你有没有想过
如果把当前免费的网页版 AI 打包成API KEY,咱是不是就Token自由了?
*UP 曾经就是这么干的,采用RPA白嫖网页版AI(豆包、KIMI)
Dulus 项目,大约 12K 行代码,打包成一个 2.5 MB 的 wheel
可以接 OpenAI、Groq、Bedrock、Ollama 本地模型
🐂🍺的是甚至直接劫持 Gemini 的免登访客会话、Claude.ai 的网页版、Kimi、通义千问、DeepSeek。
项目创建不到两周。
核心武器叫 WebBridge。
WebBridge:把网页聊天变成 API
原理很简单。
Dulus 用 Playwright 在后台打开浏览器,截获你输入的内容和模型返回的回答,把这些前后文重新注入到终端里的 CLIagent。
你打开浏览器,模型以为你在跟网页聊天。实际上 WebBridge 已经截获了完整的对话流,把模型输出解析成 tool call、代码片段、文件操作指令,全部传递给终端里的 agent 去执行。
这意味着你可以免费使用任何网页版 AI 模型,同时获得完整的 agent 能力。Gemini 的访客模式甚至不需要登录,打开网页就能开始用。
这不就是无敌了吗?还没完呢~
圆桌会议:多模型同时辩论
Dulus 有一个叫 Mesa Redonda(圆桌)的功能。
你提问题,它可以同时发送给多个模型。Claude 说一套,Gemini 说一套,通义千问再说一套。Dulus 把三个回答汇总、对比、融合,给出一个综合结论。
就像一场真正的圆桌讨论,而不是单向问答。
这个功能在 v0.2.89 版本就已经实现,目前支持通过 LiteLLM 网关接入 100+ 个后端的调度。
工具生态:30+ 内置工具
它能做的事远不止聊天。
内置 30+ 个工具,包括文件读写、Bash 命令执行、仓库 grep 搜索、网页浏览、截图 OCR、语音输入输出、Telegram 桥接、MCP 协议支持。还有一个 MemPalace 语义记忆系统,可以记住你之前的对话和项目上下文。
更关键的是,任何 Python 仓库都可以变成它的插件。
你想让它查股票价格?安装 yfinance,reload 插件,直接问 NVDA 或 S&P 500 的实时数据。它把这个能力叫做 adapt any python repository。
安装使用
# 项目网址:https://kevrojo.github.io/Dulus/# Githubhttps://github.com/KevRojo/Dulus
# Linux / macOS / WSL / Termux:curl -fsSL https://raw.githubusercontent.com/KevRojo/Dulus/main/install.sh | bash# Windows (PowerShell):irm https://raw.githubusercontent.com/KevRojo/Dulus/main/install.ps1 | iex
作者:多米尼加的独行侠
KevRojo 是唯一的贡献者。
他在 README 里用很大一段文字说明了一个问题:$DULUS 代币。
这不是社区先炒起来他再跟进的。社区有人先发了代币,但他看到早期投资者可能被割,就自己发了一个官方的,把创建者奖励锁在链上。他说自己不会卖。
代币的路线图也很清楚:开源版永远免费,代币只用于商业层——云端实例、多用户工作区、模型 credits、Pro 订阅。持有量达到阈值自动享受折扣。顶层持有者可以投票决定功能优先级。
这既是一种保护,也是一种矛盾。
一方面他坚持开源免费。另一方面他在同一项目里推销自己的代币。读者会问:这到底是公益还是营销?
它的极限在哪里
先说 ToS 红线。
WebBridge 的核心逻辑是劫持网页版 AI 服务的会话流。Gemini 和 Claude.ai 的服务条款明确禁止自动化抓取和代理转发。Anthropic 已经针对 Claude Code 的"额外使用"场景推出了付费墙。Dulus 的玩法随时可能触发账号封禁。
这已经不是工具问题了。这是一场游击战。
再说性能瓶颈。
WebBridge 依赖 Playwright 的浏览器实例,每个请求都要经历页面渲染、DOM 解析、文本提取。延迟比原生 API 高出数倍。复杂工具链调用时,一次完整的 agent 循环可能需要几十秒。
对比 Claude Code 原生的 token 级流式响应,WebBridge 的方案更像在模拟层里跑——能工作,但不够丝滑。
WebBridge 不会消失。即使用户被封、服务升级反爬,总有人会继续写下一个代理、下一个劫持、下一个免费的桥。
代码不会妥协。
这么好的东西,不点个赞吗~
信息来源
1. https://github.com/KevRojo/Dulus — 项目仓库
2. https://pypi.org/project/dulus/ — PyPI 包
3. https://x.com/KevRojox — 作者 X
Have you ever wondered
If we package the current free web version of AI into an API KEY, won't we be free from token limits?
