50.7k Stars!Agent-Reach 开源,让 AI Agent 免费获取全网信息50.7k Stars! Agent-Reach Is Open Source, Letting AI Agents Access Information Across the Web for Free


AI free token/AI免费积分

50.7k Stars!Agent-Reach 开源,让 AI Agent 免费获取全网信息

图片

AI Agent 很聪明,但一让它上网就瞎了?这个项目治好了


一、AI Agent 的尴尬:能写代码,却上不了网

过去一年,AI Agent 赛道热度持续攀升。

Claude Code、OpenClaw、Cursor、Windsurf……各路智能体层出不穷,写代码、做方案,看起来越来越全能。

但真正在日常工作中高频使用 Agent 的朋友,一定遇到过同一个扎心的问题:

让 Agent 去 Twitter 看看最新舆论?不行。 让它去小红书搜一下真实用户分享?不行。 让它去 B 站提取视频字幕来分析?还是不行。

不是 Agent 不够聪明,而是「联网」这件事,门槛远比想象中高。

每一个平台都有自己的访问限制机制、Cookie 验证流程、接口收费策略。想打通一个平台,少则折腾半天,多则踩坑三天。做方案的精力全都浪费在了调试环境上。

而今天要介绍的开源项目 Agent-Reach,就是专门来解决这个问题的。

图片

二、Agent-Reach 是什么?

一句话概括:它是一个给 AI Agent 一键装上「全网信息获取能力」的 CLI 工具。

项目地址:https://github.com/Panniantong/Agent-Reach

目前已经在 GitHub 上收获了 50.7k Stars,GitHubDaily、HelloGitHub、阮一峰周刊等知名技术媒体都推荐过这个项目,SitePoint 和 Substack 也做了专题报道,开发者社区的关注度可见一斑。

Agent-Reach 的定位非常清晰:它是一个脚手架工具,而非框架。

框架会包一层又一层,最后你连底层工具长什么样都不知道。脚手架只管帮你把环境搭好,实际干活的还是上游工具。装完之后,你的 Agent 直接调用上游工具——比如读 Twitter 就直接调 xreach,读 YouTube 就直接调 yt-dlp。中间没有包装层,性能损耗为零。

这种设计有个好处:上游工具更新了,你不用等 Agent-Reach 跟着更新,工具链的每一环都可以独立替换。

Agent-Reach 把以下麻烦事,全部封装成了开箱即用的命令行工具:

  • 各平台接口对接
  • 依赖安装与环境配置
  • 代理设置与登录配置
  • 数据清洗与格式化输出

不管你用的是 Claude Code、OpenClaw、Cursor 还是 Windsurf,只要跑得动命令行,就能无缝使用 Agent-Reach。

非常贴心的一点:完全开源免费,无需任何 API Key,采用 MIT 协议(最宽松的开源许可,可自由使用和修改)。


三、能打通哪些平台?

Agent-Reach 覆盖了日常使用频率较高的主流内容平台,分为两大类:

开箱即用(零配置)

平台
能力
任意网页
读取页面内容,自动清理 HTML
YouTube
提取字幕、搜索视频,底层调 yt-dlp
RSS
订阅并读取任意 RSS/Atom 源
微信公众号
搜索文章、全文阅读,输出 Markdown
微博
热搜榜、用户动态、评论
V2EX
热门帖子、节点浏览、详情+回复
全网搜索
语义搜索,无需 API Key

简单配置后解锁(一句话搞定)

平台
能力
GitHub
私有仓库、Issue/PR、Fork,调用 gh CLI
X(原 Twitter)
搜索推文、浏览时间线、发推
B站
本地字幕提取与搜索
Reddit
读帖子与评论
小红书
阅读、搜索、发帖互动
抖音
视频解析与内容获取
LinkedIn
个人资料、公司页面、职位搜索
小宇宙播客
音频转文字(Whisper 转录)

坦白说,这个列表已经覆盖了绝大多数内容从业者和开发者的日常信息源。一个 Agent 打通全部平台,想象空间非常大。


四、安装与使用:极其简便

安装

打开你的 AI Agent,直接复制粘贴这句话:

帮我安装 Agent Reach(项目地址:https://github.com/Panniantong/Agent-Reach)

一句话就装完了。

如果采用手动安装,pip 命令如下:

pip install https://github.com/Panniantong/agent-reach/archive/main.zip
agent-reach install --env=auto

