Claude Code Sessions Can Talk to Each Other Now. Here's the Problem It Actually Solves.

I run more Claude Code sessions in parallel than I probably should. A worktree building something on one branch, a second terminal poking at something unrelated on the same repo, occasionally a third I forgot I left open. The failure mode is always the same one: a session in one tab learns something a session in another tab needs, and the only pipe between them is me. I read what one wrote, tab over, and retype it into the other. If I forget, one session ships a change the other one is about to conflict with, and I find out from a merge error instead of a warning.

That pipe is gone now. Claude Code shipped cross-session messaging starting with version 2.1.224 on August 7, and finished the feature this week: version 2.1.232, out August 13, added the ability to name a target session directly in your prompt with an @ mention, the same way you already @-mention a subagent. Sessions on one machine now also get unique names automatically, so you're not squinting at two terminals both called my-app-3f trying to guess which is which.

What actually shipped

Two tools do the work: ListAgents finds which of your sessions Claude can currently reach, and SendMessage delivers a message to one of them by name. You don't call either one yourself. You ask in plain language, and Claude figures out the target and writes the message:

Let the session working on the payments API know the schema migration finished

Or, as of this week, name the target directly:

Let @api-worker know the schema migration finished

Type @ and Claude Code suggests your other live sessions on the same machine, the same typeahead you'd get mentioning a subagent. It's on by default the moment you're on a supported version. Nothing to enable, nothing in a settings file to flip.

That first row is the one worth sitting with. Message a session in your other terminal and the text never leaves your laptop: it travels over a per-session Unix socket, not through Anthropic's servers at all. Only cross-machine and cloud messages route through Anthropic to reach you elsewhere. If you're the sort of person who wants to know where your bytes go before you turn a feature on, that's the actual answer, not a marketing line.

It sends a note, not the conversation

The part I'd get wrong if I hadn't read the docs closely: a message is a short piece of text one Claude writes for another, not the sending session's history, files, or context. If session A wants session B to actually pick up where A left off, that's a different feature (resume), and it's been around for a while. What's new here is narrower and, honestly, more useful day to day: a nudge, a finding, a status. "The migration's done, rebasing is safe." "That endpoint moved, update your calls." Short, and it arrives between tool calls so it never interrupts a session mid-command.

That narrowness is also the safety story. A message from another session can't approve a pending permission prompt on your behalf, can't touch CLAUDE.md or settings because another session asked, and if the message contains something that looks like a slash command, it lands as plain text instead of running. If acting on what the message says needs a permission the receiving session doesn't already have, you get the same prompt you'd get from anything else.

There's also a throttle built in so two sessions can't get stuck talking past each other: Claude Code rate-limits repeated messages, drops identical repeats arriving in a short window, and caps how many accepted messages a session can have waiting unread at 50. A message loop stops itself.

Two smaller pieces from the same release

Worth a line each, since they shipped in the same window and touch the same crowd. subagent_type: "fork" subagents now inherit the full conversation and prompt cache by default, where before that was an opt-in environment variable. And spawning an agent that isn't a teammate inside an interactive session now runs it in the background by default instead of blocking your terminal. If you're building on the Agent SDK or running a lot of subagents through Claude Code, both are worth checking against whatever you were relying on the old defaults for.

What I'm actually doing differently

I'm naming my sessions on purpose. An auto-derived name like claude-camp-3f is fine until you have three of them, and an @ mention is only as useful as the name you're typing. /rename takes ten seconds and I'm doing it at the start of any session I expect to run alongside others.

I'm letting a session speak up instead of checking on it myself. The habit I'm building is asking the session doing the risky change to message the session it'll affect the moment it lands, rather than me tabbing over on a timer to see if it's done yet.

I'm treating the local-only row as the default I actually want. Same-machine messaging never touches Anthropic's servers, which is exactly the case I use most (parallel worktrees on one repo). Cross-machine and cloud messaging go through Anthropic to reach you elsewhere, which is a reasonable tradeoff, but it's a different one, and now I know which is which before I lean on it.

If you're running more than one Claude Code session at a time and haven't touched this yet, it costs nothing to try: type /list-agents in any session on v2.1.224 or later and see what it already knows how to reach.

Sources: Message your other Claude Code sessions · Claude Code changelog · Claude Code 2.1.224 Brings Cross-Session Messaging and Self-Hosted Runners