claude --worktree is great for isolating a single session, but you're still juggling terminal tabs manually — starting each one, switching between them to check progress, reviewing diffs by hand.
ChatML is basically the management layer on top of that. You get a dashboard where you can run 3-5 agents at once, each in its own worktree, and see what they're all doing without tab-switching. It also handles the worktree lifecycle for you (creation, branching, cleanup), has a built-in diff viewer with code review, tracks cost per session, and lets you open PRs directly.
Closest analogy: running containers manually vs. having Compose manage them. Same underlying primitive, but the orchestration matters once you're doing it regularly.
Yes. Each session get it's up to 5 Terminal sessions. So if I want to test a particular session I was just spinning servers for that session. I had to prepare the code in development to attempt to use multiple ports in case we had another session running.... But I had no problems.... I am writing a feature that you can configure init and tear tasks, so you can do things like "npm install", "docker-compose up", etc....
I had it working, but needed more test, so I commented out from the initial release. I will be adding that shortly after some more testing.
This article shows the results of our experiment letting AI build everything. We need this tool for our team and we ask Claude to build for us. Took us 45 days and 750+ PRs merged all running in isolated parallel claude sessions.
Yes. In November/December timeframe when Opus 4.5 came out, I realized I was trusting it to write the code. I was doing light reviews. I had tried a few MacOS products that does this orchestration, but I was needing more and they weren't answering my emails.
It was then I realized I had to build ChatML. The first commit was on January 17, so it took about 45 days to get this out, so people can use...
One weekend I closed 120 PR's with on average 10-15 Worktree sessions at any given time. The hardest was to learn how to context-switch in your mind. I built some MacOS Notifications and audible alerts to catch my attention when new plan was proposed in a session, an agent asking questions, etc....
It worked really well for ChatML, since I developed ChatML with ChatML. 750+ Pull Request done 100% by AI.
ChatML is basically the management layer on top of that. You get a dashboard where you can run 3-5 agents at once, each in its own worktree, and see what they're all doing without tab-switching. It also handles the worktree lifecycle for you (creation, branching, cleanup), has a built-in diff viewer with code review, tracks cost per session, and lets you open PRs directly.
Closest analogy: running containers manually vs. having Compose manage them. Same underlying primitive, but the orchestration matters once you're doing it regularly.