Skip to content

Back to the catalog

https://github.com/Panniantong/Agent-Reach
github.comrepo

https://github.com/Panniantong/Agent-Reach

Agent-Reach is a CLI and agent-"skill" layer that equips coding agents (Claude Code, Cursor, OpenClaw) with ready-made web-reading tool use: fetching arbitrary pages via Jina Reader, YouTube transcripts via yt-dlp, GitHub repos via gh CLI, Bilibili search, RSS via feedparser, and semantic web search via Exa through mcporter, with login-gated platforms (Twitter/X, Reddit, Xiaohongshu, Facebook, Instagram) added on request via cookies or OpenCLI browser sessions. Each platform module probes an ordered list of candidate backends and falls back automatically when one breaks (e.g. yt-dlp being blocked by Bilibili's anti-scraping), and an `agent-reach doctor` command reports which backend each channel is currently using. It positions itself as a "capability layer" that selects, installs and health-checks the right upstream tool for an agent rather than wrapping the reads itself.

TL;DR A skill/CLI layer giving coding agents one-command access to read Twitter, Reddit, YouTube, GitHub, Bilibili, Xiaohongshu and more, with self-selecting, self-healing backends.
Document

How it works

Agent Reach selects, installs and health-checks tools but does not sit in the read path. The agent reads a SKILL.md and then calls upstream tools directly. Each channels/*.py file holds an ordered list of backends. "Switching backend" means reordering that list, not rewriting code. Probes actually exercise each backend, not just check that the command exists. The first backend that works fully is selected, and broken ones get a fix prescription.

flowchart LR
  A[doctor probes next backend] --> B{fully works?}
  B -- yes --> C[selected for channel]
  B -- no --> D{more candidates?}
  D -- yes --> A
  D -- no --> E[report fix prescription]

Current backend order (the page calls these "current picks", re-checked periodically on real machines):

ChannelBackends, in order
twittertwitter-cli ▸ OpenCLI ▸ bird
bilibilibili-cli ▸ OpenCLI ▸ search API (yt-dlp retired, blocked by Bilibili)
redditOpenCLI ▸ rdt-cli (no zero-config path, login required)
facebook, instagramOpenCLI only (desktop browser session)
xiaohongshuOpenCLI ▸ xiaohongshu-mcp ▸ xhs-cli
linkedinmcp-server-linkedin ▸ Jina Reader (not in the summary)
web / youtube / github / rss / searchJina Reader / yt-dlp / gh / feedparser / Exa via mcporter

The page says a batch of single-platform CLIs stopped being maintained in March 2026, and the project re-routed around them.

How to use it

You install it by pasting this prompt into the agent:

帮我安装 Agent Reach:https://raw.githubusercontent.com/Panniantong/agent-reach/main/docs/install.md

Updating uses docs/update.md the same way.

Install details:

  • Default install changes nothing. By default agent-reach install only checks the environment. Installing dependencies, wiring Exa over MCP, and writing SKILL.md into the agent's skills directory all require agent-reach install --system.
  • Install from this repo, not PyPI. The PyPI package with the same name is not this project. The CLI bundles yt-dlp and feedparser. It checks for Node.js, gh CLI and mcporter.
  • Channels: six zero-config channels are active by default. Login-gated channels are added only when you name them.
  • OpenClaw: the default messaging profile blocks shell exec. Run openclaw config set tools.profile "coding", then openclaw gateway restart, then start a new conversation.
  • Proxy: none is needed on a local machine. On a server you need one (stated cost about $1/month).

Twitter setup:

  • Cookies are accepted only as a manual Cookie-Editor export.
  • The saved cookie is used only by doctor to check that configuration is complete.
  • To run upstream twitter commands, you must still export TWITTER_AUTH_TOKEN and TWITTER_CT0 in the process environment.

Xiaohongshu setup:

  • Agent Reach never logs in for you and never reads browser cookies.
  • OpenCLI only reuses a Chrome session you already control.
  • configure xhs-cookies does not inject cookies into OpenCLI.

Uninstall:

agent-reach uninstall
agent-reach uninstall --dry-run
agent-reach uninstall --keep-config
pip uninstall agent-reach

A full uninstall removes ~/.agent-reach/ (including all tokens and cookies), the skill files, and the mcporter MCP config. --dry-run only previews, and --keep-config removes the skill files but keeps tokens.

Limits

  • Account bans: the page warns that cookie-based access (Twitter, Xiaohongshu and others) risks platform detection and bans. Use dedicated throwaway accounts, which also limits the damage if credentials leak.
  • Reddit, Facebook and Instagram have no anonymous path.
  • Backends will change: the backend choices are explicitly provisional, and the project's reliability depends on one maintainer tracking anti-scraping changes.
  • Credentials: cookies are stored locally only.
  • No licence: the page does not state one, so check the repo before you adopt it.

Organisations named here

  • GitHub

Technologies named here

  • Agent Reach
  • OpenCLI
  • twitter-cli
  • bird
  • yt-dlp
  • gh CLI
  • bili-cli
  • rdt-cli
  • xiaohongshu-mcp
  • xhs-cli
  • mcp-server-linkedin
  • Jina Reader
  • Exa
  • mcporter
  • feedparser
  • Cookie-Editor
  • SKILL.md
  • Agent Skills Hub
  • OpenClaw
  • Model Context Protocol
  • Claude Code
  • Cursor
  • Node.js
  • GitHub
  • Python

Techniques named here

  • Sequential backend probing with fallback chain
  • Cookie-based authentication