长期登录不会因为"自动化特征"被平台封号。2、同时运营多个店铺、社媒账号或地区站点,每个账号有独立干净的环境,不会因为串号被关联。3、Cookie、登录状态长期保持稳定,不用反复扫码或重新登录。
二、执行层
这一层解决的是遇到验证码或者遇到反机器人检测,被阻隔在页面外时该怎么办的问题。
BrowserAct 可以自动处理这些常见的反机器人检测,降低被识别为自动化行为的概率。
它能帮你做的是:
1、登录需要人机验证的网站时,Agent 可以自动通过验证,不需要你手动点"我不是机器人"2、抓取需要登录态才能看到的数据(比如后台数据、会员内容),Agent 能自动跨越验证屏障3、遇到比较复杂的页面检测时,Agent 会自主尝试突破,而不是直接放弃任务
三、人工层这一层解决的是 AI 实在搞不定时,接下来该怎么办的问题。答案也很显然,人工介入。当然需要人工介入并不是 BrowserAct 的能力不行,而是有些情况只有人类可以做到。例如短信验证码、扫码验证或者涉及敏感操作需要二次授权等等。但 BrowserAct 不同的是,它会生成一个远程协作链接,在任何设备中打开这个链接完成验证后,Agent 会从刚才终端的地方继续往下进行,而不是从头再开始,陷入无限循环。它能帮你做的是:1、AI 遇到短信验证码时,不会终止整个任务,你把验证码输进去,它还能接着干2、处理企业后台需要 SSO 登录或敏感审批的场景,你帮它过一道,后面全自动3、复杂长流程任务里,AI 和人工交替配合,你不需要重新执行已经做完的部分
所以总结一下,这三层的设计可以理解为,BrowserAct 先将 Agent 行为伪装成真人,如果没有被识破,就进入网页获取数据,如果被识破了,就进入下一层。下一层的执行层来解决被识破后网站设立的“栏杆”,如果“栏杆”实在是拆不掉,就呼叫真人进场帮助解决。这三层设计,就能让 Agent 更稳定地完成浏览器操作。 额外功能设计 除了这三层设计,BrowserAct 还有一些额外的能力,这里简单总结一下它额外的功能:1、多任务并发:支持在同一个浏览器环境下同步执行多个任务,各个窗口共享 Cookie 及登录状态2、多账号隔离:每个账号都运行在独立的 Stealth 浏览器中,账号之间互不影响。3、三种浏览器模式:chrome 模式复用本地登录态,适合需要登录的场景;stealth 隐私模式可以每次会话使用新的指纹+代理,配合动态代理,适合需要大批量抓取数据的爬虫场景,可以做到自动轮换IP,没有残留。Stealth 固定身份模式可以在隐私模式基础上,提供稳定的指纹+IP,配合静态代理,适合需要多个账号独立、稳定运行的场景,可以确保账号环境不变,不被系统判定为机器人。
4、技能沉淀:这是一个独立的Skill,需要单独安装。它可以将跑通的流程沉淀为可复用的 Skill,后续再次跑相同流程时直接运行这个Skill就可以了。 使用场景 ① 电商数据监测618 过去,各位电商人肯定还在忙得不可开交。都在忙着收集数据,监测同行数据也是任务的一环。因为电商平台都有严格的反制机制,正常 Agent 的浏览器控制是无法进入获取数据的,我们尝试让 BrowserAct 来跑一下这个任务。
由于我的电脑没登录过,所以 BrowserAct 找不到登录态,触发了人机接力。
可以看到它人机接力的方式是给我们发送一条链接,这条链接是 BrowserAct 的云实例,相当于一个云端虚拟机,所以这条链接在任何设备中打开都可以。打开后会进入云实例的登录页面,这里我们直接扫码登录就可以了,登录完成后就可以直接把页面关掉。
此时 BrowserAct 就获取到了登录状态,继续开始接下里的任务。在 Agent 中可以看到 BrowserAct 的操作汇报,每一步都在告诉你具体做了什么事情。
如果搜索词输入错了,它还会自己检查并进行纠正。
然后会自己按照销量去进行商品排序。
最终成功获取了销量前10的溜溜梅价格。
整个过程除了因为没有登录态,需要进行人机接力之外,其余 反制机制都被 BrowserAct 自行解决了。如果这类任务你每天要重复好几次,那可以使用上文提到的技能沉淀能力,将这个流程沉淀为一个Skill,后续直接运行这个 Skill,任务不管跑多少次,都会是我们预期的效果,也节省了 Agent 重复探索所消耗的 Token 和时间。
使用也很简单,直接让 Agent 安装 BrowserAct Forge 这个技能:
在 Agent 输入:
安装这个技能,链接:https://github.com/browser-act/skills/tree/main/browser-act-skill-forge,并验证技能是否可用。
② 多店铺管理许多电商商家可能会同时开多家店铺,管理时需要频繁切换账号,很麻烦。使用BrowserAct可以创建多个隐私浏览器,账号相互隔离在自己的专属浏览器中,互相不会干扰。我这里直接问应该如何操作,它给了我两个方案:
我选择方案二,然后它问了我几个问题:
接着需要购买一个静态代理,这个代理的意思就是给你的浏览器分配一个不会变的IP地址,让你的账号处在一个稳定的环境中,这样才不会触发平台风控,降低封号风险。
