Team Room

A two-mind deliberation room. Claude (Anthropic) and Codex (OpenAI) collaborate on your question over short turns, address each other by name, and arrive at an answer together. Use it three ways:

• As a Claude Code plugin (/team-room "...") — agent-callable from any session via the bundled MCP server
• As a shell tool (scripts/ask-team-room.sh) — for cron, scripts, other agents
• As a browser app (./start.sh) — chat-style UI with projects, topics, live streaming

Why

Same-model self-consensus drifts toward agreement. A second voice from a different lab catches things the first model would systematically miss. The current dialogue mode frames the agents as colleagues building an answer together, not graders critiquing each other's essays.

Install as a Claude Code plugin (recommended)

If you have Claude Code installed:

/plugin marketplace add constantinexanthos/team-room
/plugin install team-room@team-room
/reload-plugins

After install, in any Claude Code session:

• /team-room "your question" — opens a working session, returns the dialogue
• /team-room:rooms — list recent topics
• /team-room:status <topic> — check an in-flight session
• The MCP tool team_room.ask is also callable directly by other agents — that's the agent-orchestration interface

The plugin bundles an MCP server (stdio, no daemon) that Claude Code launches on demand. State (transcripts) lives in ~/.team-room/ by default — override with TEAM_ROOM_DIR=....

Requirements: claude CLI authed (Anthropic Max or API key), codex CLI authed (ChatGPT Plus/Pro or OpenAI API key).

Quickstart for the browser app

One-time install:

git clone https://github.com/constantinexanthos/team-room.git ~/team-room
cd ~/team-room
./install-desktop.sh

That drops a Team Room launcher on your Desktop. Double-click it any time:

1. Terminal opens, server starts on localhost:8765
2. Browser pops to the topic picker (all topics ranked by recency, with last-spoke role)
3. Click a topic → enter the transcript viewer

Close the Terminal window to stop. Re-open by double-clicking again.

Driving the conversation

In Claude Code (or any terminal), three commands:

# You/Claude proposes something into the room
./log.sh sub-project-e claude "Draft spec for the Overview tab: ..."

# Claude asks Codex for an adversarial read (high reasoning effort by default)
./ask-codex.sh sub-project-e "Critique this spec. Find what's missing."

# You weigh in directly
./log.sh sub-project-e costa "Pick option B, skip the polling refresh."

The viewer updates within ~1.5s. Color stripes on the left mark who's talking: orange = Claude, green = Codex, blue = you, purple = system.

Files

File	Purpose
Team Room.command	macOS double-click launcher (symlinked to Desktop by installer)
install-desktop.sh	One-time: drops the launcher symlink on your Desktop
start.sh	Boots the server + opens the browser
server.py	Tiny static server with /topics.json auto-discovery
index.html	Topic picker landing page
viewer.html	Single-page transcript viewer (polls JSONL every 1.5s)
ask-codex.sh	Send prompt to Codex, append both sides to a topic
log.sh	Append an arbitrary message (claude / codex / costa / system)

Transcripts

JSONL at .team-room/<topic>.jsonl (gitignored). One line per message:

{"ts":"2026-05-22T00:30:00Z","role":"claude","model":"claude-opus-4-7","content":"..."}

Topics are arbitrary slug strings — use one per decision (e.g. sub-project-e, policy-engine-design, launch-positioning). Old topics stay in the picker; nothing is ever deleted automatically.

Override the room dir with TEAM_ROOM_DIR=/some/path if you want transcripts somewhere else.

Requirements

• macOS (open and .command handling are Darwin-specific; would need small tweaks for Linux)
• Python 3.10+
• Codex CLI authed via codex login (ChatGPT Plus or Pro)
• A Claude session (Claude Code is the natural fit)

What v1 deliberately doesn't do

• No write path from the browser. Interventions happen in the terminal — keeps the security surface tiny.
• No persistence beyond JSONL. No DB, no search, no replay UI.
• No autonomous loop. Claude prompts Codex when there's a strategic call to make; Codex doesn't poll for new messages.
• No "type into the team room itself" input. You drive via the terminal; the viewer is read-only. Lifting this is the natural v2.

If any of those become friction, lift the limit. Don't pre-build.

License

MIT.