*UP used to do this, using RPA to white-label the web version of AI (Doubao, KIMI)
The Dulus project, with about 12K lines of code, is packaged into a 2.5 MB wheel.
It can connect to OpenAI, Groq, Bedrock, and Ollama local models.
What's even more impressive is that it directly hijacks Gemini's guest sessions, Claude.ai's web version, Kimi, Tongyi Qianwen, and DeepSeek.
The project was created less than two weeks ago.
The core weapon is called WebBridge.
WebBridge: Turning web chats into APIs
The principle is very simple.
Dulus uses Playwright to open a browser in the background, intercepting the content you input and the response returned by the model. It then re-injects this context into the CLIagent in the terminal.
When you open the browser, the model thinks you're chatting on the web. In reality, WebBridge has already intercepted the entire conversation flow, parsing the model output into tool calls, code snippets, and file operation instructions, and passing them all to the agent in the terminal to execute.
This means you can use any web-based AI model for free while gaining full agent capabilities. Gemini's visitor mode doesn't even require logging in; you can start using it by simply opening the webpage.
Isn't this invincible? Not yet!
Roundtable: Multi-model debate
Dulus has a feature called Mesa Redonda (Roundtable).
When you ask a question, it can send it to multiple models simultaneously. Claude says one thing, Gemini says another, and Tongyi Qianwen says yet another. Dulus aggregates, compares, and fuses the three responses to give a comprehensive conclusion.
It's like a real roundtable discussion rather than a one-way Q&A.
This feature was already implemented in version 0.2.89 and currently supports调度 through the LiteLLM gateway to over 100 different backends.
Tool ecosystem: 30+ built-in tools
It can do much more than just chat.
There are over 30 built-in tools, including file read/write, Bash command execution, repository grep search, web browsing, screenshot OCR, voice input/output, Telegram bridging, and MCP protocol support. There's also a MemPalace semantic memory system that remembers your previous conversations and project context.
More importantly, any Python repository can become a plugin.
Want it to check stock prices? Install yfinance, reload the plugin, and ask directly for NVDA or S&P 500 real-time data. This capability is called adapt any python repository.
Installation and usage
Author: The Lone Ranger from the Dominican Republic
KevRojo is the sole contributor.
In the README, he dedicates a large section to explaining the $DULUS token.
The token wasn't created because the community hyped it up; rather, he saw that early investors might get scammed, so he issued an official token himself, locking the creator's rewards on the chain. He says he won't sell.
The roadmap for the token is clear: the open-source version remains free, and the token is only used for the commercial layer — cloud instances, multi-user workspaces, model credits, and Pro subscriptions. Holders reaching specific thresholds automatically receive discounts, and top holders can vote on feature priorities.
This is both a protection and a contradiction.
On one hand, he insists on free open-source. On the other, he promotes his own token within the same project. Readers may wonder: Is this philanthropy or marketing?
What are its limits?
First, regarding ToS red lines.
The core logic of WebBridge is hijacking the session flow of web-based AI services. The terms of service for Gemini and Claude.ai explicitly prohibit automated scraping and proxy forwarding. Anthropic has already put a paywall in place for "additional use" scenarios like Claude Code. Dulus' approach could trigger account bans at any time.
This isn't just a tool issue; it's a guerrilla war.
Second, performance bottlenecks.
WebBridge relies on Playwright's browser instances, meaning every request must go through page rendering, DOM parsing, and text extraction. The latency is several times higher than native APIs. For complex toolchain calls, a single complete agent loop can take tens of seconds.
Compared to Claude Code's native token-level streaming responses, WebBridge's solution feels like running in a simulation layer — it works, but it's not smooth.
WebBridge won't disappear. Even if users get banned or services upgrade their anti-scrapping measures, someone will always write the next proxy, the next hijack, the next free bridge.
Code never compromises.
Such a great thing, why not like it?
Information sources
1. https://github.com/KevRojo/Dulus — Project repository
2. https://pypi.org/project/dulus/ — PyPI package
3. https://x.com/KevRojox — Author's X account

