| name | agent-reach |
| description | Use public web, GitHub, social/community, video, RSS, and career search surfaces from the CLI/MCP when local repo evidence is not enough. Routes to public-only research tools while avoiding internal/company URLs and secret-bearing pages. |
| license | unknown |
| compatibility | Requires public network access; optional gh, mcporter, yt-dlp, and platform-specific CLIs unlock additional channels. |
| metadata | {"dotfiles.provenance.upstream":"https://github.com/Panniantong/agent-reach","dotfiles.provenance.mode":"adapted","dotfiles.provenance.local-changes":"Rewritten as a public-only routing skill with company URL guards, local tool fallbacks, and non-mutating health checks."} |
Agent Reach
Use this skill when the task needs current public internet evidence beyond local repository files: web search, page reading, GitHub discovery, social/community signal, video subtitles, RSS, or public career/company pages.
Provenance: Upstream: https://github.com/Panniantong/agent-reach; License: unknown; Mode: adapted; Local changes: public-only routing, company URL guards, local fallbacks, and non-mutating health checks.
Safety boundary
- Public sources only by default. Do not send company, client, localhost, private repository, authenticated, or secret-bearing URLs through hosted readers/search tools such as Exa, Jina Reader, or third-party scrapers.
- Work-scope (NAVER): for
*.navercorp.com, internal, or authenticated URLs, never use Jina (r.jina.ai) or any hosted reader/scraper. Fetch locally with defuddle or agent-browser open <URL> --profile "Default" --allowed-domains <host> instead. See the routing table in company/configs/AGENTS-company.md.
- Prefer official documentation, upstream repositories, release notes, and source pages before social summaries.
- Treat CLI/social results as leads until opened and checked. Report unavailable tools instead of claiming a channel worked.
- Use
/tmp for temporary fetched content. Use ~/.agent-reach/ only for persistent tool state.
- Do not bulk scrape, bypass access controls, or collect personal data beyond what the user explicitly asks and is allowed to process.
Route by task
| Need | Read next | Typical tools |
|---|
| General public web search or code-aware search | references/search.md | mcporter Exa tools, search snippets |
| Read a public URL, article, RSS feed, or WeChat article | references/web.md | Jina Reader, web-reader MCP, Exa crawl, RSS parser |
| GitHub repositories, PRs, issues, releases, CI logs | references/dev.md | gh search, gh repo, gh pr, gh run, gh api |
| Reddit, V2EX, X/Twitter, Xiaohongshu, Douyin, Weibo, Bilibili signals | references/social.md | platform CLIs/MCPs when installed |
| YouTube/Bilibili/podcast metadata, subtitles, or comments | references/video.md | yt-dlp, bili, transcript helpers |
| Public LinkedIn/job/company lookup | references/career.md | LinkedIn MCP if authenticated; public fallback only |
Quick checks
command -v agent-reach >/dev/null && agent-reach --version
command -v gh >/dev/null && gh auth status
command -v yt-dlp >/dev/null && yt-dlp --version
Do not use agent-reach doctor as a read-only probe. The upstream command may
re-register its generic skill under ~/.agents/skills and ~/.claude/skills,
overwriting or duplicating this adapted local skill. Run it only during explicit
upstream setup; afterward reinstall local skills and compare the active trees.
If a command is missing, use the nearest available public alternative and state the gap.
