agent-reach

Agent Reach

Agent Reach is an open-source CLI that gives an AI agent practical access to the live internet. It wraps a fleet of free / open-source backends (Jina Reader, twitter-cli, yt-dlp, OpenCLI, rdt-cli, bili-cli, gh, feedparser, Exa via MCP) behind one tool with a primary + fallback architecture: when one backend fails, it transparently switches to another, so you keep reading instead of debugging auth.

Use this skill to install Agent Reach on the user's machine, diagnose which channels are live, and then read and search content from platforms a plain WebFetch/curl can't reach.

$ARGUMENTS — if the user passed a URL or a search query, that's the target to fetch or search once the tool is ready.

When to use this

Reach for Agent Reach when the content lives behind something a simple HTTP fetch can't handle:

Login-walled or JS-heavy pages — tweets/threads, Xiaohongshu posts, LinkedIn, Reddit threads.
Media transcripts — YouTube, Bilibili, podcasts (Xiaoyuzhou).
Structured feeds and code hosts — RSS/Atom, GitHub issues/PRs/repos.
Open-web research — global search via Exa when you need to find sources, not just read a known URL.

If a plain WebFetch of a public static page already works, just use that — don't install a whole CLI for a one-off readable page.

Safety boundaries (read before installing)

Agent Reach modifies the user's environment, so honor these limits:

Stay inside ~/.agent-reach/. Config, tokens, and tool repos all live there (~/.agent-reach/tools/). Temp work goes to /tmp/. Do not clone repos or drop files into the user's project workspace.
Never sudo without explicit user approval. If a step needs elevated permissions, stop and ask first.
Credentials stay local. Cookies/tokens are written to ~/.agent-reach/config.yaml with 600 permissions and are never transmitted off-machine. Treat that file as secret — never print or commit it.
Account-ban risk on login channels. Twitter/X and Xiaohongshu can ban accounts used for automated reading. Tell the user to log in with a dedicated secondary account, never their primary.
Prefer a dry run first. Use --safe (preview, no changes) or --dry-run to show the user what will happen before mutating anything.

Install

Agent Reach is not vendored in this skill — it's installed from upstream. Detect the OS and pick an install path. Run these in the user's shell (the Bash tool); on Windows, run them under WSL or a POSIX shell since the upstream tool targets Unix-like environments.

Preferred — pipx (isolated):

pipx install https://github.com/Panniantong/agent-reach/archive/main.zip
agent-reach install --env=auto

Fallback — virtualenv (use if PEP 668 blocks a global install):

python3 -m venv ~/.agent-reach-venv
source ~/.agent-reach-venv/bin/activate
pip install https://github.com/Panniantong/agent-reach/archive/main.zip
agent-reach install --env=auto

agent-reach install detects local-vs-server environment, configures the search backend, and registers the per-channel capability docs the agent reads later. --env=auto lets it choose; pass --safe to preview without changes.

Optional channels

The default install brings up the no-login channels. Add login/extra channels explicitly:

# install specific channels
agent-reach install --env=auto --channels=opencli,twitter

# install everything
agent-reach install --env=auto --channels=all

Optional channels include opencli, twitter, xiaohongshu, reddit, bilibili, xueqiu, xiaoyuzhou, and linkedin.

Verify & configure

agent-reach doctor              # status of every channel (ready / needs login / error)
agent-reach watch               # health + update check
agent-reach configure proxy     # set an outbound proxy URL
agent-reach configure groq-key  # set a Groq API key (for backends that use it)
agent-reach update              # update an existing install
agent-reach uninstall [--keep-config] [--dry-run]

Always run agent-reach doctor after installing and report the result to the user — it's the source of truth for what's reachable right now and what still needs a login.

Channels at a glance

Platform	Login required
Web pages (Jina Reader)	No
YouTube (transcripts)	No
RSS / Atom	No
GitHub (`gh`)	No (richer with login)
Global search (Exa)	No
Bilibili	No
V2EX	No
Xueqiu	No (configurable)
Twitter / X	Yes (ban risk — use a burner)
Reddit	Yes
Xiaohongshu	Yes (ban risk — use a burner)
LinkedIn	Yes (configurable)
Xiaoyuzhou (podcasts)	Yes

Reading & searching

After install, fetch and search through the channels Agent Reach set up. The agent-facing reading workflow is driven by the capability docs that agent-reach install registers and that agent-reach doctor points to — read those for the exact, current invocation of each channel rather than guessing, because the upstream backends evolve.

In practice that means:

Run agent-reach doctor to confirm the channel for the target URL/query is ready (and log in first if it reports "needs login").
Follow the registered capability instructions for that channel to read the URL or run the search. For some channels you'll call the underlying tool the installer placed on PATH directly — e.g. yt-dlp for YouTube transcripts, gh for GitHub, bili for Bilibili, rdt for Reddit, opencli for Xiaohongshu, and mcporter/Exa for open-web search.
If a backend errors, let the primary→fallback design retry, or re-run doctor to see which backend degraded.

If the user gave a target in $ARGUMENTS, route it to the matching channel once doctor confirms it's ready, then return the extracted content.

Compatibility

Agent Reach works with any agent that can run shell commands — Claude Code, Cursor, Windsurf, OpenClaw, and others. The upstream backends target Unix-like environments; on Windows, drive it through WSL.

Credits

Agent Reach is created by Panniantong (@Neo_Reidlab, pnt01@foxmail.com) and released under the MIT License — https://github.com/Panniantong/agent-reach. All credit for the tool and its backend integrations belongs to its author and contributors. This skill only documents how to install and drive the upstream CLI; it does not vendor or modify it.

Skill packaged by Rinu (l3ad3r1) in collaboration with Claude (Anthropic).