BB-Browser

Let AI agents operate your logged-in browser directly

Browser or plug-in

https://github.com/epiral/bb-browser/blob/main/README.zh-CN.md

Visit Website

In-depth Report

bb-browser (BadBoy Browser) is an open source browser automation tool released by developer epiral in early 2026. The core concept is "your browser is the API". Different from traditional headless browser solutions such as Playwright and Selenium, bb-browser uses Chrome extensions and CDP protocols to allow the AI Agent to directly reuse the user's real browser login status to access the Internet without API Key, simulated login, or anti-crawling confrontation. Currently supporting 36 platforms and 103 commands, GitHub has received 4,600+ stars, and npm has been downloaded approximately 1,700 times a week. The Chinese developer community pays more attention to it, but the penetration rate in the English community is still low, and the overall situation is in the early and rapid iteration stage.
bb-browser was developed by independent developer yan5xu (GitHub username epiral), written in TypeScript (88.8%), and adopted the MIT open source license. The first commit of the project was on January 31, 2026, and the Chinese README was officially released on March 15 and began to be promoted. By mid-April, it had been iterated to version v0.11.3, with a total of 18 Release and 29 npm versions. The birth background of the project points directly to a core pain point in the field of AI Agent: the design goal of the Internet is to serve human browsers, 99% of websites do not provide machine interfaces, and the mainstream access method for AI Agents is still API, which means that the vast majority of websites without open APIs are almost inaccessible to Agents. The idea behind bb-browser is to flip this equation - instead of forcing websites to provide machine interfaces, they let machines use human interfaces directly. The project spread rapidly in the Chinese developer community. After the Chinese version of the README was released on March 15, WarpNav was first recommended on March 16 (845 stars at the time). Subsequently, Zhihu, Nuggets, He San Notes, Blue Ring Blog, Vibe Tools, What’s Worth Buying and other platforms successively published review articles. By the end of March, the stars had exceeded 3,300. On March 16, the project was also posted on Hacker News, receiving 7 points and a small number of positive comments.
The core capabilities of bb-browser are divided into two layers: high-level semantic interface and low-level atomic operations. The high-level semantic interface provides a one-line imperative platform call that returns structured JSON. Covering 36 platforms and 103 commands, covering search engines (Google, Baidu, Bing, DuckDuckGo, Sogou, WeChat), social media (Twitter/X, Reddit, Weibo, Xiaohongshu, Jiji, LinkedIn, Hupu), news information (BBC, Reuters, 36 Krypton, Toutiao, Oriental Fortune), technology development (GitHub, StackOverflow, HackerNews, CSDN, Blog Park, V2EX, Dev.to, npm, PyPI, arXiv), video platform (YouTube, Station B), audio and video entertainment (Douban, IMDb, Qidian Chinese Network), financial stocks (Snowball, Oriental Fortune, Yahoo Finance), job recruitment (BOSS) Direct recruitment, LinkedIn), knowledge encyclopedia (Wikipedia, Zhihu, Open Library) and other 11 categories. Station B alone has 9 adapters and is the single platform with the most commands. The underlying atomic operations provide 7 basic browser control commands: open (open the page), snapshot (get the accessibility tree), click/fill (simulate click input), eval (execute JavaScript), fetch (network request with login status), network (packet capture analysis), screenshot (page screenshot). All commands support --json formatted output, --jq inline data filtering, --tab multi-tab concurrent operations.The technical architecture adopts a four-layer design: AI Agent (Claude Code, Codex, Cursor, etc.) connects to bb-browser CLI through CLI or MCP protocol. CLI communicates with local Daemon through HTTP (default listening localhost:19824). Daemon connects to Chrome extension through CDP WebSocket. Chrome extension controls the real browser through chrome.debugger protocol. Site Adapter is essentially a JS file that is directly executed through eval() in a browser tab. It calls the page's own fetch() API to carry real cookies. It can even directly call the page's existing Webpack module or Pinia Store to complete complex signatures. This is also the most essential difference between bb-browser and traditional solutions - no matter how disguised a headless browser is, fingerprints, behavior patterns, and WebSocket handshakes will always reveal its flaws. However, bb-browser does not have this problem at all, because it is the user's real browser. Adapters are divided into three levels according to complexity: Tier 1 only requires Cookie direct fetch, which can be completed in about 1 minute like Reddit, GitHub, and V2EX; Tier 2 requires Bearer Token plus CSRF token, like Twitter and Zhihu, which takes about 3 minutes; Tier 3 involves Webpack