Not long after I published copse, a worktree tool I built for my tasks, Filippo, the CEO of Kosuke, emailed me. He'd seen copse and liked it. Then he told me about Kosuke, and the first thing what attracted me was "No more getting stuck in worktrees and local dev environments."
He had a point. It was not something I thought while making copse.
copse makes the worktrees. When a ticket touches the frontend and the backend, it puts a fresh copy of each repo in one folder, on the right branch, and I run Claude Code from there. That takes seconds now.
But then when I want to see the changes done, git doesn't copy your .env varialbes into that new worktree, so I had to install the packages, add the env vars, make sure the port doesn’t clash with default one before I can even start the app. When my claude code says the code is done, I need to start the app and check to see whether done means done. And if the task runs long, my laptop stays open until it finishes. Though, I use cmux, so a pile of terminals doesn't bother me. Everything else on that list does.
I built copse to make my workflow faster, but then I realized there is something else eating up my time. That's the slow part. It was never the folders or worktrees.
So I tried Kosuke#
Right after you sign up in Kosuke, you can connect a GitHub repo and press Deploy. Right after that, an agent reads the repo, works out what the app needs to start, which services, ports, env vars and setup scripts, and saves all of it. I was testing with a nextjs app, and it deployed it perfectly without me doing anything except clicking the button deploy.
It was like T3 Code, but with a server and steroid. Every chat session gets its own branch, its own sandbox in Kosuke's cloud, and a live preview of the app sitting next to the chat. Claude Code runs in the browser on my own subscription, and whatever it changes shows up in the preview straight away. I can hand it a task, shut the laptop, and come back to a branch and a running app.
What surprised me more was the testing. On a small event site I'm building, I asked for the countdown to turn into a live "happening now" card on the day itself. The agent wrote it in about two minutes, then spent four testing it. It opened the site as it would look the night before, during the day, after it was over, and on a phone, and took a screenshot of each. I do not know if they are using any skill or anything else to make so many tests perfectly, but I love it!
What do I think!#
I'd refer this to anyone running several agents at once at it, since every session gets its own sandbox and nothing fights over ports. Same for anyone who wants to hand off a task and close the laptop. This is so easy that even a team’s PM can spin off a sandbox and make those changes instead of making the prototype in lovable, without asking an engineer to set anything up.
I would say the pricing is great with the things they are offering. Starter is $20 a month with $25 of sandbox credit, and Pro is $190 with $250. Sandboxes are billed by the second while they're awake, from $0.51 an hour, and craziest thing is you can use your own Claude or Codex plan.
It's GitHub only for now. Most of my work lives on GitLab, so that was the first thing I asked Filippo for. Even so, out of all the agent tools I've tried, this is one of the most useful, and I'm glad Filippo reached out.
If you want to try it, sign up and enter EHSANUR2026 in the "Promo code (optional)" field right after. You get $30 in credits instead of the usual $10.