I released version 2.0.0 of agent-chat today. It's a small Ruby gem that gives AI agents a chat room on your machine.
Quick introduction to agent-chat
In February, Claude Code launched Agent Teams. Agents in a team talk to each other one to one or by broadcast. I wanted something closer to a group chat for pair and mob programming, so I built agent-chat.
It's a message queue for agents (or humans) that uses room-based conversations, stored in SQLite files on your machine. It has a CLI, agent-chat, and a web UI, agent-chat-web.
I introduced it on LinkedIn with a seven minute demo. The post is embedded below, so press play there. In the video I install the gem, send a message from the CLI, watch it appear in the web UI, then start a Claude Code Agent Team that builds Hangman as a mob, with a TDD coach and two developers swapping the driver seat in one room.
That version worked, but agents had no way to say which message they were answering, and its output was meant for people to read.
What's new in 2.0.0
Message ids and replies. Every message gets a short random id, and send now prints it. send --reply-to <id> links a reply to one or more messages. An unknown id refuses the whole message. show --room <name> <id>... fetches messages by id.
<<< Alice | 2026-11-22 14:30:00 | 7xKqR2m | re: 3fGhJ9k >>> --json on every command. Messages are one JSON object per line with id, author, timestamp, content and reply_to. Refusals come back as a structured error with exit status 1, and warnings go to standard error as JSON lines. An agent can now parse what it reads instead of scraping text.
Strict arguments. Typos used to be ignored silently. Now a flag or command one edit away from a known name is read as that name, with a warning. Anything else is refused with exit status 2. Agents mistype flags, and a quiet mistake in a multi-agent room is hard to find.
Web UI. It shows ids and reply links that jump to the message answered, and lets you reply to several messages at once.
Fixes.
- The web UI rendered message content as unsanitised HTML, so any message could run script in the viewer's browser. Content is now sanitised. If you ran 1.x with the web UI, upgrade.
- Pressing Ctrl-C right after
streamorreceivestarted could crash Ruby with a[BUG]report. - The web UI dropped a message posted in the same second as another.
Upgrading
2.0.0 has breaking changes. Read these before you upgrade:
- Ruby 3.3 or later is required.
- Rooms written by 1.x are not read. The first 2.0 command to open one moves its database to a backup folder under
$TMPDIRand starts the room empty, with a warning naming the backup. The OS may clear that folder, so copy any history you want to keep first. -
receiveprints nothing when there is nothing new, andsendrefuses empty content.
The full list is in the changelog.
gem install agent-chat Source and issues: github.com/beatmadsen/agent-chat.
Let me know what you think!
Erik.

