Agent Reach: AI Agent Web Access via One CLI
Panniantong/Agent-Reach
Reach for Agent Reach if your coding agent keeps failing at 'go read this tweet / Reddit thread / YouTube video' and you'd rather not pay for six separate APIs — it installs and health-checks the readers for you, and `agent-reach doctor` tells you what's actually working. Skip it if you need the agent to act on pages (log in, submit forms, click) rather than just read them, or you only touch one platform and can wire that single tool yourself.
Why Your Agent Goes Blind Online
AI agents can write code and edit docs, but they go blind the moment you ask them to read something online. Per the README's own examples, Twitter's API costs money, Reddit 403s server IPs, XiaoHongShu needs a login, Bilibili's anti-bot blocks generic downloaders, and a raw web page comes back as HTML soup. Each platform has its own paywall, block, login, or cleanup step, so wiring them up one by one for every new agent is a repeated time sink.
A Capability Layer, Not a Reader Itself
Agent Reach is an open-source (MIT) Python capability layer that gives an AI agent web access. It doesn't do the reading itself — it selects, installs, health-checks and routes the upstream tools an agent uses to read and search the internet, then registers a SKILL.md so the agent knows which tool to call. Coverage includes web pages, YouTube, RSS, semantic search, GitHub, Twitter, Reddit, Bilibili, XiaoHongShu, Facebook, Instagram, LinkedIn and a few China-focused sites like V2EX and Xueqiu.
Free API Access in a Pay-Per-Platform World
Giving agents real capabilities is the current center of gravity in AI tooling, and reading the open internet is a top-ranked wanted capability of those capabilities. Agent Reach packages that as a one-command install with no per-platform API bills, which is an easy sell. Its README carries a Trendshift badge claiming a #1 GitHub trending spot, and it sits at roughly 54.8k stars on GitHub.
Zero-Config Channels and Fallback Routing
- ✓One-line install: you paste a URL to your agent, which runs `pip install` and sets up Node.js, gh CLI, mcporter and the Exa search MCP
- ✓Multi-backend routing per platform (a preferred source plus fallbacks); switching a source is reordering a list, not rewriting code
- ✓`agent-reach doctor` reports each channel's live status and which backend it's currently using
- ✓Zero-config channels out of the box: web pages (Jina Reader), YouTube (yt-dlp), RSS (feedparser), Exa search, public GitHub and V2EX
- ✓Login-gated channels on request: Twitter, Reddit, Facebook, Instagram, XiaoHongShu and LinkedIn via browser login state
- ✓Local-only credential storage (config file at perms 600) plus `--safe` and `--dry-run` install modes
Paste a URL, Let the Agent Install It
Agent Reach installs by paste-and-go: you copy a one-line instruction pointing your agent at the project's install.md (a raw GitHub URL) and the agent does the rest — `pip install agent-reach`, plus Node.js, gh CLI, mcporter and the Exa search MCP. By default it activates six zero-config channels and asks before adding login-gated ones. `--safe` skips automatic system changes and just lists what's needed; `--dry-run` previews every action without touching anything. Note the agent must be allowed to run shell commands (OpenClaw users need the coding tools profile).
Talking to Agent Reach in Plain Language
You drive it in plain language through your agent: 'read this link', 'search Bilibili for AI tutorials', 'what does this YouTube video say'. The agent reads the installed SKILL.md and calls the right upstream tool — Jina Reader, yt-dlp, gh, Exa, and so on — with no wrapper layer in between. Run `agent-reach doctor` to see each channel's status and current backend. For login platforms, tell the agent 'help me set up X' and it walks you through the credentials or browser login state.
From YouTube Transcripts to Reddit Threads
- •Let an agent pull a YouTube transcript and summarize a tutorial
- •Search Reddit or Twitter for how people rate a product or where a bug shows up
- •Read an arbitrary web page as clean text instead of raw HTML tags
- •Run a semantic web search from the agent without paying for a search API
- •Read a public GitHub repo's issues during an agent-driven investigation
No API Bills for the Core Channels
- ✓No paid API keys for the core channels; the README says the only possible cost is an optional ~$1/month server proxy, which local machines don't need
- ✓One install command wires up many platforms at once instead of configuring each tool by hand
- ✓Per-platform fallback routing plus `agent-reach doctor` makes breakage visible and swappable rather than a silent failure
- ✓Credentials stay on your machine (config file perms 600), with `--safe` and `--dry-run` install modes for cautious setups
- ✓Works with any shell-capable agent — Claude Code, Cursor, Windsurf, OpenClaw
Where Platforms Fight Back
- △Everything rides on third-party platforms' terms and anti-scraping; backends break when platforms change — the README itself notes yt-dlp got blocked by Bilibili's anti-bot in 2026-06 and a batch of single-platform CLIs went EOL in 2026-03
- △Cookie/login-state platforms (Twitter, Reddit, XiaoHongShu, etc.) carry a real account-ban risk — the README explicitly tells you to use a burner account, not your main one
- △No zero-config path for Reddit, Facebook, Instagram or XiaoHongShu; they need browser login state via desktop OpenCLI
- △It only reads and searches — it doesn't operate pages (forms, logins, clicks), and the README points you to a separate browser-automation tool for that
- △Docs and platform coverage lean heavily toward Chinese platforms (Bilibili, XiaoHongShu, V2EX, Xueqiu); the primary README is in Chinese with English linked separately
Firecrawl, Skyvern, and Plain yt-dlp
Best for Chinese-Platform Coverage
People running a shell-capable coding agent who keep hitting the wall where the agent can't read a tweet, a Reddit thread, a YouTube video, or a web page. It's especially handy if you work with Chinese platforms like Bilibili and XiaoHongShu, which most Western tools ignore. Skip it if you need the agent to act on pages — submit forms, log in, click through flows — or you only touch one source and can wire that single tool yourself.
Agent Reach FAQ
Agent Reach is an open-source Python CLI that gives an AI agent web access. It installs, health-checks and routes the tools an agent uses to read and search sites like Twitter, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu, so the agent can fetch and search internet content without you configuring each platform by hand.
Yes for the core channels — the tools are open source and the README says the APIs used are free. The only possible cost it mentions is an optional server proxy at roughly $1/month, which local machines don't need.
Per the README: web pages, YouTube, RSS, semantic search and GitHub work with zero config, while Twitter, Reddit, Facebook, Instagram, XiaoHongShu, Bilibili and LinkedIn need login state or a quick setup. It also covers China-focused sites like V2EX and Xueqiu.
Yes. The README says it works with any agent that can run shell commands, naming Claude Code, Cursor, Windsurf and OpenClaw. OpenClaw users have to enable the coding tools profile so the agent can execute commands.
The README warns that platforms using cookie login (Twitter, XiaoHongShu, Reddit and similar) can detect scripted access and ban accounts, so it recommends a dedicated burner account. Credentials are stored locally with file permissions set to 600.