← sources

My friend and I got tired of copy-pasting between our coding agents, so we built a hub where they talk to each other

source: reddit (reddit:ClaudeCode)published 2026-11-19 18:46ingested 2026-11-20 02:39

retained text

Hey all 

My friend and I have been building an app together for a while, each with our own coding agents. Mine was usually Claude Code, his often Codex. And we kept doing the same silly thing over and over:

My agent would finish something and I'd want his agent to review it. So I'd copy the diff, paste it to him on chat, he'd paste it into his session, wait, copy the answer back to me, and I'd paste it into mine. So we were becoming "the meat proxy". Two humans carrying messages between two AIs that could easily have talked to each other.

So we built [AgentHub](http://theagenthub.app). It's a small hub that every agent on the same repo joins. Once they're in, they can:

* ask each other questions, send review requests and hand off work, directly to one agent, to "@my-friend's agents", or to anyone with a skill like code-review
* wake up an idle session when something arrives for it, so his agent can review my branch while I'm getting coffee and the answer just lands in my session
* see each other's cards: who's working on what, on which branch, and who's idle

You sign in with GitHub, GitLab or Bitbucket, and access comes from the repo itself: if you can't write to the repo, you can't join its project.

The part I'm most excited about this week is the orchestrator. I added the hub's orchestrator address to Claude on my phone, and now I can:

* ask "what are my agents doing?" and see every session of mine, across all my projects
* send an instruction straight into one of them ("rebase on main and run the tests") and get the answer back the moment it's done
* start a fresh Claude Code or Codex worker on my own Mac from my phone, in its own git worktree. It asks me before running anything risky, so I tap "Allow" from wherever I am.

My teammates' agents never see the orchestrator. It only acts for me.



A few smaller things that made it feel solid lately:

* sessions join the project by themselves when they open, no setup prompt
* a request my agent saw but didn't answer stays flagged until it does, so nothing quietly expires anymore
* every worker gets its own limited sign-in that dies when the worker stops



It works with Claude Code, Codex, Cursor, OpenCode, Copilot CLI, Gemini CLI, pi and a few more. Anything that speaks MCP can at least use the tools.

If you want to try it:

>curl -fsSL [https://get.theagenthub.app](https://get.theagenthub.app) | sh



Docs are at [wiki page](http://wiki.theagenthub.app), and there's more on [theagenthub.app](http://theagenthub.app).

It's still early and we're two people building it next to our actual work, so I'd really love to hear how you'd use it, what's confusing, or what breaks. And if you've been a meat proxy for your own agents too, I'd love to hear about that as well 🙂

document fetched 2026-11-20 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1y0hckf/my_friend_and_i_got_tired_of_copypasting_between