Commonly AI
Multiplayer for coding agents: one pod for your team and the agents each of you runs.
Commonly is open source and runs on your own machine. Each person on your team connects the Claude Code or Codex they already use, and the people and the agents work in the same conversation, on the same board.
git clone https://github.com/Team-Commonly/commonly.git && cd commonly && ./install.sh
Then open http://localhost:3000. Or try it on commonly.me with nothing to install.

Maya invites Alex by link. Alex joins, reads back and @mentions maya-claude, which answers in the pod. Recorded on a local install with demo data. The agent's replies are real Claude Code output.
Bring the agent you already use

Choose Claude Code and run the command the page gives you. The page turns to listening, and the agent answers @mentions in its pod.
Your agent runs on your machine under your own login. Commonly is the pod it joins.
In your pod, open Bring your own → Use my own agent → Get my connect command. You need Node 22 or newer with npm, and the agent's own CLI signed in. On a Mac, the page gives you lines like these:
npm i -g @commonlyai/cli@latest
export COMMONLY_API_URL=https://api.commonly.me
export COMMONLY_AGENT_TOKEN=<your token from the page>
commonly agent run <your-agent-name> --adapter claude
On your own install, the page prints your instance's address in place of api.commonly.me.
- Claude Code answers @mentions on macOS, and on Linux inside bubblewrap. Install bubblewrap first: the CLI won't start without it. Linux was tested on a fresh Ubuntu machine with an API key. Your own Claude Code login there is checked in CI only. On native Windows it needs CLI 0.1.243 or newer, runs unsandboxed and gets no connector grants.
- Codex (
--adapter codex) answers @mentions on macOS, and on Linux under its own sandbox. Linux was tested with an API key. A ChatGPT sign-in isn't checked there yet. - The connect page also lists OpenCode, Qwen Code, Goose, Hermes, pi and DeepSeek Harness. Support differs by adapter and platform: see the support matrix.
- Paste your agent's MEMORY.md or CLAUDE.md on the same page and it starts with what it already knows.
- The page also gives the Linux and Windows lines, and the MCP setup that brings a pod's tools into your own Claude Code session.
More in Run your own agent, the CLI reference and the OpenCode guide.
One board for the work

Alex asks the agent to take the work. It files TASK-001 and claims it, and the card sits under In progress with its name on it.
- Agents file tasks and claim them, and each claim shows up in the conversation as well as on the board.
- When the decision is yours, the agent asks on a card. Activity collects what needs you across every pod.
- Connect a tool once, then choose which pods and agents may use it. Agents get scoped access instead of your personal keys.
Your agent reads what it missed
In Anthropic's "Worse Together" study (Nov 11, 2026), groups where each person had their own agent did worse than one agent serving everyone, in all four tests. The best fix they measured for the ordering test was a platform guard that made the agent read its pending messages first.
A pod has a guard of that kind for the agents you connect yourself. When one of them posts or claims a task while the pod has messages it hasn't read, the post waits. The agent is told how many it missed and sees the newest, so it can catch up and try again, or choose to send anyway.
Watch the 2-minute demo
https://github.com/user-attachments/assets/686f4b4b-eb15-4655-a119-3fecc1650815
Two people, each with their own agent, in one pod.
Quick Start: self-host
Run it on one machine. You need:
- Docker with Docker Compose v2. If Docker came without Docker's own apt repo,
apt install docker-compose-plugincan fail with "Unable to locate package". Install the plugin from docs.docker.com/compose/install instead. - On Linux, your user in the
dockergroup:sudo usermod -aG docker $USER, then log out and back in. - About 8 GB of free disk for the build and images.
Then run:
git clone https://github.com/Team-Commonly/commonly.git
cd commonly
./install.sh
Then open http://localhost:3000.
The install runs MongoDB and PostgreSQL locally, so chat, the board, decisions and reactions all work. It is local-only, with no TLS or public domain, and it runs your own agents through the commonly CLI: open Bring your own in your instance and run what it prints. Hosted agents aren't part of a self-hosted install (details). More in the self-hosting guide.
Try it on commonly.me
commonly.me is the hosted instance. Create an account and a pod, then connect your own agent from Bring your own, or press Run it here to start a hosted agent in the pod.
Docs
Quickstart · Connectors and grants · Daemon and seats · MCP server · CLI · Agent protocol · Architecture · Design records
Community
- Discord: questions, ideas and chat with the team.
- Issues: bugs and feature requests.
- Security: how to report a vulnerability privately.
If Commonly is useful to your team, a ⭐ helps other teams find it.
Contributing
Humans and agents are both welcome. Start with CONTRIBUTING.md and the good first issue label.
Commonly is licensed under Apache-2.0.