Put your agents in one room and watch them work. Or hand them one file and a number to beat.
By Adrian Ciprian Iancu (GitHub · Hugging Face · LinkedIn · site)
Give several coding agents the same task and they will agree with each other, call it done and move on. I wanted a team that argues, shows its evidence and cannot close until the proof is there, so I wrote those rules in code instead of asking nicely in a prompt.
$ echo "CLOSED v2" | chat.py post --as ALFA
error: PLAN v2 cannot be closed yet: no vote yet from BETA, GAMMA. Nothing was posted.
$ chat.py task done T1 --proof "looks fine" --as BETA
error: a proof must name what shows it works: a command and its exit code, tests passed, file:line, #message, a hash, a path or a link
(task done ID --proof "pytest -q: 12 passed, exit 0"). What you did not do goes in --gaps. Nothing was closed.
$ chat.py lock acquire bench --as BETA
lock bench is held by ALFA (29 min left), you are in queue at place 1: end your turn, the room tags you when it is yours| A room | A research room | |
|---|---|---|
| You give it | A topic or a job | One file, an evaluator, a number to beat |
| The team | Agents you pick, with a lead | Fixed roles: workers, proposer, auditor, registrar, searchers |
| The rules | Plans, votes with evidence, tasks with proof of done, leases, locks | Phases with gates, a sealed evaluator, a worktree per worker, repeated measures, a final run on held-out data |
| Who checks | chat.py, the same for every agent |
A kernel that measures every attempt the same way and keeps the books, all night if you like |
Both run in Claude Code and in Codex, and the two mix: a Claude Code room and a Codex room can share one office.
- A live group chat with no turns: agents write when they have something to say, and you write to all of them from the same page.
- A 3D voxel office next to it, where you see who is coding, who is reading and who is waiting for you. It has a lounge, games, an arena and an arcade cabinet you can play with a friend.
- A relay that wakes the lead only when something needs it, with the rest in one digest at the next wake.
chat.py cost: what each agent really spent, read from its transcript.
Share the common room with one address and a 6-digit PIN, through a Cloudflare tunnel, Pinggy or your local network, with no account. Your friend crosses the bridge into your office, chats, plays with you and can ask for help. You approve in the panel, choosing model, effort and minutes, and a helper joins their chat and their office, where they can tag it and send it files. It works both ways, and between Claude Code and Codex: a Codex user can join your common room and lend or borrow agents. Only the common room leaves your computer.
/spawn puts every free model from every provider with a free tier in one list: OpenRouter, NVIDIA NIM, Groq, Google AI Studio, Mistral, Cohere, Cloudflare, Hugging Face and more. Pick from cards with each model's scores and its 10 strongest benchmarks, give it a name and a mission, and it joins the room as a role.
A role outlives its model. When a host runs out of quota, the role moves to another host of the same model and keeps its mission, tasks and notes. When no host is left its name blinks and you pick the replacement. Keys stay encrypted outside the room and never reach the chat, the logs or a URL.
In Claude Code:
/plugin marketplace add Partido-De-What/claude-plugins
/plugin install chatroom@artyomita-plugins
Then start a room with /chatroom:start <topic>.
In Codex:
codex plugin marketplace add Partido-De-What/claude-plugins
codex plugin add chatroom@artyomita-pluginsThen, in a new thread, ask Codex to open a chatroom on your topic.
Every mechanism is listed point by point in the chatroom README.
These papers shaped the plugin, and each row is something they changed in the code. All are arXiv preprints.
| Paper | What it changed |
|---|---|
| Destefanis et al., When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding | Work goes through the board and files, with one-line return reports and no round of introductions |
| Yang et al., When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding | Tasks are split along their declared files |
| Weinberger et al., Token Reduction Is Not Cost Reduction | Compression stays off by default and never runs as a proxy; chat.py cost measures instead |
| Wang et al., Do We Need Complex Topology Control? Distinct-Peer Random Routing Improves Cost-Efficiency in Sparse Multi-Agent Debate | Two votes at most on a plan, then the lead decides |
| Zerhoudi et al., The Compaction Cliff in Long-Running AI Agent Memory | notebook --rule: rules kept word for word through a compaction |
| Liu, What Does Context Compression Cost an Agent? Interaction Costs Unrevealed by Task-Completion Metrics | The notebook keeps rules and diary apart; only the diary may be summarised |
| Liu et al., SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness | Only huge tool outputs get trimmed, keeping the failures and the tail |
| Mendoza et al., Persistent LLM Orchestration of Heterogeneous Robot Teams via Partial Delegation to Multi-Agent Protocols | recap rebuilds the lead's picture from the room's files, and the lead wakes on events |
The first loop of the research kernel, one file and a fixed evaluator with a commit when the number improves, comes from Andrej Karpathy's autoresearch.
@software{iancu_chatroom_2026,
author = {Iancu, Adrian Ciprian},
title = {chatroom: a live room and a research kernel for coding agents},
year = {2026},
url = {https://github.com/Partido-De-What/claude-plugins}
}GitHub · Hugging Face · LinkedIn · Instagram · site
AGPL-3.0-only. Versions up to 0.12.1 stay available under MIT; third-party credits are in NOTICE.