Agent-Reach is a capability layer tool that provides AI Agents with internet access capabilities, with the goal of enabling Agents to read and search content on platforms such as Twitter, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu. The project is provided in CLI form, and the official statement is that reading and searching incur no API fees.
[Purpose and Positioning] AI Agents can already write code, edit documents, and manage projects, but they are often limited when going online to find information: YouTube subtitles are unavailable, the Twitter API requires payment, Reddit rejects server IPs with 403, Xiaohongshu requires login, Bilibili is blocked by risk control, web pages come back with a pile of HTML tags, and RSS requires writing your own code. Agent-Reach handles these barriers in a unified way. It does not build a new scraping tool; rather, it sits one level above specific implementations and is responsible for selection, installation, health checks, and routing. The actual reading is done by the Agent directly calling upstream tools, with no wrapper layer.
[Main Capabilities and Supported Channels] Zero-configuration channels (6 activated by default) include: reading any web page via curl https://r.jina.ai/URL , viewing a GitHub repository with gh repo view owner/repo, extracting YouTube subtitles with yt-dlp, searching Bilibili with bili search (no login required), performing semantic search across the web with Exa, and parsing subscribed RSS with feedparser. Channels that require a logged-in state include Xiaohongshu, Twitter, Reddit, Facebook, and Instagram. The Agent will present a menu and ask the user which ones are needed, and will install them only after they are named. Each platform maintains an ordered list of preferred plus backup backends. The channel file actually probes each candidate backend in order (not just checking whether the command exists), and the first fully available one is selected; broken ones are given a repair prescription. Changing the access method only requires adjusting the order of the list, not rewriting code. The agent-reach doctor command can show the status of each channel and which backend is currently in use.
[How to Use] The user copies a specified one-sentence instruction to an AI Agent (such as Claude Code, OpenClaw, Cursor, etc.), and the Agent will complete the subsequent installation on its own. For users who have already installed it, updating likewise requires only one sentence. The installation process includes: installing the agent-reach command line from this repository (it comes with yt-dlp and feedparser; the official notice says not to install the package of the same name from PyPI, as that is not this project); checking system infrastructure such as Node.js, gh CLI, and mcporter and providing installation methods for missing items; installing dependencies and connecting to Exa via MCP when --system is explicitly passed; detecting whether the environment is local or a server; writing to the Agent's skills directory when --system is explicitly passed; and by default activating only zero-configuration channels. The official emphasis is that agent-reach install by default only checks the environment and will not automatically install system packages or write configuration; only after explicitly allowing system modifications should agent-reach install --system be used. Users of OpenClaw need to first confirm that exec permission is enabled, because Agent-Reach depends on the Agent executing shell commands; you can set "tools": { "profile": "coding" } in ~/.openclaw/openclaw.json, restart the Gateway, and start a new conversation.Other platforms are not subject to this restriction. In terms of configuration, users do not need to memorize commands. After reading SKILL.md, the Agent will determine the invocation on its own; for platforms that require login, just tell the Agent "help me configure XXX."
[Security and Restrictions] Cookies are stored only locally, are not uploaded or shared externally, and the code is fully open source. Twitter only accepts content manually exported by the user via Cookie-Editor; Agent-Reach does not perform Xiaohongshu login on behalf of the user, nor does it read Xiaohongshu browser cookies; OpenCLI only uses a Chrome session that already exists and is explicitly controlled by the user; agent-reach configure xhs-cookies will not inject cookies into OpenCLI/Chrome. After Twitter cookies are saved, they are only used by agent-reach doctor to check whether the configuration is complete. Before directly running the upstream twitter command, TWITTER_AUTH_TOKEN and TWITTER_CT0 still need to be explicitly set in the current process environment. Official reminder: For platforms that log in using cookies (Twitter, Xiaohongshu, etc.), calling them through scripts/APIs carries the risk of being detected by the platform and having the account banned. Be sure to use a dedicated alt account, not your main account, because of the risk of account bans and the security risk that cookies are equivalent to full login permissions. A local computer does not need a proxy; a proxy is only needed when deploying to a server (about $1/month). Uninstalling will clear ~/.agent-reach/ (including all tokens/cookies), each Agent's skill files, and the MCP configuration in mcporter; to uninstall the Python package itself, use pip uninstall agent-reach.
[Ecosystem and Dependencies] The upstream tools involved include OpenCLI, twitter-cli, rdt-cli, xiaohongshu-mcp, xhs-cli, bili-cli, yt-dlp, Jina Reader, Exa, mcporter, feedparser, and mcp-server-linkedin. The project author states that it will continue to be maintained, tracking platform changes, integrating new channels, and ensuring the channels are as usable, practical, and free as possible.
Regarding specific fees, prices, the complete list of platforms, detailed steps for each channel, etc., the official materials do not provide them one by one, and they are not provided for now.
