claude-rooms

claude-rooms

Multiplayer Claude Code. Code with a friend in your own terminals, your Claude sessions aware of each other in real time.

Install

/plugin marketplace add hyattc1/claude-rooms
/plugin install claude-rooms@claude-rooms

Try it

connor> /claude-rooms:rooms-create
        Room: kite-frog
ryan>   /claude-rooms:rooms-join kite-frog
        Joined room kite-frog as ryan. 1 teammate online: connor.
connor> /claude-rooms:rooms-status
        Room: kite-frog
        You: connor (online)
        Teammates: ryan (online)

That is it. From this point on, both Claudes have continuous live awareness of each other. Anything either agent does (edits, focus changes, file locks) propagates to the other in the background. Either agent can call the read_room_state MCP tool at any time during its own turn to pull a fresh snapshot.

Architecture in three paragraphs

Peer-to-peer, no server. Every developer's Claude Code session runs its own claude-rooms MCP server, which holds a live CRDT document for the room. Sessions sync that document directly to each other over WebRTC, using y-webrtc on top of yjs. There is no relay we run, no account to make, no infrastructure to deploy. The room code is the rendezvous string used by the public signaling servers; whoever has the code is in the room.

Shared state, not messages. Agents do not message each other. They share a single document whose state describes every agent's current focus, branch, recent files, recent actions, and active file locks. Each agent reads from this live document via the read_room_state MCP tool, which is a sub-100ms in-memory query. The agent decides how to react. The PreToolUse hook on Write/Edit/MultiEdit also checks the document deterministically before any file edit, so a conflicting edit is denied with a structured reason naming the teammate who holds the file.

Hooks for plumbing, MCP for context. The plugin ships a small set of hooks (SessionStart, PreToolUse, PostToolUse, Stop, SessionEnd) that publish each agent's current state to the local MCP server over a Unix domain socket (or named pipe on Windows), and an MCP server that owns the y-webrtc connection and exposes two tools (read_room_state, update_my_focus) to the agent. Hooks fail open: if the MCP server is unreachable or the room is offline, every hook exits 0 silently and your local Claude Code session keeps working. claude-rooms going offline is a feature loss, not an outage.

Full design rationale (state schema, lock semantics, awareness reaping, the six lock cases, fail-open principles, out-of-scope decisions) lives in CLAUDE.md.

Slash commands

Command	What it does
/claude-rooms:rooms-create	Generate a new room code and join it. Share the code with a teammate.
/claude-rooms:rooms-join <code>	Join an existing room. Warns if no teammates are detected after 3 seconds.
/claude-rooms:rooms-status	Print the current room code, you, teammates, active file locks, and any territory overlaps.
/claude-rooms:rooms-leave	Release your locks, drop your territory claim, mark offline, and clear the local room state.
/claude-rooms:rooms-doctor	Print a diagnostic page (WSL2 status, ICE config, MCP socket, recommended next step). Works outside a room.

Inside the plugin, the agent also has these MCP tools:

Tool	What it does
read_room_state	Snapshot of every teammate's live state. Call it liberally.
update_my_focus	One-phrase status, so teammates see what you are currently working on.
update_my_plan	Compact summary of a multi-step plan plus done/total counts.
claim_territory	Declare which directories/files you intend to touch for the current task.
release_territory	Drop your current territory claim.

What v1.2 adds

Better WSL2 ergonomics and a real diagnostic. Three things: