[feature request] Remote Claude Code

This is a dream feature - but with the remote agents released by Cursor and Augment, maybe it's not so distant a dream! And since I asked for an internal todo list and, to my delight, it happened, I thought, why not ask again?

My dream remote claude code setup:

• I specify some instructions for how to duplicate the repo. Since I code in Rails and there are things like the credentials master key and yarn packages and gems that get installed after cloning, I probably need to specify some kind of script to help set up the test instances. Perhaps it involves pointing it to a Claude.Dockerfile that sets up the image, and then a .claude/.remote that tells it how/where to deploy the Docker. Up to you, I'll adapt, so long as it can run Rails.
• Then, once that is set up and working, I know that I can do something like claude --remote implement-feature-x and it will spin up a remote image and do the work there in a branch without user interactivity. When it's done, it submits a pull request to the repo so I can review it, and notifies me that the job is done.
• I then review the pull requests and, if they are good, I accept them, if they need changes, I make a comment and then I tell a claude session to resume working on implement-feature-x remotely so it can apply my comments, and then I accept them. Or if it is garbage, which will happen sometime and is no biggie, I discard the PR.

Benefits:

• Instead of having to sit there and watch Claude code do its thing in series, and basically watch one instance of Claude Code implement one thing at a time, I can e.g. spend a day creating various implementation plans, then in the evening I launch them all in triplicate, and then in the morning I review a bunch of pull requests and apply the best ones. Then when I'm done I work on a bunch of new plans, and before checking out for the day, I hit send, and off we go again.
• This would enable me to "scale my compute effort" massively. I can focus my time on the work Claude can't do (decide what to do, decide the desired architecture for it, and review that the PRs don't break the consistency of the codebase), and, I think, multiply my output.

You know this is the future, right? And Claude Code is ideally positioned for this because it already doesn't rely on IDEs and the like to manage the flow.

I hope you've got this in the works already and this helps inform the direction/vision! <3

Thanks for all the great work either way!!