Agent Reach: sixteen platforms, six of them free of configuration

The install is a URL you paste to your agent, not a command

The quick start is two fenced blocks containing a single sentence each, and the sentence holds a raw.githubusercontent.com address pointing at docs/install.md, with a second one for docs/update.md. You paste that sentence to Claude Code, OpenClaw, Cursor or Windsurf, and the agent does the rest. The repository has a CHANGELOG.md, a version in pyproject.toml and three tagged releases, and the install path points at the main branch rather than any of them.

Consequence for the reader: what executes is decided by whatever is in docs/install.md at the moment you paste the link, so there is no versioned way to install this. You can pin a release to read the source and you can pin a release to build it, but the documented path for a user always fetches the head of the branch. There is a first-run prerequisite worth knowing before that, and it is agent-specific: on OpenClaw the agent cannot run commands until you change the tool profile, and the README gives both the command and the file to edit.

openclaw config set tools.profile "coding"

Afterwards the gateway needs a restart, and the other named agents are not affected.

install checks by default and only writes when you pass --system

The install is described in six numbered steps, and three of them are gated on an explicit flag. Installing the CLI from this repository is step one, bundled with yt-dlp and feedparser. Checking system infrastructure is step two, covering Node.js, the gh CLI and mcporter and reporting what is missing and how to get it. Steps three, five and the dependency half of step four only happen when the system flag is passed, which is what installs dependencies, connects the Exa search backend over MCP, and writes SKILL.md into the agent's skills directory.

Consequence for the reader: the default really is read-only, and it is unusual for a tool in this category, so the risk of a first run modifying your machine is low. The cost of that design is that the default also leaves most of the product switched off. Step six says the default activates six zero-config channels and puts the login-gated ones into a menu you pick from by name, so a first run gives you a working subset and a list of things it is asking permission for.

Six channels work with nothing, and the table says which

The platform table is the honest part of this README. It has sixteen rows and two useful columns. Working out of the box with no configuration are: reading any web page, YouTube subtitle extraction and video search, reading any RSS or Atom feed, and V2EX, which covers hot posts, node posts, post details with replies and user information. GitHub reads public repositories and searches. LinkedIn reads public pages through a reader service. Bilibili search and video details work through a separate command line tool with no login.

Consequence for the reader: nine rows have something in the out-of-the-box column, but three of those are partial. Twitter reads a single tweet and nothing else without configuration. LinkedIn gives you public pages and not profiles, companies or jobs. One row gives a link health check rather than search. That is where six comes from, and it is the number the install step uses. So the count to plan against is six fully working channels, and the three rows that look available in the table are the ones most likely to disappoint.

Reddit has no zero-config path, and the table admits it

One row states outright that there is no zero-config path, because the anonymous interface has been blocked. What you get instead is search plus reading posts and comments, and only after you either install a desktop tool that uses your browser's login state or set up a separate command line tool with a cookie. Facebook and Instagram are the same shape with nothing out of the box at all: both need the desktop tool, and both explicitly reuse the Chrome session you already have. Xiaohongshu also has nothing out of the box, and its row is more restrictive still.

Consequence for the reader: the platform you are most likely to install this for is often the one that cannot be set up from zero, and the table says so before you try rather than after. Every one of these paths terminates in a manual step performed by you in a browser. That is a deliberate design, and the next section explains why, but the practical effect is that a login-gated channel is a project rather than a toggle.

The cookie rules are narrow, and one of them stores a credential that does nothing

The README is explicit about what it will not do. Cookies stay on your machine and are not uploaded or passed on. The project states it does not perform the Xiaohongshu login for you and does not read Xiaohongshu browser cookies, that the desktop tool only uses a Chrome session you already have and clearly control, and that the command to configure Xiaohongshu cookies will not inject them into the browser or the tool. Where a cookie is required you are told to export it yourself with a separate extension.

The Twitter arrangement is the one to read twice. Twitter accepts only content you exported by hand. Once saved, the cookie is used for one purpose: letting the diagnostic command check whether your configuration is complete. To run the underlying twitter command yourself you still have to set the token and the second token explicitly in the current process environment.

Consequence for the reader: the boundaries are sensible and they are the reason this is not a credential-harvesting tool, but they leave a real cost. No channel reaches a logged-in site on its own, and the Twitter path stores a value that does not then do anything, which is confusing unless you read the note. Budget for doing the login work yourself on every gated platform.

Free holds for the zero-config channels, and the paid ones are the rest

One of the summary rows claims everything is free, that all tools are open source and all APIs are free, and that the only possible cost is a server proxy at about a dollar a month which a local machine does not need. The example environment file complicates that in a specific way. It asks for an Exa search key with a note of a free monthly allowance, a GitHub token described as optional for higher rate limits, a Reddit proxy in the form of a URL with a username, password, host and port, a Groq key for transcription, and an OpenAI key described as an optional fallback for when Groq is rate-limited.

Consequence for the reader: the free claim and the platform table are consistent, and the pattern is the point. The channels that work with nothing configured are the free ones, and every channel that costs money or draws on a quota is a channel that is not configured for you. The Reddit proxy line is the clearest case: full access is the paid path, and free access is the path nothing is set up for. The same pattern holds for search, which needs a key or an MCP connection, and for podcast transcription, which needs a Whisper key.

The same PyPI name is not this project, and three releases landed in two days

Step one of the install says to take the agent-reach command from this repository, and warns in the same sentence not to install the package of the same name from PyPI, because it is not this project. The release history explains the sensitivity. Three releases shipped inside two days in June 2026: a fix for source installation failing because of duplicate wheel packaging, then a release whose title describes honest slimming by removing unmaintained channels and adding transcription and a JSON output for the diagnostic command, then a release for multi-backend routing, real health checks and the desktop tool. No release has appeared since, although commits continued to 2026-09-15.

Consequence for the reader: a name collision on the package index plus a wheel packaging bug means the documented install path is the only one to trust, and a user who searches the index instead of the repository can end up running unrelated code under the same name. The packaging metadata does not help you tell them apart, because the project declares its own repository address as the packaging Homepage, so the index record has no separate site to compare against.

Editorial conclusion

Adopt Agent Reach for the six channels that work with no account and no key, which is web pages, YouTube subtitles, RSS, V2EX and public GitHub, and treat everything past that as a separate decision you make per platform. Do not install it for Reddit, Facebook, Instagram or Xiaohongshu expecting the documented flow, because every one of those needs a session you create by hand and the project deliberately refuses to create it for you. Before you run anything, check three things: that your agent can execute shell commands at all, which on OpenClaw means setting the tools profile to coding and restarting the gateway, that the install is reading docs/install.md from the main branch at the moment you paste it, since the repository has tags and the install path deliberately does not use them, and that you have not pip-installed a package of the same name from PyPI, which the README states is not this project.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. Release notes