ai-room

One terminal chat room for every AI CLI on your machine.

Claude, Codex, Gemini (Antigravity), Ollama — all in one conversation. Talk to all of them at once, address one directly, or make them debate each other. Zero dependencies.

ai-room — one terminal, every AI on your machine

In the room: Claude, Codex, Gemini, Ollama

you › which database should I use for a small SaaS?
Claude: For a small SaaS, Postgres — boring, battle-tested, and you...
Codex: Postgres. SQLite if you're truly single-server and want zero ops...
Gemini: Postgres is the safe default; consider SQLite + Litestream if...

Run it

npx ai-room

(or from GitHub before the npm release: npx github:kirpajeet12/ai-room)

No setup. It detects which AI CLIs you already have installed and logged in — anything missing is simply not in the room.

Supported out of the box:

| Agent | CLI it drives | Install | | ------ | -------------------------- | -------------------------------------- | | Claude | claude (Claude Code) | https://claude.com/claude-code | | Codex | codex (OpenAI Codex CLI) | npm i -g @openai/codex | | Gemini | agy (Google Antigravity) | https://antigravity.google | | Ollama | ollama (local models) | https://ollama.com |

Use it

| You type | What happens | | --------------------------------- | ------------------------------------------------ | | hello everyone | every agent replies, in parallel | | @codex review this idea: ... | only Codex replies | | /debate tabs vs spaces 3 | agents argue 3 rounds, then one gives a verdict | | /agents | who's in the room | | /mute ollama / /unmute ollama | take an agent out of / back into the room | | /export notes.md | save the whole transcript as markdown | | /help, /quit | you can guess |

Each agent sees the shared transcript, so they respond to each other, not just to you — ask a question, then ask another agent "do you agree with Codex?"

Add your own agent

Any CLI that takes a prompt and prints a reply can join. Create ~/.config/ai-room/config.json:

{
  "agents": [
    {
      "name": "mistral",
      "label": "Mistral",
      "color": "34",
      "bin": "ollama",
      "cmd": ["ollama", "run", "mistral", "{prompt}"],
      "out": "stdout"
    }
  ]
}

• cmd is an argv array — {prompt} is replaced with the outgoing message (no shell, no quoting bugs)
• out: "stdout" reads the reply from stdout; "file" for CLIs with noisy stdout that support writing their final message to a file ({outfile} in cmd)
• color is an ANSI color code (31–37)
• Reusing a built-in name overrides that built-in

Notes

• Replies take as long as the underlying CLI takes — a cloud agent call can be 10–30 seconds. Local Ollama models are usually faster.
• Each agent uses your existing login/subscription for that CLI. ai-room never sees a key or token; it just runs the same commands you could type yourself.
• Agents are called fresh per message with the recent transcript as context (last 30 messages).

License

MIT © Kirpajeet Singh Gill