另外,Agent-Reach 还支持作为 Skill 安装:

npx skills add Panniantong/Agent-Reach@agent-reach

这种方式适合已经在使用 Skill 体系的用户,装完之后跟其他 Skill 一样统一管理维护。

更新也很简单,对 Agent 说一句「帮我更新 Agent Reach」即可。

安全模式

如果你担心自动安装带来的风险,也可以使用安全模式:

帮我安装 Agent Reach(安全模式)

安全模式下,Agent-Reach 不会自动安装系统包,只会告诉你需要装什么,由你来决定是否安装。

平台配置

配置各平台的方式也极其统一——直接对 Agent 说一句话即可:

  • 「帮我配小红书」
  • 「帮我配 Twitter」
  • 「帮我配代理」

需要 Cookie 的平台,最推荐的配置方式是使用 Chrome 插件 Cookie-Editor 导出 Cookie 后进行配置。

以 Twitter/X 为例,不走官方 API,而是通过 Cookie 认证模拟浏览器登录。你需要从浏览器开发者工具中手动复制 auth_token 和 ct0 两个 Cookie 值,贴到配置文件里。需要注意的是,Chrome v130 以上版本加入了 Cookie 加密,自动提取工具已经用不了了,只能手动从 DevTools 中复制。

小红书则可以通过 npm 安装,扫码登录即可:

npm install -g rednote-mcp
rednote-mcp init

Cookie 保存在本地,下次直接使用。

诊断命令:doctor

装完之后跑一下 agent-reach doctor,它会列出所有渠道的状态——绿色的就是能用的,灰色的需要配置。哪个工具没装、哪个 Cookie 过期、哪个代理没设,全部标出来。这个功能非常实用,可以帮你一步步把所有平台都调通。


五、架构一览

从源码结构来看,Agent-Reach 的设计思路很清晰:

每个平台独立为一个 Python 文件——twitter.py 管 Twitter,youtube.py 管 YouTube,各管各的。加新平台就加一个文件,删平台就删一个文件,模块化做得非常干净。

每个模块里有一个 check() 方法,这个方法就是给 agent-reach doctor 用的。它会检查依赖工具是否安装、认证信息是否配置。实际的数据读写不经过 Agent-Reach,装完之后 Agent 直接调用上游工具,中间没有包装层。


六、各平台配置要点

根据实际使用经验,各平台的配置难度不尽相同,以下按难度分级说明:

真正开箱即用

YouTube 底层调用 yt-dlp,视频字幕和评论都能抓,装完直接好使。 GitHub 调用 gh CLI,仓库、Issue、PR 原生支持。 RSS 和网页读取也都内置支持,给个 URL 就能提取正文。

需要简单配置

B站需要配个代理,配置好之后国内平台的内容也能被 Agent 读取了。 Reddit 需要配代理加 Exa 搜索引擎,技术讨论一搜就有。

需要稍微折腾

Twitter/X 是重头戏。不走官方 API,用 Cookie 认证模拟浏览器登录。需要从浏览器开发者工具手动复制 Cookie 值。拿到的 Cookie 有有效期,过期了需要重新提取。建议使用小号操作,不要用主号冒险。

Exa 搜索 功能很好,全网语义搜索,对 AI 非常友好。但在 Windows 上有一个编码 bug——mcporter 输出的 Unicode 字符(比如勾选符号)会导致 Python 的 GBK 解码崩溃。Mac 和 Linux 用户没有这个问题。手动修改源码,在相关调用处添加 encoding='utf-8' 即可修复。


七、必须说的风险

虽然 Agent-Reach 解决了很大的痛点,但也有一些风险需要了解:

Cookie 认证的封号风险

Twitter 和小红书使用的都是 Cookie 认证方式,说白了就是模拟你的浏览器登录状态去抓取数据。如果平台检测到异常行为,是有可能封号的。官方文档也明确建议使用专用小号。

认证信息的本地存储

Cookie 和 Token 存储在 ~/.agent-reach/config.yaml 文件中,权限设置为 600,只有你自己能读。但如果你的电脑被入侵,这些认证信息就暴露了。建议定期检查这个文件,不再使用后及时清理。

平台接口的稳定性

xreach CLI 通过绕过官方 API 的方式来获取 Twitter 数据,Twitter 随时可能更改接口策略。届时可能需要等待社区更新。这类依赖第三方工具的方案,稳定性无法与官方 API 媲美。

