One prompt, one research brief from YouTube, Reddit and X. But this tool isn't an MCP, and it doesn't go in Connectors

The repo is open source and MIT licensed, with over 86k stars on GitHub. If you'd rather go straight there, the button opens the English docs. The places where the install quietly goes wrong, the unrelated same-name package on PyPI and the Windows commands, are covered separately in 106.

What's extra below: where the tool actually lives, which sources are ready and which want a login, and the corrected version of the prompt that pulls one brief out of three platforms at once.

Most guides for this tool say "install the MCP, add it in Claude under Customize, Connectors, and Claude gains new tools like search Reddit." The repo's own docs say something else: Agent Reach is a CLI and a skill. There's no MCP address to add and no Connectors setting. What it does is install tools like yt-dlp, gh and twitter-cli on your machine; the reading is done by the agent itself, running those commands in a terminal.

The upshot: this tool doesn't work in the claude.ai web chat or the ChatGPT app, because neither can run commands on your computer. It works in agents with terminal access, like Claude Code, Cursor, Windsurf and OpenClaw.

Paste the line below into Claude Code as-is. This is the repo's current recommended wording; the agent first checks what's missing and asks you before installing anything on the system.

Safely check and install Agent Reach: https://raw.githubusercontent.com/Panniantong/agent-reach/main/docs/install.md

If your agent supports skills there's a second route: you install the skill first, and the agent fills in the CLI itself if it's missing.

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

Guides list three sources as "free and reliable": YouTube, Reddit, GitHub. True for the first two, not for Reddit. Reddit closed anonymous access and its official API is approval-only; the repo's own table says there's "no zero-config path" for Reddit. So if your research prompt mentions Reddit, you need to log in first.

Reddit has two routes. On a desktop, OpenCLI rides the Reddit session already open in Chrome. On a server, rdt-cli gets installed and you log in with rdt login. You can have the agent do either by telling it "set up Reddit."

This is where the work happens. Research that used to need five tabs and three subscriptions, in one prompt. I added two lines to the original: say which source each point came from, and if a source won't open, say so instead of filling the gap. The second one matters, because when a source comes back empty the model sometimes fills that part from general knowledge and you don't notice.

Using the Agent Reach tools:
1. Pull the transcript of this YouTube video: [URL]. Summarize the 5 key takeaways.
2. Search Reddit for what people actually complain about with [product/topic] and group the complaints into themes.
3. Find 3 recent X threads on [topic]; tell me the consensus and the contrarian take.

Rules:
- Next to each point, say which source it came from.
- If a source won't open or comes back empty, say so plainly; don't fill that part from general knowledge.

Give me one combined brief.

If you haven't set up Reddit or X yet, drop items 2 and 3 and start with YouTube, GitHub and web pages; that's the part that works on day one with no login at all.

The gain is speed. Competitor research, pulling content ideas out of real complaints, grabbing a video's transcript and turning it into a post: normally an afternoon of copy-paste, here one prompt.

X is its own case: the platform has tightened access, and free tools can stall now and then. If it stalls, tell the agent to skip that part and continue with the other sources; the "say so if it won't open" rule in the prompt already makes this visible.