9to5mac.com web signal

Claude Code v2.1.224 lets sessions message each other on macOS and Linux

3 sources tracking this story

TL;DR

  • Same-machine messages travel over per-session Unix domain sockets and never reach Anthropic servers; cross-machine replies route through Anthropic as reply-only.
  • Inbound messages cannot approve pending permission prompts, change configuration, or execute slash commands in the receiving session.
  • The feature is absent from Bedrock, GCP Agent Platform, AWS Claude Platform, and Microsoft Foundry, limiting it to direct-API and Claude.ai subscribers at launch.

Anthropic shipped a small but genuinely interesting change to Claude Code this week: sessions can now talk to each other. 9to5Mac reports that as of v2.1.224, released August 7, one Claude Code session can deliver a message to another running session on the same account, on macOS or Linux. Windows is not supported at launch. It lands a week ahead of the August 14 switch to auto mode by default.

The mechanic is deliberately restrained. You do not compose the message yourself; you tell Claude what you want the other session to know, and Claude writes the summary that gets delivered. Anthropic describes the payload as text only, never conversation history or files. The intended jobs, in the company's own framing, are handing over a finding, coordinating parallel worktrees, getting status from long-running work, and replying across machines. That is a real gap in how people actually use Claude Code today, where the standard workaround is a human ferrying context between terminals.

Why this is worth paying attention to beyond the changelog: the primitive Anthropic is exposing here, an agent that can notice something and nudge another agent about it, is the thing most third-party multi-agent frameworks are built to sell. Folding a basic version into the tool most Claude developers already run reprices that market a little. It also makes long-running background work more useful, because a research session or a slow refactor can now interrupt the foreground when it finds something instead of waiting to be checked on.

The scope here is narrow by design. Anthropic has walled off the actions that would make cross-session messaging dangerous: it cannot approve permission requests, cannot change configurations, and slash commands like /compact arrive as plain text rather than executable instructions. Anything privileged still prompts the receiving user. 9to5Mac's writeup skips the mechanics: how the messages travel between sessions, whether there are rate limits or token costs, and how authentication works when the two sessions sit on different machines. Those are the details that will decide whether teams treat this as a convenience or as infrastructure to build on.

What others are reporting

Coverage cluster as of 24h after publish

  1. Anthropic (Claude Code Docs) Read →

    First-party spec covering Unix socket delivery, reply-only cross-machine routing, inbound controls, provider exclusions, and per-session permission isolation in full technical detail.

    Between two ordinary interactive sessions with default settings, Claude Code delivers the message. Delivery isn't guaranteed in every configuration, though.
  2. MacRumors Read →

    Consumer-focused coverage emphasizing the privacy architecture and what the feature explicitly cannot do: no permission approval, no config changes, no history sharing.

    Messages between sessions on the same Mac stay local and never reach Anthropic's servers.