Windows 兼容性问题

Agent-Reach 的底层工具很多是为 Mac 和 Linux 设计的。Windows 用户可能会遇到一些奇怪的问题,比如编码错误、子进程找不到可执行文件等。有些问题搜索都搜不到,因为 Mac 用户根本遇不到。如果你是用 Windows 的开发者,需要有一定的排查耐心。


八、和官方 API 的成本对比

以 Twitter/X 为例,看一下成本差距:

官方 API 方案:

免费 tier 极度受限,基本只能发推不能读。2 月份新出了按量计费,综合下来每月成本约 200 美元以上(约 1400+ 人民币)。

Agent-Reach 方案:每月 0 元。

当然,免费的代价是前文提到的那些风险——Cookie 可能过期、平台接口可能变动、有一定封号风险。

但对于个人开发者和小团队来说,每月节省 200 美元起步,这个账非常好算。如果还在探索期、验证期,先用 Agent-Reach 跑起来验证需求,等业务稳定了再考虑是否升级到官方 API,这是一个非常务实的路径。


九、为什么我觉得这个项目值得关注?

当下 AI Agent 赛道,大家都在追求「更智能」、「更自主」、「更像人」。但一个连网都上不了的 Agent,再聪明又有什么用?

各大社交媒体和内容平台早已筑起了严格的数据访问限制。接口收费、跨域限制、Cookie 验证、代理配置……这些技术门槛叠加在一起,让绝大多数 AI Agent 实际上处于「断网」状态。

Agent-Reach 没有追逐花哨的技术噱头,而是回到了一个最朴素的问题:如何让 Agent 真正获取信息?

它把跨平台的信息获取能力封装成了开箱即用的工具。不需要手动调试环境,不需要对接接口,不需要关注复杂的访问限制。哪怕你没有技术背景,也能轻松让 Agent 帮你从 Twitter、小红书、B站、公众号上获取你需要的信息。

对于内容创作者、市场运营、产品经理、独立开发者来说,这无疑是一个生产力倍增器。


十、写在最后

AI Agent 的未来,不应该只停留在「写代码」和「做方案」这两个场景里。

当 Agent 能够真正自由地获取全网信息,实时整合内容——那时候,它们才有可能成为我们真正的数字助手,而不仅仅是一个更聪明的代码补全工具。

Agent-Reach 迈出了这一步。但需要客观看待的是,这不是一个成熟的商业产品,而是一个开源脚手架。Cookie 认证方案有风险,平台接口可能变动。它适合愿意折腾、能接受偶尔抽风的开发者。

如果你每个月在 Twitter API 上花 200 美元以上,值得试试。 如果你想让 Claude Code 或者 Cursor 能读读网页内容,也值得试试。 但如果你在做企业级产品,需要 99.9% 的稳定性,还是建议使用官方 API。

项目地址:https://github.com/Panniantong/Agent-Reach

参考来源:GitHub Agent-Reach 项目主页、社区讨论及相关技术报道

如果觉得有帮助,欢迎点赞、在看。

图片

 

50.7k Stars! Agent-Reach is open source, letting AI Agents access information across the web for free

图片

AI Agents are clever, but the moment you let them online they go blind? This project cures that

1. The AI Agent's awkward spot: it can write code but cannot get online

Over the past year the AI Agent track has kept heating up.

Claude Code, OpenClaw, Cursor, Windsurf… agents keep multiplying, writing code and building plans, looking more all-round capable by the day.

But anyone using Agents frequently in daily work has surely hit the same painful problem:

Ask the Agent to check the latest sentiment on Twitter? No. Ask it to search Xiaohongshu for real user reviews? No. Ask it to extract subtitles from a Bilibili video to analyse? Still no.

It is not that Agents lack intelligence, but that getting 'online' has a far higher threshold than people imagine.

Every platform has its own access restriction mechanisms, cookie verification flows and interface pricing strategies. Getting one platform working takes anywhere from half a day of fiddling to three days of troubleshooting. All the energy for building a plan gets burnt on debugging the environment.

The open-source project Agent-Reach, introduced today, exists precisely to solve this problem.

图片

2. What is Agent-Reach?

In one sentence: it is a CLI tool that gives an AI Agent one-click 'access to information across the web'.

Project URL: https://github.com/Panniantong/Agent-Reach

It has already gathered 50.7k Stars on GitHub, and well-known tech media including GitHubDaily, HelloGitHub and Ruan Yifeng's newsletter have recommended it; SitePoint and Substack also ran feature coverage. The level of attention from the developer community is evident.

Agent-Reach's positioning is very clear: it is a scaffolding tool, not a framework.