如果是需要长期使用的商业用途,可以买一个用着,如果只是用一次的话,也可以不绑定静态代理,用本地自己已有的地址就可以。
接着它会把静态代理的购买链接发给我们,去购买就行。当然也不一定要购买BrowserAct的静态代理,在其他地方购买后也是一样能用的把ID告诉Agent就行。确认后,一个额外的隐私浏览器就创建好了,在使用Agent管理多个店铺账号时,也不用频繁切换登录了。③ 自媒体分发许多自媒体会将稿件分发到多个平台,提高自己的影响力,但是同一个稿件需要重复在多个平台发多次,有点麻烦。我们尝试用BrowserAct来跑一下这个任务。
我的浏览器已经登录过小红书和知乎,所以我让 BrowserAct 用 Chrome 模式来跑这个任务,可以自动获取我的登录态。
它也是成功进入了我的小红书主页,没有需要人机接力让我重复登录,也没有触发反制机制,直接开始发贴流程。
之后就是比较长的发帖流程,找发送按钮,编辑标题,编辑正文等等。
最后成功发了出来,但让我比较意外的是,它还做了一个封面图,然后内容也不是以文字形式发送的,而是做了一张图。
顺便还收获了一个收藏。
知乎也顺利发送完成。
傻瓜式安装
BrowserAct 的技能开源免费,基本大部分功能都能免费使用。
只有在需要使用代理时,例如上文创建绑定了静态代理的隐私浏览器,才会产生付费。
安装起来非常简单,打开 BrowserAct 官网,点击中间的白色按钮 👇
将复制的提示词发送给你所使用的Agent,就会自动安装。之后在执行浏览器操作的任务时,就会使用技能来完成了。
当然,也可以直接告诉你的 Agent:
在 Agent 输入:
安装browser-act这个技能,链接:https://github.com/browser-act/skills/tree/main/browser-act,安装完成后验证一下它是否可用。
写在最后
在这些具体的使用场景中,所遇到的问题并不是单纯的模型够强就能解决的。
我很庆幸有无数的开发者,做出了类似 BrowserAct 这种可以解决具体场景问题的技能,让 Agent 可以在实际工作中完成任务。
不过 BrowserAct 也不是万能的,验证码、短信确认、企业审批这类环节,依然需要人类介入,简而言之自动优先、人工兜底、无缝续接。
对于正在尝试将 Agent 引入生产环境的团队来说,这种务实的设计思路,可能比全自动化的噱头更有长期价值。
BrowserAct 官网:
https://www.browseract.ai/ML
Github:
https://github.com/browser-act/skills
Before
Now, when we talk about the capability boundaries of AI Agents, most people focus on the model's reasoning power and context length.
But one often overlooked point is: the Agent's execution capability in the real web environment.
The Agent tools we currently use still have some barriers when browsing web pages — for example, they often struggle when faced with login requirements, CAPTCHAs, or dynamically rendered pages.
That's why I'm introducing a CLI tool specifically for browser automation today — it's very popular on GitHub, with over 2.8k stars!
BrowserAct
Most Agents rely on visual perception, while BrowserAct directly executes via command line.
It tackles the success rate of browser-related tasks by addressing environment camouflage, execution strategy, and human collaboration.
A Three-Tiered Execution Strategy
Let's first look at BrowserAct's design philosophy. As mentioned earlier, it operates in three main layers: environment, execution, and human. When executing a task requiring browser manipulation, these layers are traversed in sequence.