injection or Pinia Store operations, like Twitter search, and Xiaohongshu, which takes about 10 minutes. Officials have tested 20 AI Agents running concurrently, each independently reversing different websites and producing available adapters. The marginal cost of making new websites accessible to the Internet by Agents is approaching zero. The product provides three usage methods: CLI direct command line use, OpenClaw extension-free mode (directly use the OpenClaw built-in browser through the --openclaw parameter, without installing extensions and starting daemon), MCP Server mode (adding npx -y bb-browser --mcp to the MCP configuration file to access AI coding tools such as Claude Code and Cursor).Compared with competing products, the core differentiation of bb-browser lies in the two dimensions of Token efficiency and anti-detection capabilities. Traditional CDP solutions (such as Browser Use, Playwright MCP) may consume 50K+ tokens for a single DOM snapshot, while bb-browser directly returns structured JSON through the site adapter, which only requires a few hundred tokens. Hacker News user genedai verified this after a few days of actual use: "Way less token waste compared to the usual CDP approach." In the reviews of 10 browser automation tools, bb-browser was classified as "an interesting route,
bb-browser is a completely open source and free project, using the MIT license without any paid functions or commercial restrictions. The npm package is downloaded about 1,700 times a week, and there is no SaaS service, no cloud-hosted version, and no paid plans. Developers maintain the project through the GitHub repository, and the community contributes platform adapters through the bb-sites repository. This pure open source model means that the project has no direct path to monetization. In the long term, possible commercialization directions include enterprise services, hosted cloud browser solutions or a professional adapter market, but these are currently not on the project roadmap.
The Chinese developer community's overall evaluation of bb-browser is positive and cautious. Technology blogs generally appreciate the innovation and practicality of its ideas. The author of He San's notes commented, "Without bb-browser, the world of AI Agent is files + terminals + several APIs that require keys; with it, the world of AI Agents is files + terminals + the entire Internet." Users in the Linux.do community summarized it as "It's really unscrupulous, but it's really useful." The author of the Nuggets article gave five stars for functional completeness, five stars for ease of use, and five stars for scalability, but only three stars for security. Specific positive comments focus on several dimensions. Token efficiency is extremely high. The traditional solution requires 50K+ tokens for a single operation, while bb-browser only requires a few hundred tokens. This is an advantage that has been repeatedly verified by actual users. It works out of the box with zero configuration. You can start using it with just npm install -g bb-browser without applying for any API Key. The anti-crawling capability fundamentally solves the problem. Using a real browser and a real login state, the website cannot distinguish whether it is a human or an AI agent. The adapter mechanism is flexible, with three complexity levels covering adaptation scenarios from 1 minute to 10 minutes. It can also realize automated batch adaptation with AI Agent. Negative feedback mainly focuses on two directions: safety and maturity. Privacy and security concerns are the most discussed issues. bb-browser has gained near-complete control over the browser through the CDP protocol. Community-contributed adapters may contain malicious code without official auditing. There are XSS or local code execution risks when eval is executed within a browser tab. High-frequency automated operations may still trigger platform risk controls, although they are technically undetectable. The quality of adapters varies, and some platform adaptations may become invalid due to website revisions. The overall project is still in its early stages. There is almost no interaction between "What is Worth Buying" and the article comment area of Sohu IT Channel, indicating that awareness among non-technical communities is still very low.
