Want to try it? Head over to GitHub: jtsternberg/claude-plugins → hotline
Hotline is an agent plugin — a set of skills that give AI agents a shared communication protocol, so one agent session can call into another workspace’s session directly instead of duplicating what it already knows. It works from Codex and (probably) other agents too, though right now it can only dial out to Claude.
That’s different from the multi-agent tooling Claude ships with out of the box. Built in, an agent can see which other agents are currently active and ask one a question directly — which means dropping that question into whatever live session is already running there, context and all.
Hotline defaults the other way: a first call starts a brand-new session in the target workspace, and every call after that, from that same caller, picks up the conversation Hotline started for it there — a different caller dialing the same workspace gets its own separate session, not yours. You can still point it at somebody else’s session, but you have to hand it the session ID on purpose, and even then it forks that conversation by default — the callee inherits all the context, the original transcript stays untouched. That session ID does double duty, too: say “dial into that conversation” and it’s a handle Hotline forks; say “dial whatever workspace that session lives in” and it’s just an address, routed like any other target, no fork involved.
To understand why I built it, here’s the post I put up on LinkedIn about it, copied here as the canonical writeup:
I keep running into the same problem with AI workspaces.
One workspace gets really good at one thing.
The blog workspace knows how I write.
The API workspace knows the contracts, scripts, and quirks.
The frontend workspace knows the design system and component patterns.
At first, the temptation is to copy that “magic” everywhere.
Bring over the same skills.
Duplicate the same plugins.
Paste the same instruction files.
Try to make every workspace equally capable.
But the more I did that, the more it felt like I was watering down what made each one strong in the first place.
The specialist stopped being a specialist.
So I’ve been leaning into a different idea:
What if the answer isn’t duplication?
What if the answer is collaboration?
I built a Claude Code plugin called Hotline around that idea.
Instead of trying to make one workspace know everything, it lets one Claude instance call another workspace that already has the right context, tools, and muscle memory.
Almost like calling in the specialist.
“Hey docs workspace, can you draft the release notes?”
“API workspace, can you sanity-check this contract?”
“Design workspace, help me think through this component.”
It feels like a small shift, but I think it points to something bigger:
Maybe the future isn’t one giant super-agent.
Maybe it’s a network of specialists that know when to call each other.
Curious if others are seeing the same pattern emerge in their AI workflows.
Huge thanks to Sandhya Simhan for the nudge to share this more broadly 🙏
Plugin install instructions here: github.com/jtsternberg/claude-plugins/tree/main/plugins/hotline