I want to build an agent that my sister can use.
我想做一个我姐姐都能用的 agent。
She's not a developer. She doesn't know what an API is. But she knows what she wants: "help me find good restaurants people are talking about on Xiaohongshu and Douban." That should just work.
她不是开发者，不知道 API 是什么。但她知道自己想要什么："帮我看看小红书和豆瓣上大家在聊哪些好吃的餐厅。"这件事应该直接就能跑。
She can't use Opus or GPT-5. Not just because of the cost — some models she simply has no way to access. But she can run an open-source model. If the agent only works with frontier models, it doesn't work for her. It doesn't work for most people.
她用不了 Opus 或 GPT-5，不只是因为贵，有些模型她根本就没办法用。但她可以跑开源模型。如果 agent 只能跟前沿模型配合，那对她来说就是不能用。对大部分人来说都是不能用。
AI Agent for everyone. Every model. Every language. Every website.
AI Agent 平权。任何模型，任何语言，任何网站。
Today's AI agents have a class divide:
今天的 AI Agent 存在一个阶级分化：
-
Frontier models (Opus, GPT-5) can write code to interact with websites — reverse-engineer APIs, handle auth, parse responses. They work, but not everyone can access them.
-
Open-source models (Qwen, GLM, Llama, DeepSeek) are smart enough to reason and plan, but they can't reliably write complex scraping code on the fly. So they're locked out of the real web.
-
前沿模型（Opus、GPT-5）能写代码跟网站交互，逆向 API、处理认证、解析响应。能用，但不是每个人都能用到。
-
开源模型（通义千问、智谱、Llama、DeepSeek）足够聪明，能推理能规划，但没法现场写出复杂的爬虫代码。所以它们被挡在了真实互联网之外。
This is not a model problem. It's an interface problem.
这不是模型的问题，是接口的问题。
bb-agent is built on bb-browser site adapters. 95 adapters have already done the hard work — reverse-engineering APIs, handling cookies, parsing responses, all packaged into one-line CLI commands that return structured JSON.
bb-agent 基于 bb-browser 的 site adapter 构建。95 个 adapter 已经做完了难的部分，逆向 API、处理 cookie、解析响应，全部封装成一行 CLI 命令，返回结构化 JSON。
# No code needed. Just CLI. Any model can do this.
# 不需要写代码，就是 CLI。任何模型都会。
bb-browser site twitter/search "AI agent" → JSON
bb-browser site twitter/following elonmusk → JSON
bb-browser site xiaohongshu/search "旅行攻略" → JSON
bb-browser site bilibili/search "深度学习" → JSON
bb-browser site youtube/search "machine learning" → JSON
The intelligence requirement drops from writing code to calling commands:
对模型的智力要求从写代码降到了调命令：
Before / 之前:
Agent needs Opus → (write code to reverse-engineer Twitter API) → maybe works
After / 之后:
Agent needs any model → bb-browser site twitter/search "query" → always works
Use frontier models to BUILD the adapters. Use any model to RUN them.
用前沿模型来造 adapter，用任何模型来跑任务。
95 adapters across 35+ platforms. Twitter, YouTube, Reddit, Bilibili, Xiaohongshu, Douban, Zhihu, Weibo, HackerNews, GitHub, and more.
95 个 adapter 覆盖 35+ 平台。推特、YouTube、Reddit、B站、小红书、豆瓣、知乎、微博、HackerNews、GitHub 等。
-
Cross-platform workflows — search Twitter, check Xiaohongshu, read Reddit, all in one task
-
Social graph analysis — follow chains, find key people, map relationships
-
Content monitoring — track keywords across platforms
-
Research automation — collect and synthesize from multiple sources
-
跨平台工作流 — 搜推特、查小红书、读 Reddit，一个任务搞定
-
社交图谱分析 — 追踪关注链，找关键人物，画关系图
-
内容监控 — 跨平台关键词追踪
-
调研自动化 — 多源信息收集与综合
bb-agent (any LLM: Qwen, GLM, DeepSeek, Llama, GPT, Claude...)
│
│ "search Twitter for AI agents, then check who the top authors follow"
│
├── bb-browser site twitter/search "AI agent"
├── bb-browser site twitter/user {top_author}
├── bb-browser site twitter/following {top_author}
│ ...just CLI commands, structured JSON in/out
│
▼
Results → LLM reasoning → next action → repeat
The browser is already logged in. bb-browser runs in your Chrome with your cookies, your sessions, your auth. No API keys needed. The adapter runs JS in the page context — it IS the page.
浏览器已经登录了。bb-browser 跑在你的 Chrome 里，用你的 cookie、你的会话、你的登录态。不需要 API key。adapter 在页面上下文里跑 JS，它就是页面本身。
Early stage. Building in public.
早期阶段，公开构建中。
- bb-browser — The browser automation CLI that powers everything
- bb-sites — Community-maintained site adapters (95 adapters, 35+ platforms)
MIT