agent reach

About

Agent Reach is an open-source capability layer for giving AI agents access to information across web pages, search engines, social networks, video platforms, GitHub, RSS feeds, and other online sources. The project describes itself as a selector, installer, health checker, and router rather than another scraping wrapper: after setup, the agent normally calls the selected upstream tool directly.

This distinction is important for directory classification. Agent Reach is not itself an MCP server. Some routes use MCP infrastructure, including Exa through mcporter and optional MCP-based backends for platforms such as XiaoHongShu and LinkedIn, while other routes use tools including Jina Reader, OpenCLI, twitter-cli, rdt-cli, bili-cli, yt-dlp, GitHub CLI, and feedparser.

The repository's current Python metadata declares version 1.5.0, Python 3.10 or newer, and the MIT license.

Key Features

• Multi-backend routing with an ordered primary and fallback list for supported channels.
• agent-reach doctor checks whether platform backends actually work and reports the active route and remediation when a backend fails.
• Web page reading through Jina Reader.
• Web semantic search through Exa via mcporter.
• GitHub repository and code access through GitHub CLI.
• YouTube metadata and transcript workflows through yt-dlp.
• Bilibili access through bili-cli with OpenCLI as an additional route for some capabilities.
• Twitter/X, Reddit, Facebook, Instagram, and XiaoHongShu support through combinations of dedicated CLIs, OpenCLI, cookies, and optional MCP backends.
• RSS and Atom parsing through feedparser.
• Agent Skill support so compatible coding agents can learn the routing rules and invoke the appropriate upstream command.
• Safe-by-default installation mode that checks the environment without installing system dependencies or writing configuration unless system changes are explicitly approved.

Use Cases

Agent Reach is designed for agents that need to research across sources that are difficult to cover with one generic web-search API. Examples include reading a URL, searching Twitter/X discussions, inspecting Reddit conversations, retrieving YouTube subtitles, searching GitHub, reading XiaoHongShu or Bilibili content, querying technical communities such as V2EX, following RSS feeds, and performing broader semantic web search.

It is particularly useful when a team wants one maintained routing layer instead of independently selecting, installing, authenticating, and troubleshooting a different tool for every website.

Installation and Requirements

The current installation guide recommends installing the project directly from its GitHub archive with pipx:

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

The default agent-reach install --env=auto behavior is a read-only environment and dependency check. System-level installation and configuration require explicit approval:

agent-reach install --env=auto --system

A virtual environment plus pip is also documented as an alternative. The project currently warns users not to install the unrelated same-name package from PyPI and instead points installation at the GitHub repository.

Agents that support the Agent Skills format can also install the skill with:

npx skills add Panniantong/Agent-Reach@agent-reach

The installation process may check or configure dependencies such as Node.js, GitHub CLI, mcporter, yt-dlp, and platform-specific upstream tools. Several basic channels work without account configuration, while optional platforms can require cookies, an existing logged-in browser session, a free API key, or other platform-specific authentication.

Compatibility

The official documentation explicitly describes compatibility with Claude Code, OpenClaw, Cursor, and Windsurf, and more generally with AI agents that can execute shell commands. Agent Reach can additionally be installed as an Agent Skill in Claude Code, OpenClaw, and other environments supporting the Skills format.

This compatibility should not be confused with MCP-client compatibility. Agent Reach does not expose one universal MCP server endpoint. Instead, it can install and route to MCP and non-MCP backends depending on the platform being accessed.

Limitations

Some supported websites require a logged-in browser session or exported cookies, so availability depends on the user's local environment and the target platform's current restrictions. The project warns that scripted access to cookie-authenticated services such as Twitter/X and XiaoHongShu can create account-restriction or suspension risk and recommends using dedicated accounts rather than primary accounts.

Several backends are environment-dependent. OpenCLI routes that reuse a logged-in Chrome session are primarily desktop-oriented, while server deployments may require different fallback tools or proxies.

Agent Reach depends on upstream projects and websites whose behavior can change. Its routing architecture is specifically intended to replace or reorder those backends when an access method stops working, so agent-reach doctor is more authoritative for a particular installation than assuming one backend will remain permanent.

Sources

• Official repository: https://github.com/Panniantong/Agent-Reach
• Installation guide: https://github.com/Panniantong/Agent-Reach/blob/main/docs/install.md
• English documentation: https://github.com/Panniantong/Agent-Reach/blob/main/docs/README_en.md
• Agent Skill: https://github.com/Panniantong/Agent-Reach/blob/main/agent_reach/skill/SKILL.md
• Python project metadata: https://github.com/Panniantong/Agent-Reach/blob/main/pyproject.toml
• Releases: https://github.com/Panniantong/Agent-Reach/releases