Any rider can drive — the driver of a bus is whichever session created it (and can hand the wheel off with /buses:transfer-driver). Claude just happens to be in the driver's seat in this picture.
buses
Get your AI windows talking. Across screens, across machines, near-zero tokens.
0 idle tokens · tamper-proof messages · bash + jq + openssl · 15 test rounds, all green · Claude · Codex · Gemini · any local LLM
About the name — a bus is a named channel (general, deploy-watch, mobile-team, whatever you want). You can spin up as many buses as you need; each session subscribes only to the ones it cares about. Within a bus, messages target all, a specific recipient, or a comma-list of recipients — and @-mentions in the body let one AI tag another by name or short-id, the way you'd @ someone in Slack. Replies, threads, broadcasts, side conversations — all over a folder.
Claude Code gets auto-delivery via the plugin hook. Codex, Gemini, and local-LLM orchestrators get the same auto-delivery by prefixing their model invocation with buses-wrap. See Cross-CLI ridership.
You know how you sometimes open three Claude Code terminals — one for the backend, one for the frontend, one to run tests — and end up copy-pasting between them like a hostage negotiator? buses makes that go away. Your AI sessions can leave each other notes, broadcast updates, and tag each other into specific threads. Everything flows through a folder they all see, and the messages just appear the next time you type into the other window.
Who this is for
- You run multiple Claude Code terminals on one machine and want them to actually coordinate, not just exist
- You hop between a laptop and a server and want both AI sides on the same page
- You're a small team sharing a project and want your AIs to chatter so the humans don't have to repeat themselves
- You want to leave yourself a queue — "next-future-Claude, when you wake up, here's where I left off"
- You want agent-to-agent task handoff — Claude on your laptop tells Codex on your server "deploy UAT", Codex deploys, Codex replies "done" → all autonomous, all over a shared folder
A 60-second demo
In the first terminal:
/plugin marketplace add /path/to/buses
/plugin install buses@buses
/buses:start
/buses:start is a guided wizard — it asks you 3 short questions (shared-folder path, your terminal name, which bus to join) and runs the right commands for you.
In a second terminal, on the same or another machine: same three commands. Pick a different name when asked.
Then in terminal A:
/buses:send all <name-of-B> "hey can you check the build log?"
Terminal B sees a 📬 buses: 1 new message(s) from <A> line above Claude's next response, the moment you type anything in there. No polling. No wasted tokens.
What it costs
| Activity | Tokens | Why |
|---|
| Idle terminal, no traffic | 0 | The hook prints empty output → nothing added to context |
| Receiving a 1-line message | ~30 | Just the body + a short framing tag, injected once |
| Sending a message | one slash-command's worth | Same as any /buses:* invocation |
| Background watcher (desktop pings) | 0 ever | It's a plain bash polling loop, never calls Claude |
Autonomous task handoff (buses-react) | 0 when idle, one wrapped-AI call per drain | Polling daemon fires buses-wrap <your-ai> only when unread > 0 |
That "0 when idle" is the whole point. Most terminals are sitting there waiting for you to type. They cost you nothing. The instant someone has something to say, the message lands in your next prompt's context and Claude tells you.
Full breakdown of every delivery path: docs/COSTS.md.
What's a bus
