The Agent Reach Setup
Give your AI agent the live internet in one pasted sentence.
Agent Reach is a free, open source tool that gives AI agents like Claude Code the ability to read and search the live internet across fifteen channels, from YouTube transcripts to Reddit threads to Twitter, with no API fees. It hit 62,683 GitHub stars within five months of its February 2026 release. Install is one sentence you paste to your agent, six channels work with zero setup, and the rest unlock when you ask. I run it inside my own research workflow. The one rule that matters: logged-in platforms get a dedicated spare account, never your main.
Read time: 9 min
Why can't the smartest software you use read a YouTube video?
Because the platforms your research lives on were never built to be read by tools. Ask your agent what a YouTube tutorial covers and it can't get the transcript. Ask it what Twitter thinks of a product and the official API wants money. Send it to Reddit and the anonymous route is blocked. Every platform has its own wall, and climbing each one means its own tool, its own configuration, and often its own bill.
Agent Reach is the free, open source answer, and the traction is real: 62,683 stars on GitHub within five months of its late February 2026 release, a number that puts it among the fastest growing AI tools of the year, plus a #1 spot on GitHub's trending list along the way. It's MIT licensed, built by a developer called Panniantong, and its own README states the economics plainly: every tool it uses is open source, every API it touches is free, and the only possible cost is about a dollar a month for a proxy if you run it on a server, which a personal computer doesn't need.
The barrier that used to exist was the setup itself. Wiring an agent into even one platform meant finding the right scraper, installing dependencies, and debugging authentication. Agent Reach collapsed all of that into a single pasted sentence, because the agent installs it for you.
The idea in plain words
Three terms and you know everything this playbook requires.
An AI agent is an AI that can run commands on your computer rather than just chat. Claude Code is the one most people reading this use, and Cursor, Windsurf, and OpenClaw all count too. If your AI can create files on your machine, it's an agent.
A terminal is the text window where those commands run. For this setup you barely touch it, because the agent does the typing.
A cookie is the small file that proves to a website that you're logged in. A few platforms in this playbook only open up when the agent can borrow a login, and that detail drives the single most important safety rule below.
What Agent Reach actually is: a layer that picks the best current method for reading each platform, installs it, health-checks it, and routes your agent's requests to whichever method works today. When a platform blocks one route, the project swaps in another, and you don't have to care. That already happened in practice: when Bilibili blocked the popular downloader yt-dlp in June 2026, the project switched that channel to a different tool with zero action needed from users.
The install, which is genuinely one sentence
Open your agent and paste this:
Help me install Agent Reach: https://raw.githubusercontent.com/Panniantong/agent-reach/main/docs/install.md
That's the entire install. The agent reads the instructions at that link and does everything itself: it installs the command line tool, adds the supporting programs it needs, connects a free semantic search engine with no API key required, checks whether you're on a personal computer or a server, and registers a skill file so it knows to use these abilities on its own in future conversations. At the end it asks which of the login-required platforms you want, and only installs the ones you name.
If you're cautious about a tool changing your system, and on a shared or work machine you should be, there's a built-in safe mode: add "install it in safe mode with the --safe flag" to the sentence, and it will list what it wants to install and let you approve each piece instead of proceeding automatically. There's also a preview mode, --dry-run, that makes no changes at all and just shows you the plan.
The rest of this playbook is below.
Enter your email to read it. You also get one new playbook every week.
One email a week. Unsubscribe any time.
Updating later is the same trick with a different link:
Help me update Agent Reach: https://raw.githubusercontent.com/Panniantong/agent-reach/main/docs/update.md
One platform-specific note. If your agent is OpenClaw, it needs permission to run commands before any of this works, which its default configuration blocks. Ask it to set its tools profile to "coding" and restart, or the install will stall at the first step. Claude Code, Cursor, and Windsurf need nothing.
What works the moment install finishes
Six channels are live with zero configuration, and you use them by asking in plain English rather than remembering any commands.
Any web page: "read this link for me," and it comes back clean instead of as a wall of code. YouTube: "what does this video cover," and it pulls the transcript, or searches YouTube directly. GitHub: "what does this repository do," including issues and discussions on public projects. RSS: "subscribe to this feed and tell me what's new." Bilibili search and video details, no login needed. And full-web semantic search through Exa, which the installer wires up automatically with no key, so "search the web for recent comparisons of X" just works.
A useful seventh: single Twitter posts read without any setup. Hand it a tweet link and it can read that tweet. Searching Twitter or browsing timelines is a different tier, covered next.
Run one command after install to see all of this for yourself: ask the agent to run agent-reach doctor. It reports every channel, whether it's working, which route it's currently using, and how to fix anything that isn't. Make it a habit, because it's also your first diagnostic when something stops working months from now.
