UtilKits

Agent Reach

Read-only CLI that gives AI agents a single JSON interface to GitHub, YouTube, Bilibili, V2EX, Reddit, X and other platforms that built-in web search handles badly.

Agent Reach
Summarize with:ChatGPTPerplexityGrokClaude

Agent Reach: give your coding agent the parts of the internet it cannot open

Claude Code, Codex and similar CLI agents ship with web search and page fetching. That covers public articles well. It fails on pages behind a login, on what is said inside a video, and on Chinese platforms that general search barely indexes. Ask an agent to read a Reddit thread, pull a YouTube transcript or check a V2EX discussion and you usually get a blocked page or a guess.

Agent Reach fills that gap. It is an MIT-licensed Python CLI by Panniantong (108k+ GitHub stars, version 1.6.1 when I tested it). It installs and wires up the upstream tools for each platform, then exposes them all as agent-reach <channel> <action>. Every command prints JSON, so the agent never has to scrape or parse HTML.

This kit covers which platforms work out of the box, how a request is routed, and what a real run looks like on a Windows machine.

1. Channels and what each one needs

Agent Reach groups platforms into channels with three levels: default works right after install, optional turns on when you ask, and experimental can break whenever the platform tightens anti-bot rules.

Channel Level Actions Needs
github default search repos, read repo and README, read issues and PRs with comments nothing
youtube default video info, search, subtitles, speech-to-text nothing; speech-to-text needs a Groq or OpenAI key
web default web search, page reading nothing
v2ex default hot topics, nodes, topics with replies, users nothing
bilibili default search, video info, trending, subtitles nothing; subtitles need a QR login
reddit, linkedin, xiaohongshu, boss, xueqiu optional search and read posts, jobs or quotes a login for most actions
xiaoyuzhou optional podcast transcription Groq key
twitter, facebook, instagram experimental search, profiles, posts login

Run agent-reach channels to list every channel with its actions and parameters. The agent can read that output and pick the right command without a manual.

2. How a request is routed

Each action on each platform has an ordered list of tools. Agent Reach tries the first, and if it fails it moves to the next. Reading a Bilibili video tries bili-cli, then Bilibili's public API. Reading a web page tries Jina Reader, then Exa.

Agent Reach workflow from agent through the CLI and router to a primary or fallback tool, the platform and a JSON result
Agent Reach workflow from agent through the CLI and router to a primary or fallback tool, the platform and a JSON result
  • Agent to CLI. The agent runs one shell command, for example agent-reach github read owner/repo.
  • Router. It picks the channel from the command, or from the URL when you use agent-reach read <url>.
  • Primary, then fallback. A broken upstream tool only reorders that list. The way the agent calls Agent Reach does not change.
  • JSON back. Each result has ok, channel, action, backend and a list of items, so you can see which tool answered.

A real run

I ran the commands below in PowerShell on Windows with the default channels. The screenshot shows the full, unedited output. doctor reads local state only and makes no network calls, so most channels show unverified. github shows ready because I had run agent-reach check github first. Run agent-reach check <channel> to test one live.

Terminal running agent-reach doctor, github search and youtube search with full JSON output

Other commands that worked on the same machine: github search, github read, web read, youtube subtitle and bilibili hot. The youtube subtitle call returned the full caption text of a Claude Code tutorial (about 8,300 words) in one response, something a normal fetch tool cannot do.

3. Safety and access

Agent Reach is read-only. It searches and reads, and it does not post, comment, like or send messages.

  • Logins. It never logs in for you and never asks for a password. For channels that need one, it uses its own background browser and copies the login for that one site from your Chrome. Your Chrome profile stays unchanged and the password store is not read.
  • Use a secondary account. The copied login is usually your main account, and platforms may restrict accounts that automate access. The README recommends a spare account.
  • One folder. Tools, logins and settings live under ~/.agent-reach/, so removal is one command: agent-reach uninstall --dry-run lists what would go.
  • Local browser only. The dedicated browser accepts connections from your own machine.

[!IMPORTANT] The agent-reach package on PyPI is not this project. Install only from the GitHub URL in install.md.

4. Install and first commands

Let your agent install it

The README's supported path is to hand your agent the install guide. Paste this into Claude Code or any agent that can run commands:

Install Agent Reach: https://raw.githubusercontent.com/Panniantong/Agent-Reach/main/docs/install.md

The agent installs uv if it is missing, sets up the default channels and tells you what works. Read install.md once before you approve the commands.

Check the install

agent-reach doctor --text
agent-reach check github

First commands

agent-reach channels
agent-reach github search "agent reach" --limit 2
agent-reach youtube subtitle "https://www.youtube.com/watch?v=<id>" --lang en
agent-reach read https://www.v2ex.com/t/1253999

Add and remove channels

Tell your agent "add Reddit" or run agent-reach add reddit. Then run agent-reach login reddit and sign in yourself in the window that opens. agent-reach remove reddit clears the channel, its login and its private tools.

Keep it current

agent-reach update --check
agent-reach update
|Share:

Links