📡 Agent Reach — MCP Fork

13 Internet Platforms as MCP Tools for Your AI Agent

Quick Start · MCP Server · Origin · What's Different · Platforms

---

Origin

This is a fork of Panniantong/Agent-Reach (v1.5.0), an excellent project that gives AI agents read/search access to 13+ internet platforms.

The original project focuses on CLI + skill-based access for coding agents. This fork adds a production-grade MCP (Model Context Protocol) server as a first-class interface, making all 13 platforms available as MCP tools that any MCP-compatible host (Claude Desktop, Cursor, Windsurf, etc.) can call directly.

What's Different

Area	Original (Panniantong)	This Fork
Primary interface	CLI (agent-reach read/search/doctor)	MCP server (8 tools) + CLI
MCP server	Basic proof-of-concept	Production-grade, 8 tools, 70+ tests
Web search	DuckDuckGo via Jina Reader (403 on servers)	Exa via mcporter (works from any IP)
Stock quotes	Xueqiu-only (needs browser cookies)	yfinance → Yahoo Finance v8 fallback (no auth needed)
Subprocess calls	16 ad-hoc blocks	Single _run_cli() helper (consistent error handling)
Tool descriptions	Minimal	Verb-first, return format, negative instructions (per MCP best practices)
Testing	Subprocess-based	70 in-memory FastMCP tests (deterministic, fast)
V2EX doctor	Always reported "ok"	Honest probing, returns "warn" when blocked
Error messages	Generic	Actionable: suggests alternatives when a platform is blocked

All upstream credit to @Neo_Reidlab and the Panniantong team for building the channel architecture and CLI tooling.

---

Why Agent Reach?

AI agents can already write code, edit documents, and manage projects — but ask them to find something online, and they're lost:

• 📺 "Check what this YouTube tutorial says" → Can't, can't get captions
• 🐦 "Search what people are saying about this product on Twitter" → Can't, Twitter API requires payment
• 📖 "Go check Reddit if anyone had the same bug" → 403 blocked, server IP rejected
• 📕 "Check this product's reviews on XiaoHongShu" → Can't open, must log in to view
• 🔍 "Search online for the latest LLM framework comparison" → No good search tool, either paid or low quality
• 🌐 "Check what this webpage says" → Gets back a mess of HTML tags, unreadable
• 📡 "Subscribe to these RSS feeds" → Need to install libraries and write code

Agent Reach turns this into one sentence: give your Agent the install URL, and minutes later it can read Twitter, search Reddit, watch YouTube, and browse XiaoHongShu.

---

Supported Platforms

Platform	Available Via	Auth Needed
🌐 Web	Jina Reader / Direct HTTP	No
📺 YouTube	yt-dlp caption extraction	No
📡 RSS	feedparser	No
🔍 Web Search	Exa via mcporter	No (free tier)
📦 GitHub	gh CLI	Optional (for private repos)
🐦 Twitter/X	twitter-cli	Cookies
📺 Bilibili	bili-cli	No
📖 Reddit	OpenCLI / rdt-cli	Cookies
📕 XiaoHongShu	xhs-cli / OpenCLI	Cookies
💼 LinkedIn	linkedin-scraper-mcp / Jina Reader	Optional
💻 V2EX	Web scraping (API often blocked on servers)	No
📈 Xueqiu	yfinance (US stocks) / Xueqiu API (Chinese stocks)	Cookies for Chinese stocks
🎙️ Xiaoyuzhou	RSS + Whisper	Groq API key

---

MCP Server

This is the main feature of this fork. Run the MCP server and your AI host gets 8 tools:

agent-reach-mcp

8 MCP Tools

Tool	What It Does	Best For
doctor	Check all 13 platforms' status	Diagnosing setup issues
read_url	Read any web page as clean text	Articles, docs, RSS feeds
search	Search by keyword across platforms	Finding content by topic
trending	Hot/trending content	What's popular right now
stock_quote	Real-time stock prices	AAPL, TSLA, etc. (US stocks work without setup)
get_details	Deep-dive by ID	V2EX topics, Bilibili videos, GitHub repos
transcribe	YouTube audio → text transcript	Video summaries
install	Configure a platform or API key	Setting up auth

Register with Claude Desktop

Add to your claude.json:

{
  "mcpServers": {
    "agent-reach": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/agent-reach-mcp",
        "python", "-m", "agent_reach.integrations.mcp_server"
      ]
    }
  }
}

Register with OpenCode

Add to your opencode.json:

{
  "mcpServers": {
    "agent-reach": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/agent-reach-mcp",
        "python", "-m", "agent_reach.integrations.mcp_server"
      ]
    }
  }
}

---

Quick Start

# Clone
git clone https://github.com/M0-AR/agent-reach-mcp.git
cd agent-reach-mcp

# Install
pip install -e .

# Check status
agent-reach doctor

# Run MCP server (for AI hosts)
agent-reach-mcp

---

Design Philosophy

This project is a capability layer, not a wrapper. It sits one level above any concrete implementation — responsible for selection, installation, diagnostics, and routing, not for the underlying reading itself. Reading is done by the AI agent calling upstream tools directly.

Multi-Backend Routing

Each platform has an ordered list of backends:

channels/
├── web.py          → Jina Reader ▸ Direct HTTP
├── twitter.py      → twitter-cli ▸ OpenCLI ▸ bird
├── youtube.py      → yt-dlp
├── bilibili.py     → bili-cli ▸ OpenCLI
├── xueqiu.py       → Xueqiu API ▸ yfinance ▸ Yahoo Finance v8
├── v2ex.py         → API 2.0 ▸ Legacy API ▸ Web scraping
├── rss.py          → feedparser
├── exa_search.py   → Exa via mcporter
└── ...

When one backend fails, the next is tried. agent-reach doctor shows which backend is active for each platform.

---

Testing

# All tests
pytest tests/ -v

# MCP server tests only
pytest tests/test_mcp_server.py -v

# With coverage
pip install pytest-cov
pytest tests/ --cov=agent_reach

---

Contributing

PRs are welcome. This is a fork focused on making the MCP server production-ready.

• Found a bug? Open an Issue
• Want a new channel? Fork the original Panniantong/Agent-Reach and contribute there, then we can merge upstream changes here

Development

pip install -e ".[dev]"
pytest tests/ -v
ruff check .

---

License

MIT — see LICENSE.

Original work Copyright (c) 2025 Agent Eyes (Panniantong/Agent-Reach). Fork modifications Copyright (c) 2026 M0-AR.