The platforms that unlock with a login
The richer platforms open when the agent can act as a logged-in user. For each one, you just tell the agent "help me set up X" and it walks you through what that platform needs. Twitter search and timelines. Reddit search, posts, and comment threads, which now require this since the anonymous route was blocked. LinkedIn public pages work through the built-in reader already, with deeper profile and company detail behind a configured route. Facebook and Instagram open through a desktop tool that reuses an existing Chrome login, covering profile lookups, recent posts, and browsing. The Chinese platforms XiaoHongShu, Weibo-adjacent communities, and the Xiaoyuzhou podcast transcriber unlock the same way, and a stock market channel for Xueqiu rounds out the fifteen-channel table.
Now the rule that matters more than anything else in this playbook.
Any platform that works through a login gets a dedicated spare account, never your main one. This comes straight from the project's own security notes, and the reasoning is simple. Platforms can detect tool-driven access on a logged-in account and restrict or ban it, and a cookie is equivalent to your full login, so if a credential ever leaked, you want the blast radius to be an account that doesn't matter. Five minutes creating a spare account is the whole cost. Your credentials stay on your own machine in a permission-locked file the project never uploads anywhere, which is good design, but local storage protects you from the developer, and only the spare account protects you from the platform.
One more boundary worth stating plainly, because it's the difference between research and trouble: reading a LinkedIn profile someone sent you is research, while bulk-harvesting profiles for lead lists is against LinkedIn's user agreement and a fast way to lose an account. This tool is for giving your agent eyes, and what you point them at is on you.
The part most coverage skips: what to actually do with it
Every video about this tool shows the install and stops. The value shows up in the workflows after, so here are the three I'd set up first, in order of payoff.
Research briefs from video. "Pull the transcript of this YouTube video and give me the five claims worth checking, with timestamps." An hour-long talk becomes a two-minute read, and the transcript route is one of the most reliable channels in the whole tool.
Community due diligence. "Search Reddit for people discussing [product or tool] in the last few months and summarize what users complain about." Vendor marketing tells you the top line, and the complaints thread tells you the truth. This is the single workflow that best justifies the Reddit login setup.
Standing watch. "Subscribe to these three RSS feeds and this creator's channel. When I ask what's new, check them all and brief me." That's a monitoring system that costs nothing and lives inside the agent you already use daily.
The reason this compounds: the installer registered a skill file, which means the agent now recognizes on its own when a request needs live internet access and picks the right route without being told. The tenth research request works better than the first not because the tool changed, but because you've learned what to ask for.
The three artifacts you need
One, the install sentence, saved wherever you keep snippets. It's also the update mechanism and the thing you'll share when someone asks how you set this up.
Two, a spare account for each logged-in platform you actually use. Named for what it is, secured with its own password, and never mixed with your real identity.
Three, a written research prompt template, 150 to 250 words, and almost nobody has one. The difference between a usable brief and a wall of paste is the instructions you wrap around the request: what you want checked, how many sources, what format the answer comes back in, and what to flag versus ignore. Write it once with slots for the topic and the platforms, and every research task starts by filling it in. The template outlives every individual tool in this stack.
If something goes wrong
The first move is always the same: ask the agent to run agent-reach doctor, read which channel is down and which backend it's on, and follow the prescription it prints. Beyond that, the common failures map cleanly. If the command isn't found right after install, restart the agent so it picks up the new tools. If a platform that worked starts failing, the route probably broke, so update with the update sentence, since swapped routes ship through updates. If Reddit returns nothing, you're on the anonymous path that no longer exists, and it needs the login setup. If Twitter search fails while single tweets work, the search tier needs its configuration finished. If OpenClaw won't run anything, it's the tools profile permission from section 2. And if a login platform suddenly logs out, the cookie expired, which is routine, so redo that platform's setup with the spare account.
Honest limits
Scraping-based access breaks by design. Platforms change, routes die, and the project's whole architecture, ordered backup routes per platform, exists because of that. The practical meaning for you is that any individual channel can be down on any given day, and doctor plus an update is the fix, not a support ticket. In March 2026 a batch of single-platform tools went unmaintained at once and the project rerouted around them, which is the pattern you're signing up for.
The ban risk on logged-in platforms is real and the project says so itself. A spare account converts that risk from serious to trivial, and skipping that step because setup was going smoothly is the one mistake this playbook exists to prevent.
The zero-config experience covers six channels, and the platforms most people get excited about, Reddit, Twitter search, Instagram, all sit behind login setups that take real minutes and carry the rules above. The one-sentence install is true, and the five-minute version of this tool is the shallow version.
And this gives your agent reach, not judgment. It can now read the live internet, and it will read confident nonsense with the same fluency as truth. The research template in section 6 matters precisely because "summarize what you find" produces slop, while "find the three most repeated complaints and link the threads" produces evidence.
What this changes
The gap between an AI that knows things and an AI that can go look has been the quiet ceiling on every research workflow since these tools arrived. What 62,000 developers voted on in five months is that the ceiling is now optional, and the price of removing it is pasting one sentence. Your agent already knew how to think about what it reads. Now it can do the reading.
