Claude-peers: Enables multiple Claude Code sessions to communicate directly with each other, forming an AI collaboration team running locally.

A project called claude-peers-mcp has sparked widespread discussion within the developer community. It enables multiple Claude Code sessions to communicate directly with each other and synchronize task states without any cloud services or API intermediaries—the entire system runs locally.

Core Concept: AI Colleagues, Not AI Tools

Traditional multi-agent systems require an “orchestrator” to coordinate various AI agents. claude-peers takes a completely different approach—it makes each Claude Code session an equal “colleague” that can communicate directly, without a central management layer.

Example real-world scenarios:

Claude A (handling poker engine): “Which files are you editing?”
Claude B (handling frontend): “Editing auth.ts + UI state”
Claude A: “Okay, I’ll avoid auth logic”

No manual coordination needed—AI syncs itself.

Technical Architecture

• Local broker daemon (localhost:7899): acts as a message relay
• SQLite peer registry: records all active Claude sessions
• Each session’s own MCP server: provides a tool interface
• Real-time channel for message pushing: zero-latency communication
• Automatic peer discovery: new sessions join the network automatically

Four core commands:

• list_peers: Find all active Claude sessions
• send_message: Send a message to another Claude
• set_summary: Describe your current task
• check_messages: Manually view received messages

Each session also automatically summarizes its status, allowing other Claudes to see: current working directory, git repo, current task, files being modified.

Typical use cases:

• One Claude writes backend, another writes frontend, automatically avoiding conflicts
• One Claude handles debugging, another performs refactoring simultaneously
• A research Claude gathers data and feeds it to a builder Claude
• Large projects split among multiple AI workers working in parallel

Limitations and Challenges

The community has raised some practical issues: running five Claude Code sessions simultaneously can significantly strain local hardware; when multiple sessions try to run npm install or git operations at the same time, additional conflict resolution mechanisms are needed; context length limits may cause sessions to lose synchronization over long periods.

However, overall, this tool signifies a broader shift: evolving from “a single AI assistant” to “an AI collaboration team,” all running locally on your machine, without cloud dependencies or extra costs.

This article on claude-peers: enabling direct communication between multiple Claude Code sessions, a locally-run AI collaboration team, was first published by Chain News ABMedia.