← Back to blog
AiAbout 7 min read

Agent Reach Hands Your Agent the Internet It Cannot Log Into

Published Nov 14, 2026
Agent Reach Hands Your Agent the Internet It Cannot Log Into

A coding agent with a search tool can find a blog post. Ask it what people are saying on Xiaohongshu about a skincare product and it will either fail or invent an answer. Ask it what is said in the first four minutes of a Bilibili video and it has nothing to read.

Agent Reach, a project that hit the top of GitHub's trending list this week and sits near 9,500 stars, is built around that gap. The English one-liner is "give your agent eyes to see the entire internet". The Chinese one is more precise: let the agent reach the parts of the internet it cannot get to on its own.

It is MIT licensed, needs Python 3.10 or newer, and is not on PyPI. That last point is in the README in bold because someone else registered the name agent-reach there.

What it actually wraps

The tool is a channel router. It installs and tests the scrapers and API clients behind a list of platforms, then exposes one command shape:

agent-reach channels
agent-reach bilibili search "大模型"
agent-reach read https://www.v2ex.com/t/123

The read command inspects the URL and picks the channel itself, which is the part that saves the most typing in practice.

Fourteen channels are listed. Six are on by default and need no account: Bilibili (search, video metadata, trending, subtitles with a login for subtitles), V2EX (trending, nodes, posts and replies, users), GitHub (repo search, repo and README reads, issues and PRs with comments), YouTube (metadata, search, subtitles, and audio transcription with a Groq or OpenAI key), plus a generic web search and reader as a fallback.

The optional ones are where the interesting data lives: Xiaohongshu for notes, body text and comments; Xueqiu for live quotes and trending posts; BOSS Zhipin for job listings and descriptions; Xiaoyuzhou for podcast transcription; LinkedIn for public job posts; Reddit for posts, bodies, comments and subreddits. X, Facebook and Instagram are marked experimental, which in scraping terms means they work until the platform changes something.

The install is a prompt, not a command

This is the part that tells you who the project is for. The README does not lead with a pip line. It leads with a sentence to paste into your agent:

帮我安装 Agent Reach:https://raw.githubusercontent.com/Panniantong/Agent-Reach/main/docs/install.md

The agent reads that document and does the work: installs uv if it is missing, then the program and the channel tools, then reports which channels are live and which optional ones can be added. Adding Xiaohongshu later is a sentence: "add Xiaohongshu". Removing it is another sentence, and removal also wipes that channel's login and tools.

Everything lives under ~/.agent-reach/. That single directory is a deliberate design choice. Agents accumulate half-installed dependencies across a dozen locations, and the only reliable uninstall is a fresh machine. One folder means upgrade, inspect, and remove are all one command, and it means a second agent can be given the same capabilities by pointing it at the same folder.

The session problem, stated honestly

Most of the platforms people actually want are behind a login. Agent Reach handles this in a way that deserves a careful read rather than a summary.

The first time you use a site, it copies your existing Chrome session for that site into its own dedicated browser, which runs in the background without popping windows. It does not read your password manager. Your original browser is untouched.

That is a reasonable engineering answer to an unreasonable problem. Logging an automated browser into X or Xiaohongshu by script is a fast route to a ban, so session reuse is what everyone does. The README is candid about the consequence: what gets copied is usually your main account, and it recommends swapping in a secondary one. There is a security section that explains how.

Two things follow. First, an agent with your Xiaohongshu session can read your notes, your saves, and your DMs if the channel exposes them. Second, your account is now the rate-limit identity for whatever the agent does, which means a loop that runs two hundred searches is your account running two hundred searches.

Neither problem is unique to this project. Both become easier to ignore when installation is a single sentence to an agent, which is exactly why the README puts the warning where you cannot miss it.

Why Chinese platforms are the real target

For an English-speaking developer, the headline feature is Reddit. For the project's actual audience, it is the Chinese half of the list.

General web search is bad at Chinese social platforms in a specific way. It indexes the pages that are public and linkable, which excludes most of the content people care about, and it has no idea what a note on Xiaohongshu says because that content is rendered client-side behind a login wall. An agent asked to research Chinese consumer sentiment on general search will produce a summary of three marketing blog posts and call it a trend.

Bilibili subtitles are the other half. A forty-minute technical talk in Chinese is invisible to an agent unless someone transcribes it, and the platform already has the transcript. Getting at it turns a video into text, which is the only form an agent can reason over.

I have been running research for this blog against Chinese platforms for months, and the standard toolchain fails on exactly these surfaces: Weibo returns HTML that parses to nothing, Zhihu demands auth, Bilibili answers 412 to anything that looks automated. A project that packages working clients for them, with per-channel health reporting, is solving a problem that people who do this work recognize immediately.

The tradeoff nobody puts in the tagline

"Zero API fees" is the phrase in the listing, and it is accurate in the narrow sense: you are not paying per call. You are paying in a different currency.

Scraping means you inherit the platform's tolerance. Channels break when a site changes its markup, and the README classifies X, Facebook and Instagram as experimental precisely because their anti-bot defenses move. A research pipeline built on a scraped channel has an availability profile closer to a hobby project than to an API, and it fails at the moment you least want it to.

It also means the answers carry the biases of the surface being scraped. Xiaohongshu search returns what the ranking favors, which is not a sample of opinion, it is a sample of what performed. An agent that reads the top twenty notes and summarizes them will report the consensus of people who write notes that rank, which is a narrower group than it sounds.

That argues for treating the tool as a source of raw material rather than an oracle, which is also the only way to notice that a channel has started returning garbage.

What I would use it for

Three jobs, in descending order of confidence.

Pulling subtitles from conference talks and product demos, because the transcript is ground truth and the platform already computed it.

Reading a specific thread or note you already have a link to, which is a retrieval task with a known answer and no sampling problem.

Cross-checking a claim across two markets. If you are writing about a tool's reception, reading the Reddit thread and the Xiaohongshu notes on the same week gives you two distributions instead of one, and the differences are usually the story.

What I would not do is point it at an open-ended question and trust the summary. The channels are good at fetching and bad at representing, and an agent will not tell you which one it did.

Related articles