Claude Code sessions gain direct messaging across terminals
A new channel lets one Claude Code session hand findings to another over a local socket, with cross-machine exchanges limited to replies only.
Anthropic has shipped cross-session messaging in Claude Code, letting one running session pass a written message to another instead of the user copying context between terminals. The feature requires version 2.1.224 or later, runs on macOS and Linux including WSL 2, and is active with no setup once those conditions are met.
The model gets two tools: ListAgents to enumerate reachable sessions and SendMessage to address one by name. Users can prompt a send — asking whether a migration finished in another terminal, for instance — or the agent can initiate one on its own when it judges that a change it just made breaks work another session depends on. What travels is plain text only: never conversation history, files, or the sender's context. Sessions on the same machine communicate over a per-session Unix domain socket restricted to the operating-system user, with nothing routed through Anthropic's servers. Reaching a session on another machine, or on Claude Code on the web, does go through Anthropic and is reply-only — a session cannot initiate a conversation across machines.
Anthropic has fenced the channel with several restrictions. An incoming message is explicitly labelled as coming from another session rather than the user, so it cannot satisfy a pending permission prompt, cannot instruct the receiver to alter permission settings or CLAUDE.md, and any slash command inside it arrives as inert text. The receiving session's own permission rules still apply to anything the message asks for. A crossSessionInbound setting accepts, holds or refuses arriving messages; by default a session that bypasses permission prompts holds messages for approval. Repeat sends are rate-limited and identical messages deduplicated so two agents cannot spin into a loop, with at most 50 queued and 100 held. Administrators can disable both directions through managed settings, and the feature is unavailable on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry.
Why it matters
Developers increasingly run several coding agents at once across worktrees, and the coordination has been manual. Giving sessions a peer channel turns a set of isolated agents into a loosely coupled system — and creates a new propagation path for prompt injection, which is why the design deliberately strips the message down to text with no authority to approve or reconfigure anything.