| name | web-reach |
| description | Read or search a SPECIFIC internet platform the right way - YouTube (transcripts, not the page), Twitter/X, Reddit, Facebook/Instagram feeds, GitHub repos/issues, Bilibili, XiaoHongShu, LinkedIn, RSS, podcasts - by ROUTING to the best FREE backend instead of blindly scraping the SPA. Recognize the platform from the URL or the ask, pick the specialized tool (yt-dlp for YouTube subs, gh for GitHub, agent-reach for login/social), and only fall back to web_fetch markdown for a generic page. For "doc/lay/tom tat/tim tren youtube, twitter/X, reddit, facebook, instagram, github, bilibili, tiktok, linkedin, RSS...". (doc va tim noi dung theo tung nen tang - dung dung cong cu, khong cao SPA). |
| match | youtube\.com|youtu\.be|twitter\.com|reddit\.com|facebook\.com|instagram\.com|bilibili\.com|tiktok\.com|linkedin\.com|github\.com|xiaohongshu|\byoutube\b|\btwitter\b|\breddit\b|\bfacebook\b|\binstagram\b|\bbilibili\b|\btiktok\b|\blinkedin\b|\bgithub\b|\btweet|transcript|rss feed|\brss\b |
Skill: Web reach (route each platform to the right free backend)
Scraping a platform's web SPA is the WORST way to read it (see the web-reading skill for why it thrashes).
Almost every platform has a specialized free backend that returns clean, structured data directly - a
YouTube transcript, a tweet's text, a Reddit thread, a repo's README/issues. Identify the platform first,
route to its tool, and you skip the whole render/scrape/extract mess.
Rule 0 - identify the platform, then route (do NOT default to scraping)
| Platform / ask | Best FREE backend | Needs login? |
|---|
| YouTube (what does this video say) | yt-dlp -> subtitles/transcript (NOT scraping the page) | no |
| GitHub (repo / issues / PRs) | gh CLI (already here) or the API | no |
| RSS feed | feedparser (pip install feedparser) | no |
| Generic article / public page | web_fetch (returns markdown; set scrape_backend: "jina" for public JS pages) | no |
| Reddit / Twitter-X / Bilibili / XiaoHongShu / LinkedIn / podcasts | specialized CLI (below) | some |
| Facebook / Instagram FEED (logged-in) | browser session bridge (agent-reach/OpenCLI) OR the browser MCP + your session | yes |
The fast path: agent-reach if it's installed (covers ~15 platforms, free)
Agent-Reach is a router that installs + health-checks free
backends for ~15 platforms and reuses your logged-in browser session for social ones. If the user has it:
agent-reach doctor --json - shows which platform backends are live (never assume).- Then call the routed tool it points to (
yt-dlp, gh, twitter, bili, mcporter/Exa search, ...)
via bash, or agent-reach <platform> <url> per its docs.
It's not installed by default - pip install agent-reach (the user's choice; it pulls in per-platform CLIs).
Native routes that need NO agent-reach (use these first if present)
- YouTube:
yt-dlp --write-auto-subs --skip-download --sub-format vtt -o - "<url>" (or --get-...) for
the transcript; summarize that, don't open the video page.
- GitHub:
gh repo view, gh issue list, gh api ... - clean + authenticated already.
- Generic web:
web_fetch(url) -> markdown; large pages paginate (page:N).
Check availability honestly first: which yt-dlp gh / agent-reach doctor. If a backend is missing, say so
and offer the one-line install, don't silently fall back to fragile scraping.
Login-required feeds (Facebook feed, Twitter home, private IG) - read this first
- These need YOUR authenticated session. Two ways: (a) a browser-session bridge (agent-reach/OpenCLI's Chrome
extension + daemon) reuses your live login; (b) the browser MCP driving real Chrome that's already logged in.
- Before collecting a feed—especially a requested count such as 50/100—load the
web-reading skill. Its
capture-before-scroll + disk accumulator preserves rows that a virtualized feed unmounts and bounds context.
- Account risk + ToS: automating a logged-in social account can get it checkpointed or banned, and a
session bridge acts AS you. Warn the user, keep it minimal, prefer public/RSS/API routes, and never do it on
an account you can't afford to lose. This is the user's call - surface the risk, don't just do it.
Honesty
Report which backend you used and what you could/couldn't get (not installed, login wall, rate limit, ToS).
Never invent content to fill a gap.