I. Environment Layer
This layer effectively disguises the Agent as a real person.
BrowserAct uses browser fingerprint spoofing and dynamic network identities to make the Agent's browsing environment indistinguishable from that of a human user.
Additionally, each account can be bound to an independent login space and network path, ensuring that multiple accounts running simultaneously do not interfere with one another.
What it can achieve for you:
1. A single account staying logged in for a long time won't get banned by the platform for "automated behavior." 2. Run multiple stores, social media accounts, or regional sites simultaneously—each account has its own isolated, clean environment and won't be linked through cross-account association. 3. Cookies and login states remain stable long-term, eliminating the need for repeated QR code scans or re-logins.
2. Multiple stores, social media accounts, or regional sites can be operated simultaneously, with each account maintaining an independent, clean environment, preventing association through cross-account tracking.
3. Cookies and login states remain stable over the long term, eliminating the need for frequent re-scanning or re-logging.
II. Execution Layer
This layer addresses what to do when you encounter a CAPTCHA or an anti-bot check that blocks access to the page.
1. When a website requires human verification at login, the Agent can automatically pass the verification without you manually clicking "I'm not a robot." 2. It can automatically bypass authentication barriers to scrape data that requires login status (e.g., backend dashboards, member-only content). 3. When encountering complex page detection, the Agent will autonomously attempt to break through instead of simply abandoning the task. Section 3: Human Layer — This layer addresses what to do when AI absolutely cannot handle a situation. The answer is obvious: human intervention. Of course, needing human intervention doesn't mean BrowserAct's capabilities are lacking; rather, some scenarios are simply beyond what machines can do alone. For example: SMS verification codes, QR code scans, or sensitive operations requiring secondary authorization. But unlike other tools, BrowserAct generates a remote collaboration link—open it on any device, complete the verification, and the Agent will resume from exactly where it left off instead of restarting from scratch and falling into an infinite loop. What it can help you with: 1. When the AI encounters an SMS verification code, it won't terminate the entire task—you enter the code and it continues. 2. For enterprise backend scenarios requiring SSO login or sensitive approvals, you handle one step and everything after runs fully automated. 3. In complex, long workflow tasks, AI and humans alternate cooperation—you don't need to re-execute parts that have already been completed. So to summarize, this three-layer design works as follows: BrowserAct first disguises the Agent's behavior as human. If not detected, it proceeds to the webpage to collect data. If detected, it moves to the next layer. The execution layer handles the "barriers" set up by websites after detection. If those barriers truly can't be broken through, it calls in a human to assist. This three-layer design enables the Agent to perform browser operations more stably. Additional Feature Design — Beyond these three layers, BrowserAct has several extra capabilities, briefly summarized here: 1. Multi-task concurrency: Supports executing multiple tasks simultaneously within the same browser environment, with all windows sharing cookies and login states. 2. Multi-account isolation: Each account runs in an independent Stealth browser, completely isolated from one another. 3. Three browser modes: Chrome mode reuses your local login states, ideal for scenarios requiring existing logins; Stealth privacy mode generates a new fingerprint + proxy for each session, paired with rotating proxies, ideal for large-scale scraping where IP auto-rotation is needed with no residual traces; Stealth fixed-identity mode provides a stable fingerprint + IP on top of privacy mode, paired with static proxies, ideal for running multiple accounts independently and stably, ensuring account environments remain unchanged and undetected as bots by the system. 4. Skill consolidation: This is an independent Skill that needs to be installed separately. It can capture a successfully completed workflow into a reusable Skill, so subsequent runs of the same process can just execute the Skill directly. Use Cases Scenario ①: E-commerce Data Monitoring — 618 is over, and every e-commerce operator was likely swamped with busywork. Collecting data and monitoring competitors were both key tasks. Since e-commerce platforms have strict anti-bot mechanisms, normal Agent browser control can't access the data. We tried letting BrowserAct handle this task. Since my computer wasn't logged in, BrowserAct couldn't find a login state and triggered the human-machine handoff. You can see the handoff method sends us a link—this link is BrowserAct's cloud instance, essentially a cloud virtual machine, so it can be opened on any device. Opening it brings up the cloud instance's login page, where we simply scan to log in. After logging in, we can close the page directly. At this point BrowserAct has obtained the login state and continues with the rest of the task. In the Agent, you can see BrowserAct's operation report—every step tells you exactly what it did. If the search term was entered incorrectly, it even self-corrected. Then it automatically sorted products by sales volume. Finally, it successfully retrieved the top 10 selling Liu Liu Mei prices. The entire process, except for the human-machine handoff due to missing login state, had all anti-bot mechanisms resolved autonomously by BrowserAct. If you need to repeat this type of task multiple times a day, you can use the skill consolidation feature mentioned above to save this workflow as a Skill, then simply run the Skill afterward—regardless of how many times it runs, it will produce the expected results, saving the tokens and time the Agent would otherwise consume exploring repeatedly. Using it is straightforward—just ask the Agent to install the BrowserAct Forge skill:










What it can achieve for you:
1. When logging into websites requiring human verification, the Agent can automatically pass the test without you needing to manually click "I am not a robot."
