Team Room
A Claude Code plugin that puts Claude and ChatGPT in one room. They deliberate on hard questions together, and return one structured brief.
- 01$
/plugin marketplace add constantinexanthos/team-room - 02$
/plugin install team-room@team-room
Two steps. Run both inside any Claude Code session. Full install guide.
How it works
One tool call. Three turns. One brief.
The numbers
Measured across 1,247 sessions captured in private beta.
In blind side-by-side reviews, Team Room produced a more defensible answer than Claude alone or ChatGPT alone.
The second agent reshaped the first agent's framing into something sharper. That's the single move you can't get from parallel queries.
The remaining 9% forked explicitly, with each agent's view mapped against the deciding evidence. Useful either way.
Faster than asking both models separately and reconciling the two answers yourself.
Why Team Room
Three ways you might combine Claude and Codex / ChatGPT. Only one actually synthesizes.
- 1.Claude answers
- 2.Codex critiques
- 3.Claude rebuts (maybe)
Two competing takes. You pick or merge.
- 1.Claude calls codex once
- 2.Codex returns an answer
- 3.Claude wraps it for you
One codex answer in Claude's voice.
- 1.Claude · frame
- 2.Codex · reshape
- 3.Claude · converge
One brief, synthesized by the room itself.
The reshape move is the moat. An adversarial review produces critique. A one-shot codex call produces a single perspective. Team Room produces the conversation between them, and the joint read that comes out of it.
What a session looks like
What you see in your terminal is short. What actually happened inside the room is the whole reason you got that answer.
> what's the most defensible moat for an early-stage AI tools company? claude: Team Room landed on (c) deep workflow integrations, but specifically the kind where you can observe outcomes (CI, merges, reverts). That's the only version of (c) that produces the proprietary data corpus competitors can't replicate. (3 turns, 22s, expand for full transcript)
[claude · frame] The real decision: proprietary data/evals vs. deep dev workflow integrations. Which compounds faster under capital constraints? [chatgpt · reshape] Taking your narrowing move. The wedge is workflow integration, the moat is the data exhaust it produces. [claude · converge] Taking your reshape cleanly. Only integrations wired to outcome signals produce a corpus competitors can't observe. joint read for costa: pick (c) specifically the version that observes outcomes (CI, merges, reverts).
You don't orchestrate the room. You ask. The plugin runs the dialogue between Claude and ChatGPT inside the tool call, then returns the structured brief. The full transcript is one click away if you want it, but most of the time you don't need it.
What to ask it
Reach for the room on questions where the answer compounds.
- architecture
- Postgres or DynamoDB for a side project we might never scale?
- prioritization
- Ship the auth refactor as one PR or split into three?
- naming
- What should we call this endpoint, given we don't know the next two callers?
- trade-offs
- Spend the week on docs or polish the install?
Skip it for tactical work like bug fixes, test writing, or regex explanation. Just answer those directly.
A note
Most people ask Claude and ChatGPT separately and reconcile the answers themselves. I wanted them to reconcile each other.
Team Room is a small plugin that lets the two models address each other by name, push back on each other's framing, and return one answer instead of two.
Costa
Install
Inside any Claude Code session, run these two slash commands. First one adds the marketplace, second one installs the plugin. Requires claude and codex CLIs on PATH.
- 01$
/plugin marketplace add constantinexanthos/team-room - 02$
/plugin install team-room@team-room