Roundtable

I run Claude Code with a lot of subagents and couldn't tell what they were doing, so I built an office for them. Try it in your browser, no install: https://kostakurta8.github.io/roundtable/ (a staged replay of a 24-agent session). Every subagent is a person. They walk in when they're spawned, sit at a desk while a tool runs, walk over to report back, and leave through the door when they're done. When one agent checks another's work you get a green CONFIRMED or a red REFUTED bubble. What else it does: - Rewind to any second. The simulation is deterministic, so clicking the timeline rebuilds the room exactly as it stood. - Tokens and a cost estimate per agent, cache included. - A Share button that turns a session into a GIF or a 1200×630 card of its numbers, rendered in the browser tab. Nothing is uploaded, and one switch hides all transcript text first. The part I didn't expect: counting the transcripts right is hard. Summing the usage field line by line over-reported my output tokens by 94%, and my subagents turned out to have written more output than all my main sessions combined. So there's a --stats command that counts it properly on your own machine. Stack: TypeScript, a small Node hub that tails the JSONL files and serves the page over a loopback WebSocket, and a GIF encoder written from scratch. Two runtime dependencies. Read-only, local, MIT. Run it on your own sessions (Node 22.12+): npx https://github.com/Kostakurta8/roundtable/releases/latest/download/roundtable.tgz Source: https://github.com/Kostakurta8/roundtable I've only used it on Windows. If you try it on a Mac or Linux, I'd like to hear whether it works, good or bad.