Claude Relay: Enable local Claude Code sessions to communicate with each other

Claude Relay addresses a specific friction point for power users of Claude Code: the isolation of sessions. When working across a distributed codebase—such as a frontend and a backend—developers typically run separate sessions to maintain clean context windows. Relay breaks these silos, allowing a developer to ask one session to query another (e.g., "ask the backend session if the auth token shape changed"), effectively turning separate AI instances into a coordinated swarm on a single machine. Technically, the implementation is lean and pragmatic. It deploys a detached hub daemon that manages a registry of active peers, routing messages via a Unix socket. The integration is handled through an MCP server that provides a set of `relay_*` tools for outbound communication and utilizes the `notifications/claude/channel` capability for inbound triggers. This allows the AI to 'hear' a message from another session between turns, maintaining the natural flow of the conversation. While the utility is high, the current state is experimental. The requirement for the `--dangerously-load-development-channels` flag signals that the underlying infrastructure is still in research preview. Furthermore, the lack of persistence and restriction to single-host communication limits its use to local development. However, for a builder managing complex, multi-repo local environments, the ability to broadcast status updates or cross-reference session context in natural language is a significant productivity win. This tool is primarily for developers who have fully integrated Claude Code into their workflow and find themselves manually copying context between terminal tabs. It is a clever use of MCP to build a lightweight communication layer that feels like a native feature of the AI agent.