Murmur: Claude Code (and other CLI) sessions talking to each other
retained text
I'm often having several Fable/Opus sessions as orchestrator/reviewer having their subagents working on different parts of big projects and often times their work overlap and they need to communicate to make sure we don't duplicate workloads. The problem is I found myself having to chime in, paste messages on different sessions/windows to have them all aligned, and also the duplicated messages: a Fable/Opus orchestrator would message other two orchestrators separately for guidance, and each of the ones that were contacted didn't know what the other had replied. Imagine 3 different Claude Sessions working on a same project, each having their own Sonnet subagents. Claude #1 sends a message to each peer Claude #2 and Claude #3, separately, then waiting for a reply. Now imagine Claude #1 sends only 1 message, both Claude #2 and #3 receive that same message, they reply asynchronously, Claude #1 receives each message and on top of that: Claude #2 knows what Claude #3 replied and vice versa. If I have to chime in to send 1 message to all 3 Claude instances (say, to align all of them at once on a decision I made) I just `/murmur <group | @group | handle[,more]> <msg>`. Some details: - One broker process, one SQLite file, one Unix socket. - Each session joins as a named peer (I recommend naming your sessions first) through its CLI's own hooks plus an MCP server. [Hummin](https://github.com/rennf93/hummin) (my own CLI built on top of Pi's) and opencode join through an extension or plugin. - One `murmur_send` reaches several sessions and can wait for their replies in the same call. Each reply goes back to whoever asked, not to the whole thread. - Idle sessions get woken up: Claude Code through its own messaging socket, Codex through `codex queue`, Hummin and opencode through their plugin APIs. Gemini CLI and zcode have no inbound channel, so they get the message at their next prompt or turn end. - It works across CLIs. I had a Claude Code session on Opus 5.5 and a Hummin session on GLM-5.3-flash ask each other a question, and both replies came back through murmur. So if you have different LLMs for different types of tasks, they can communicate. - A reply wakes only whoever asked; the other sessions in the thread aren't interrupted by it, and any of them can read the whole thread with `murmur_thread` when they need it. - Every delivery has a state: queued, sent, delivered (only with evidence from the receiving side, like the message tag in its transcript), acknowledged, and replied. - Each reply has a verdict flag: agree, disagree, info, or choice (for "pick A or B" questions). - Sessions are told to verify what a peer says against the files before agreeing, to protect against bias/yes-man behaviour. Results: 29% fewer tokens and 4 fewer API calls than the native setup. Method and raw numbers: [https://rennf93.github.io/murmur/bench/consult/](https://rennf93.github.io/murmur/bench/consult/) [Repo](https://github.com/rennf93/murmur) [Docs](https://rennf93.github.io/murmur/)
document fetched 2026-10-22 00:24 · canonical https://reddit.com/r/ClaudeAI/comments/1xbgbxx/murmur_claude_code_and_other_cli_sessions_talking