Scenario ②: Multi-store Management — Many e-commerce sellers operate multiple stores simultaneously, making account switching frequent and tedious. With BrowserAct, you can create multiple privacy browsers, each account isolated in its own dedicated browser without interfering with one another. I simply asked how to proceed, and it gave me two options: I chose Option 2, then it asked me a few questions: Next, a static proxy needs to be purchased. A static proxy assigns your browser an unchanging IP address, keeping your account in a stable environment so as not to trigger platform risk controls and reduce the chance of account bans. If it's for long-term commercial use, you can purchase one. If it's just for a one-time use, you can skip binding a static proxy and use your local IP instead. It then sends us the static proxy purchase link. Of course, you don't have to buy BrowserAct's static proxy—you can purchase one elsewhere and just provide the ID to the Agent. After confirmation, an additional privacy browser is created. When using the Agent to manage multiple store accounts, there's no need for frequent login switching. Scenario ③: Social Media Distribution — Many content creators distribute articles across multiple platforms to expand their reach, but posting the same article repeatedly on each platform can be cumbersome. We tried using BrowserAct for this task. My browser was already logged into Xiaohongshu and Zhihu, so I had BrowserAct use Chrome mode to automatically pick up my existing login states. It successfully entered my Xiaohongshu homepage without requiring any human-machine handoff for re-login, and without triggering any anti-bot mechanisms—it went straight into the posting workflow. Then came the relatively lengthy posting process: finding the send button, editing the title, editing the body text, etc. Finally, it successfully published the post. What surprised me was that it even created a cover image, and the content wasn't posted as plain text—it was turned into an image. It also accidentally earned a bookmark. Zhihu was also posted successfully.










3. In the case of complex page detections, the Agent will attempt to break through independently rather than abandoning the task.
*III. Human Layer*
This layer deals with what happens if the AI simply cannot proceed on its own. The answer is clear: human intervention. However, requiring human intervention does not indicate a failure of BrowserAct's capabilities; rather, it reflects that certain situations — such as SMS verification codes, QR code logins, or sensitive operations requiring secondary authorization — necessitate human involvement.
What it can achieve for you:
2. For enterprise backends requiring SSO logins or sensitive approvals, you can handle the authentication once, after which the entire subsequent process becomes automated.
3. In complex, long-duration workflows, AI and humans can collaborate iteratively, saving you the time of repeating already-completed steps.
So to summarize: this three-layered design mimics the behavior of a human operator first; if detected, it moves to the execution layer to resolve the barriers; and if those barriers are unbreakable, it calls for human assistance. This structure significantly enhances the stability of browser operations for Agents.
Additional Features
Beyond these three layers, BrowserAct offers additional capabilities:
1. Multi-task concurrency: supports simultaneously executing multiple tasks within the same browser environment, sharing cookies and login states across windows.
2. Multi-account isolation: each account runs in its own independent stealth browser, ensuring no interference between them.
3. Three browser modes:
- Chrome mode: reuses local login states, suitable for scenarios requiring prior login.
- Stealth privacy mode: utilizes new fingerprints + proxies per session, ideal for large-scale scraping without leaving residuals.
- Stealth fixed identity mode: provides stable fingerprints + IPs for long-term operation of multiple independent accounts.
4. Skill repository: a feature (requires separate installation) that allows for the codification of successful workflows into reusable skills, streamlining future task execution.
Use Cases




