Connect AI agents to internet platforms and content sources

Agent Reach is a capability layer that gives an AI agent internet access without its owner having to configure each platform by hand. It supports 15 platforms - web pages, YouTube, RSS, full-web semantic search, GitHub, Twitter/X, Bilibili, Reddit, Facebook, Instagram, Xiaohongshu, LinkedIn, V2EX, Xueqiu, and Xiaoyuzhou podcasts. It does not do the reading itself: it selects, installs, and health-checks the best available upstream tool for each platform (yt-dlp for YouTube, gh CLI for GitHub, Jina Reader for web pages, Exa via mcporter for search, and so on), and the agent calls that tool directly with no wrapper layer in between. Each platform routes through an ordered list of primary and fallback backends, and agent-reach doctor reports which backend each platform is currently using and diagnoses failures.

Use it when you want a coding agent (Claude Code, OpenClaw, Cursor, Windsurf, or anything that can run shell commands) to be able to read a webpage, pull a YouTube video's subtitles, search Reddit, or query a dozen other sites without hand-rolling API integrations and login flows for each one yourself. Six of the 15 channels, including web reading, YouTube, RSS, and full-web semantic search, work with zero configuration; platforms needing a login session (Twitter/X search, Reddit, Facebook, Instagram, Xiaohongshu) require telling the agent to configure that platform first. It is not a fit if you need guaranteed, contractual API access - it deliberately favors free, unofficial routes such as scrapers, browser-session reuse, and community CLIs, which can and do get rate-limited or blocked, and it performs no reading on its own without an agent driving it.

Input is a one-line instruction the agent receives pointing at the project's install doc; the CLI then exposes commands such as:

agent-reach install --env=auto
agent-reach doctor

The first is a safe, read-only check by default (--system opts into installing dependencies and writing config); the second reports per-channel backend status and fixes. Output for each channel is whatever the underlying tool returns - transcripts, search results, post text - consumed directly by the agent, while credentials are cached only locally in ~/.agent-reach/config.yaml with file permissions restricted to the owner.

Agent Reach routes to real upstream tools per platform: Jina Reader for web pages and LinkedIn's public pages, yt-dlp for YouTube, gh CLI for GitHub, bili-cli for Bilibili (with OpenCLI as a fallback after yt-dlp was blocked by Bilibili's anti-scraping measures in mid-2026), Exa via mcporter for full-web semantic search, feedparser for RSS, and OpenCLI, reusing an existing desktop browser session, for Reddit, Facebook, Instagram, and Xiaohongshu. It is MIT-licensed, written in Python 3.10+, and OpenClaw users must switch to the "coding" tools profile before Agent Reach can execute shell commands.

Developers and builders who want their coding agent to reach the open web and social platforms without personally sourcing, installing, and maintaining a separate scraper or API client for every site - and who are comfortable using a secondary account on any platform that needs a saved login cookie.