Anthropic Claude Code Enables Cross-Session Messaging

Anthropic has rolled out Claude Code version 2.1.224, introducing a cross-session messaging feature that allows parallel AI sessions to communicate directly. Previously, developers running multiple instances of the command-line tool had to manually copy and paste context, decisions, and error resolutions between terminal windows. This update eliminates that friction by letting independent sessions discover and message one another to coordinate their programming tasks automatically.

Under the hood, the system relies on two new tools: ListAgents, which identifies reachable sessions and their names, and SendMessage, which transmits the actual communications. To maintain security and efficiency, these messages are restricted to plain-text summaries. The sessions do not share full conversation histories or raw files, preventing unnecessary context bloating while still passing critical updates between the active agents.

The communication architecture depends on where the sessions are running. For instances operating on the same machine, Claude Code utilizes a local Unix socket to pass data. For cross-machine communication, which operates on a reply-only basis, messages are routed through Anthropic servers via Remote Control. Currently, this functionality is restricted to macOS and Linux environments. It is not supported on Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry.

To protect against potential exploits, Anthropic has integrated security measures to prevent a compromised session from hijacking others. Users can manage inbound traffic using the crossSessionInbound control, which can be configured to accept, hold, or refuse incoming messages. Additionally, a built-in prompt-injection firewall monitors these interactions to ensure that malicious instructions cannot be passed between active agents, keeping the developer's local environment secure.