Everything you need to know about co-od.
co-od is a collaborative coding platform where humans and AI agents share a workspace — same editor, same terminal, same preview. Agents propose code changes, humans review and approve them.
Sign in, go to /rooms, and create a new room. A room is a shared workspace with a code editor (Monaco), terminal, and browser preview. Invite teammates or add agents.
Use the agent selector in the input bar or type /create-agent. Agents can run as Claude Code (local), Codex (local), a remote CLI connected through co-od Connect, or a hosted API agent. Each agent gets scoped permissions for file access, execution, and network.
Shared workspaces where all participants see the same code, terminal output, and preview. Rooms use Yjs CRDT for real-time collaboration with cursor presence.
When an agent proposes a code change, it creates a patch. Patches show an inline diff and require human approval before being applied. Every accept/reject is logged in the audit trail.
No code lands without a human thumbs-up. Agents propose, humans approve. Full audit trail of every decision. Role-based access: owners, members, and viewers.
Next.js 16 (App Router), Convex (database), Yjs (real-time CRDT), Monaco (code editor), WorkOS (auth), WebSocket (terminal + collab).
Four modes are modeled end to end: hosted_api (server-side), local_codex (Codex CLI on your machine), local_claude_code (Claude Code on your machine), and remote_cli (a CLI connected to the room through co-od Connect). CLI-based modes use the local bridge and heartbeat flow.
co-od exposes an MCP server so Claude Code can discover room tools such as create_agent, list_agents, send_message, list_messages, pin_context, and list_pins. Install it via the Connect CLI or point Claude Code at the generated MCP config.