In the AI browser automation track, bb-browser has taken a unique technical route. Unbrowse's 2026 browser automation landscape analysis divides the market into five major categories. bb-browser's "site adapter + real login state" route forms differentiated competition with Browser Use's "LLM autonomous navigation", Playwright's "programmed control", and Steel Browser's "cloud infrastructure". The AI browser market is expected to grow from $4.5 billion in 2024 to $76.8 billion in 2034, and is at an inflection point in 2026. Industry analysis points out three trends: cloud infrastructure is converging (the differences between Browserbase, Steel, and Browserless are getting smaller and smaller), the SDK layer is being commoditized (natural language-controlled browsers are becoming "features" rather than "products"), and browsers are becoming underlying implementation details (API-native solutions that bypass browsers have overwhelming advantages in cost and speed). bb-browser's route - executing the adapter directly within the browser and returning reduced JSON - is in line with industry trends in terms of cost and efficiency, but its dependence on the local real browser also means it is not suitable for large-scale cloud deployment scenarios. Sumike's analysis article provides the most in-depth review of security compliance, pointing out that the core mechanism of this tool is essentially to bypass the website's security protection measures. When using it, you must be clearly aware of the high compliance risks (high probability of violating the terms of service of the target platform and may violate relevant laws), high account risks (may lead to the account being permanently banned), and the security responsibility lies with the user (the tool itself does not provide compliance protection). This article is the Chinese community’s most serious independent look at the project.
Security and privacy are the core controversial issues of bb-browser. The project's PRIVACY.md clearly states: all communication takes place locally on localhost, data only exists in memory, no browsing history is logged, no cookies/credentials are extracted or stored, no telemetry, no analytics, no cloud services. But these statements do not completely eliminate the risk. Specific risks include several levels. In terms of complete browser control, almost complete browser control is obtained through the CDP protocol, and malicious adapters can theoretically steal any page data. In terms of adapter code security, community-contributed adapters have not been officially audited, and there is a serious risk of supply chain attacks when eval executes arbitrary JavaScript in browser tabs. In terms of login state exposure, all user login states are exposed to the AI Agent. Once the Agent is hijacked or behaves abnormally, the consequences will be serious. In terms of platform compliance risks, accessing the platform with an automated identity will almost certainly violate the terms of service of each platform and may result in account ban. In terms of legal risks, China's "Cybersecurity Law" and "Data Security Law" have strict regulations on unauthorized access, and in the United States, the CFAA may apply. In terms of moral impact, "undetectability" is a double-edged sword. If "one-sentence reversal" is used maliciously, it can achieve brushing, trolling, and data reselling. The concurrent feature of 20 Agents makes large-scale automated abuse possible. It is recommended that users install an independent Chrome browser instance specifically for bb-browser use, keep the main browser clean, and pay attention to controlling the frequency and scope when using it for personal study and research.
bb-browser is most suitable for the following types of users: developers who use AI coding tools such as Claude Code and Cursor, who need the Agent to have real-time network access capabilities; researchers who conduct cross-platform information aggregation and search for the same topic on multiple platforms at the same time with one command; operators who need to automatically monitor the dynamics of multiple platforms, such as grabbing hot lists and news on a daily basis; and technology enthusiasts who want to lower the threshold for AI Agents to access the Internet. Unsuitable scenarios include: production environments that require large-scale commercial data collection (high compliance risk), scenarios with extremely high privacy and security requirements (the tool obtains full control of the browser), cloud scenarios that require interfaceless server deployment (relying on local real browsers), and enterprise production environments that require 100% compliance (may violate the platform ToS). In terms of alternatives, if you only need browser automation capabilities without login state reuse, agent-browser (Vercel Labs) is a more mature and lightweight choice; if you are deeply dependent on the MCP ecosystem, Playwright MCP is the safest default option; if you need cloud browser infrastructure, Steel Browser provides the most generous free quota.
bb-browser uses an ingenious way to solve the core pain point of AI Agent's "unable to access the Internet" - instead of adapting the website to the machine, it allows the machine to use a human browser. Its site adapter route far exceeds traditional solutions in terms of token efficiency and anti-detection capabilities. 4600+ Stars and rapidly growing development iterations prove the community’s recognition of its direction. However, as an early project with v0.11 version, its security model, adapter stability and legal compliance still need time to be verified. For individual developers and technology enthusiasts, this is a tool worth trying; but when used in a production environment, it is necessary to fully assess the risks and implement security isolation.