A framework wraps layer after layer, until you no longer know what the underlying tool looks like. Scaffolding only helps you set up the environment; the actual work is still done by upstream tools. After installation, your Agent calls the upstream tools directly — reading Twitter calls xreach directly, reading YouTube calls yt-dlp directly. There is no wrapper layer in between, so performance loss is zero.

This design has a benefit: when an upstream tool updates you do not have to wait for Agent-Reach to catch up, and every link in the toolchain can be replaced independently.

Agent-Reach packages all of the following tedious work into ready-to-use command-line tools:

per-platform interface integration

dependency installation and environment configuration

proxy settings and login configuration

data cleaning and formatted output

Whichever you use — Claude Code, OpenClaw, Cursor or Windsurf — as long as you can run a command line, Agent-Reach works seamlessly.

A particularly thoughtful point: it is fully open source and free, requires no API key, and uses the MIT licence (the most permissive open-source licence, free to use and modify).

3. Which platforms does it reach?

Agent-Reach covers mainstream content platforms you use frequently, divided into two categories:

Out of the box (zero configuration)

Platform

Capability

Any web page

Read page content, automatically clean the HTML

YouTube

Extract subtitles, search videos, using yt-dlp underneath

RSS

Subscribe to and read any RSS/Atom feed

WeChat official accounts

Search articles, read full text, output Markdown

Weibo

Hot search list, user posts, comments

V2EX

Hot posts, node browsing, details plus replies

Web-wide search

Semantic search, no API key required

Unlocked with simple configuration (one line of config)

Platform

Capability

GitHub

Private repositories, issues/PRs, forks, via the gh CLI

X (formerly Twitter)

Search tweets, browse the timeline, post tweets

Bilibili

Local subtitle extraction and search

Reddit

Read posts and comments

Xiaohongshu

Read, search, post and interact

Douyin

Video parsing and content retrieval

LinkedIn

Personal profiles, company pages, job search

Xiaoyuzhou podcast

Audio to text (Whisper transcription)

Honestly, this list already covers the daily information sources of the vast majority of content workers and developers. An Agent connected across all these platforms opens up enormous possibilities.

4. Installation and usage: extremely simple

Installation

Open your AI Agent and copy-paste this one line:

Installed in a single line.

If you prefer to install manually, the pip command is:

Agent-Reach also supports installation as a Skill:

This suits users already on a Skill-based system — once installed it is managed and maintained like any other Skill.

Updating is simple too — just say 'help me update Agent Reach' to your Agent.

Safe mode

If you are worried about the risks of automatic installation, you can use safe mode:

In safe mode, Agent-Reach will not automatically install system packages; it only tells you what needs installing and you decide whether to install it.

Platform configuration

Configuring each platform is extremely uniform too — just say one line to your Agent:

'Help me set up Xiaohongshu'

'Help me set up Twitter'

'Help me set up the proxy'

For platforms that need cookies, the most recommended approach is to export cookies with the Chrome extension Cookie-Editor and then configure them.

Taking Twitter/X as an example, it does not use the official API, but instead simulates a browser login via cookie authentication. You need to manually copy the auth_token and ct0 cookie values from your browser's developer tools and paste them into the config file. Note that Chrome v130 and above added cookie encryption, so automatic extraction tools no longer work and you must copy manually from DevTools.

Xiaohongshu can be installed via npm, and you just scan the code to log in:

The cookie is stored locally and used directly next time.

Diagnostic command: doctor

After installing, run agent-reach doctor and it will list the status of every channel — green means usable, grey needs configuring. Which tool is missing, which cookie has expired, which proxy is unset, all get flagged. This feature is very practical and walks you through getting every platform working.

5. Architecture overview

Judging from the source structure, Agent-Reach's design thinking is very clear:

each platform is a single Python file — twitter.py handles Twitter, youtube.py handles YouTube, and each keeps to its own. Adding a platform means adding a file, removing one means deleting a file. The modularisation is very clean.

each module has a check() method, which is what agent-reach doctor uses. It checks whether dependency tools are installed and whether authentication is configured. The actual data reading and writing does not pass through Agent-Reach — after installation the Agent calls upstream tools directly, with no wrapper layer.

6. Key configuration points per platform

Based on hands-on experience, configuration difficulty varies per platform, so here it is graded by difficulty:

Genuinely out of the box

YouTube calls yt-dlp underneath and can fetch both subtitles and comments — install and it just works. GitHub calls the gh CLI, giving native support for repositories, issues and PRs. RSS and web page reading are also built in — give it a URL and it extracts the body text.

