What Happened
A developer released Cross-Claude MCP, an open-source MCP server that creates a shared message bus between multiple Claude instances. The project, published on GitHub on March 7, 2026, solves a specific problem: Claude sessions are completely isolated from each other. If you have Claude Code running in three terminals, none of them know the others exist.
The server works by letting each Claude instance register with a name, then communicate through channels. Instances can send direct messages, broadcast to channels, share structured data, and block while waiting for replies. Think of it as a lightweight Slack, but for AI sessions instead of people.
The setup runs as a standard MCP server. Each Claude instance connects to it, picks a name, and immediately gains access to the message bus. From there, one instance can ask another to handle a subtask, share context about what it has found, or coordinate work on a larger project.
The code is available on GitHub under the rblank9 account.
Why It Matters
Anyone running Claude Code across multiple terminals has hit this wall. You have one instance researching a bug, another writing tests, and a third refactoring a module. They are all working on the same codebase but cannot talk to each other. You become the message router, copying context between sessions manually.
This matters more as multi-agent workflows become common. Claude Code users routinely spin up parallel sessions for speed. The Agent tool inside Claude Code already supports subagents, but those exist within a single session. Cross-session coordination has been a gap.
The MCP protocol is what makes this possible without hacks. MCP servers are the standard way to extend Claude's capabilities, and Anthropic has been pushing the ecosystem hard. A message bus built on MCP slots cleanly into existing setups without requiring any modifications to Claude itself.
For teams, the implications are worth noting. Multiple developers could have their Claude instances share findings in real time rather than duplicating research. A "lead" instance could delegate tasks to specialist instances and collect results.
Our Take
This is a clever use of MCP that addresses a real friction point. I have had sessions where I wanted one Claude Code instance to hand off a research summary to another, and instead I ended up pasting blocks of text between terminals like it was 2005.
That said, the value depends heavily on how you work. If you run a single Claude session at a time, this does nothing for you. If you regularly have two or three instances open on the same project, the coordination layer removes genuine overhead.
The bigger signal here is what this tells us about where MCP is heading. The protocol was designed for tool integration, but developers are now building infrastructure on top of it. Message buses, shared state, inter-agent coordination. MCP is becoming the Unix pipe of AI tooling, connecting processes that were never designed to talk to each other.
One concern: reliability under load. A shared message bus between AI instances that can each generate rapid-fire messages needs careful flow control. The README does not mention rate limiting or message queue depth, which would matter in any serious multi-agent setup.
Worth watching. If Anthropic sees enough traction with projects like this, native multi-instance coordination could land in Claude Code directly. Until then, this is a solid stopgap for power users who have already outgrown single-session workflows.