User Reviews

Sharon_TaylorII12

—

跟browser-use对比了一下，bb-browser在Token消耗上确实有压倒性优势。browser-use做一个10步的工作流要4美元的LLM费用，bb-browser几百tokens就搞定了。不过browser-use是通用型的，能操作任意网页，bb-browser只能操作有适配器的平台。

SAbro

—

bb-browser的PRIVACY.md写得还挺详细的，说所有通信都在本地、不存cookie、不上传数据。不过代码审计这种事一般人谁会去做呢，还是得信开源社区的监督。

Philip_RogersX

—

问一下这个支持Firefox吗？看文档好像只支持Chrome

Raymond_Simmons520

—

这东西是不是有安全问题啊，把浏览器完全控制权交给一个开源工具，总感觉不太放心

dONALDrEED

—

试了一下bb-browser的MCP接入，直接加到Claude Code里就能用了，一条命令查知乎热榜，真的挺方便的。之前用browser-use查一个知乎问题要花好几万tokens，现在几百个就够了，这个token效率差距太明显了。

Gerald.Diaz_20228

—

最近在做RAG相关的调研，用bb-browser一条命令同时在arxiv、GitHub、StackOverflow、知乎、Twitter、36氪六个平台搜同一话题，返回的结构化JSON直接丢给LLM分析。这个工作流如果人工做至少要半天，现在几分钟搞定。不过有一点要注意，不同平台的适配器质量差异挺大的，GitHub和Reddit的比较稳定，小红书和微博的偶尔会出问题。

DPerez007

—

npm install -g bb-browser 然后配置MCP就完事了上手成本几乎为零

ShirleyMorales

—

yyds，从此Claude Code可以帮我刷知乎了

Anna.Ward_2023

—

只支持Chrome 原理是CDP协议其他浏览器没这个

JoeDiaz_20240

—

作为一个用了一个月的人说几句实话：好处是token效率高、接入简单、反爬无敌；坏处是有些适配器不稳定会突然失效、项目还很早期文档不完善、安全性需要自己把关。整体评价：值得折腾但不建议当生产主力。

JordanEvans_20211

—

当浏览器变成API的时候，反爬这个行业是不是要失业了

DRamirezII0

—

请问这个能在服务器上跑吗？我没有桌面环境

Melissa.JonesII

—

bb-browser跟Playwright MCP该怎么选？我的理解是如果你需要登录态复用就选bb-browser，如果只是通用浏览器控制就选Playwright MCP。两个不冲突可以同时装。

Ashleybarker

—

不行它依赖真实Chrome浏览器服务器上没桌面环境用不了除非你用Xvfb或者noVNC之类的方案

贾萱

—

4600星的开源项目，免费+MIT协议，这种开发者真的该respect

WAnderson_88

—

今天刷搜狐看到bb-browser的文章，评论区0互动说明大众用户根本不知道这东西啊哈哈，但作为Claude Code用户，这玩意真的打开了新世界的大门

Michelle.Jimenez_Pro

—

有人拿anybrowse和bb-browser做对比，我觉得这两个定位不太一样。anybrowse更偏通用浏览器自动化，bb-browser的核心是站点适配器+登录态复用。看你的需求选。

Timothy_Ortiz168

—

v0.11.3更新了，适配器比之前稳定了不少，之前bilibili的有些命令会报错现在好多了

NancyGonzalez_7722

—

新版本修复了不少bug 但Twitter的适配器还是偶尔抽风

LauraWrightQ

—

说真的這東西的思路太簡單粗暴了直接在瀏覽器tab裡面跑js 用你登入的cookie發fetch 這跟傳統爬蟲完全是兩個路線

KVasquez780

—

台灣有篇介紹文把它翻成繁體了，看起來順眼多了哈哈哈

APrice007

—

装了独立Chrome专门跑bb-browser，登录了常用网站，现在Claude Code让我查什么就查什么，跨平台研究效率拉满了

Andrew.Murray_7

—

Unbrowse那篇2026浏览器自动化格局分析提到了行业三大趋势，bb-browser的路线跟「浏览器变成底层实现细节」这个趋势是吻合的。不过它依赖本地真实浏览器这个特点，在云端部署场景下确实是个硬伤。

天涯138

—

笑死真的很丧良心但真的很好用

JessicaWright520

—

知乎有人把bb-browser列为免费开源软件推荐，跟一堆知名工具放在一起。说实话这个项目确实有点东西，至少思路是全新的。

NFTCollector357

—

有人担心20个Agent并发适配网站这个特性被滥用，我觉得这个担心不是没有道理。降低逆向门槛是好事也是坏事，就看用在什么地方了。

Marilyn_Williams_66490

—

Tier1的适配器1分钟就能写完这个太夸张了，配合AI Agent读一遍guide自己就能逆向，边际成本趋近于零不是吹的

飞鸟246

—

苏米客那篇文章分析得太深入了，法律风险、技术风险、道德风险三个层面都讲到了。特别是提到CFAA可能适用这一点，很多开发者可能根本没想过这个问题。

JRichardson

—

3.3k星了，一个月前才800多，涨得真快。不过话说回来，项目还这么早期，适配器质量参差不齐，有的能用有的直接报错。

莲花776

—

没有bb-browser的话，AI Agent的世界是「文件系统+终端+几个有API key的服务」；有了它，Agent的世界是「文件+终端+整个互联网」。这个比喻太到位了。

BB-Browser

In-depth Report

User Reviews

Similar Products