Needs simple configuration

Bilibili needs a proxy configured; once set up, content from domestic platforms becomes readable by the Agent too. Reddit needs a proxy plus the Exa search engine, so technical discussions are easy to find.

Needs a bit of work

Twitter/X is the main event. It does not use the official API, instead simulating a browser login via cookie authentication. You need to copy cookie values manually from your browser's developer tools. Cookies have an expiry and must be re-extracted when they expire. Use a secondary account; do not risk your main account.

The Exa search function is excellent — web-wide semantic search, very friendly to AI. There is however an encoding bug on Windows — Unicode characters output by mcporter (such as checkbox symbols) crash Python's GBK decoding. Mac and Linux users do not have this problem. Manually editing the source and adding encoding='utf-8' to the relevant call fixes it.

7. Risks that must be mentioned

Although Agent-Reach solves a big pain point, there are risks worth understanding:

Account ban risk from cookie authentication

Twitter and Xiaohongshu both use cookie authentication, which essentially simulates your browser login state to scrape data. If a platform detects abnormal behaviour it may ban the account. The official documentation explicitly recommends using a dedicated secondary account.

Local storage of authentication details

Cookies and tokens are stored in the ~/.agent-reach/config.yaml file with permissions set to 600, readable only by you. But if your computer is compromised, these credentials are exposed. Check the file periodically and clean it up when you no longer need it.

Stability of platform interfaces

The xreach CLI obtains Twitter data by bypassing the official API, and Twitter may change its interface policy at any time. You may then have to wait for a community update. Such solutions depending on third-party tools cannot match the stability of the official API.

Windows compatibility

Many of Agent-Reach's underlying tools are designed for Mac and Linux. Windows users may hit odd problems such as encoding errors and subprocesses failing to find executables. Some of these cannot even be found by searching, because Mac users never encounter them. If you are a developer on Windows you need some patience in troubleshooting.

8. Cost comparison with the official API

Taking Twitter/X as an example, let us look at the cost gap:

Official API option:

the free tier is severely limited — basically posting but not reading. In February they introduced pay-as-you-go pricing, which works out to roughly 200 dollars or more per month (about 1,400+ RMB).

Agent-Reach option: 0 yuan per month.

Of course, the price of free is the risks mentioned above — cookies expiring, platform interfaces changing, some risk of account bans.

But for individual developers and small teams, saving 200 dollars a month from the start makes this an easy calculation. If you are still in the exploration and validation phase, get things running with Agent-Reach to validate the need, and consider upgrading to the official API once the business is stable. That is a very pragmatic path.

9. Why do I think this project is worth attention?

In AI Agents today, everyone is chasing 'smarter', 'more autonomous' and 'more human-like'. But an Agent that cannot even get online — how useful is being smart?

Major social media and content platforms have long since built strict data access restrictions. Interface fees, cross-origin restrictions, cookie verification, proxy configuration… these technical barriers stack up, leaving the vast majority of AI Agents effectively 'offline'.

Agent-Reach did not chase flashy technical gimmicks, but returned to the most primitive question: how do you actually let an Agent get information?

It packages cross-platform information access into ready-to-use tools. No manual environment debugging, no interface integration, no worrying about complex access restrictions. Even without a technical background you can easily have an Agent fetch the information you need from Twitter, Xiaohongshu, Bilibili and WeChat official accounts.

For content creators, marketing operations, product managers and independent developers, this is undoubtedly a productivity multiplier.

10. Wrapping up

The future of AI Agents should not be confined to the two scenarios of 'writing code' and 'making plans'.

Only when Agents can genuinely and freely access information across the web and integrate content in real time can they become our true digital assistants, rather than just a smarter code completion tool.

Agent-Reach has taken that step. But it should be viewed objectively: this is not a mature commercial product, but an open-source scaffold. Cookie authentication carries risk and platform interfaces may change. It suits developers willing to tinker and accept occasional breakage.

If you are spending more than 200 dollars a month on the Twitter API, it is worth trying. If you want Claude Code or Cursor to be able to read web content, it is worth trying. But if you are building an enterprise product needing 99.9% stability, you should use the official API.

Project URL: https://github.com/Panniantong/Agent-Reach

Sources: the GitHub Agent-Reach project homepage, community discussions and related technical coverage

If you found this helpful, a like and a share are welcome.

图片

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry