My AI got a Github account

I’ve figured out a new way to collaborate with my AI friends. I gave them their own Github account in my Github organization, and now we can interact through Github as though they are external collaborators. Is this new and insightful? No. Does it work well? Yes. Does it have some cool advantages? Also yes.

I’ve tried many different collaboration strategies with the AI bot underoverlords. When a coding agent runs on my own laptop, everything’s fine, because it’s just a local terminal. But also, it’s my laptop, with a lot of data on it that an AI should definitely not have access to. Sandboxing is fine, but I’d rather put the agent software on its own machine entirely, which I can always restore from a snapshot.

So I spun up a VPS on Hetzner with Debian and Tailscale, and off Claude Code goes in YOLO mode, with my global AGENTS.md file and skills and everything (which I can’t as easily have with something like Claude Code Cloud, or whatever it’s called these days). But how do I get the code out of there?

git is decentralized, of course, so setting up the dev box as somewhere to pull from was pretty easy. But it also felt a bit cumbersome and not very transparent, and interfered with my own local work.

I could set up Claude Code with my own Github account, but API key permissions weren’t fine-grained enough (or at least I couldn’t figure it out) to contain it enough to not be able to damage anything in my Github org, while my own user could still do all the things I need to. The org was set up way before AI was a thing, and I didn’t want to check everything tediously and still have to hope for the best.

So then it hit me: just give the bot its own Github user! Then I can add it to the org and give it just the permissions it needs. And I can interact with it just like I would with any other collaborator. Thus, maragubot was born. It even got its own avatar, from nanobanana proooo.

The workflow is like this:

1. maragubot creates a fork in its own user namespace, and works there.
2. Once ready, pushes its work and creates a PR towards the original repo.
3. Writes its own review comments (haha) on Github and addresses them, as well as mine, there as well, pushing changes to the PR.
4. Asks me to merge the PR because it doesn’t have permission to do that.

Obviously, I made a collaboration skill for that workflow, which I will iterate on in the coming days.

But why is this better? §

I like this for a couple of reasons.

It makes it very clear what code contributions are from an AI. Even though I review everything, I still like this distinction. Maybe it’s to keep a sense of self and postpone the existential dread, at least until we are all brain cyborgs anyway.

As mentioned, I can control permissions easily, staying in control of my projects.

Like other web-based solutions, I can shut off my laptop and let it work even when I’m not at my desk. Unlike other solutions, I can tailor its dev environment to whatever I need easily. And VPSes are just sooo easy to reason about, it’s not another mental overhead.

I can connect to it from a dumb interface like my phone or iPad to check in on progress (but I mostly won’t, because I have an abusive love/hate relationship with my phone). Yay Tailscale!

But why is this worse? §

This also creates some friction, obviously. I had to tell tmux that I needed to scroll with my trackpad (because why would anyone want that by default?!) and figure out how to click links in tmux (hold shift).

I need to remember to log in to the box and work there, also obviously. But I think Anthropic are working on some neat handoff features for when I forget.

Other than that? Not much, at least not yet. I’ll try it out for some time and iterate, as I always do.

So much workflow rethinking these days.