[_doctor_love]

My current power move in the age of AI: do nothing.

I had an idea like this and thought I could vibe code it, but then I figured someone else would care more and do it first. I was right!

This looks like a great app and I'm excited to try it out.

Free idea: I would like to be able to "jail" an agent inside a VM and send instructions to the harness from outside the VM to agent(s) installed inside. Ideally there is no Codex/Claude/etc. installed on the host.

More awesome: let me provision multiple user accounts inside the VM and restrict filesystem / network policy by user. Then I can have a dev agent, QA agent, etc. each with its own view of the work. That would be a powerful base layer for further automation.

Of course I should be able to provision various resources "attached" to the VM that agents can use on a permissioned basis; e.g., DB, queue, external volume, and so forth.

[mkagenius]

> Free idea: I would like to be able to "jail" an agent inside a VM and send instructions to the harness from outside the VM to agent(s) installed inside. Ideally there is no Codex/Claude/etc. installed on the host.

You can do exactly that with coderunner

https://GitHub.com/instavm/coderunner

[_doctor_love]

Nice! I will be sure to check it out. IMHO it's a good thing if there's a bit of competition.

[topgrain2]

> My current power move in the age of AI: do nothing.

This was also my strategy before AI. At some point in my late 20s or early 30s I all but completely stopped doing any development in my free time, because I was entirely over any fun I derived from coding per se (in truth, I'd never been that into it, I'd just been really bad at guessing what would or would not be worth spending time on) and, as they say, the "juice wasn't worth the squeeze" for almost anything that popped into my head that might be a nice program or script to have (like that xkcd chart about the payoff time for developing programs that save X minutes per week or whatever) or else it was something that wasn't necessary but might just be interesting or fun to have, but nowhere near worth the many hours it'd take to make it happen. If someone made what I wanted and released it, awesome! If not, oh well.

The big change with LLMs is now I can shit out little scripts and such in a few minutes and for pennies, maybe a couple dollars. I'm dragging old extremely-niche ideas out of mothballs because what would have been several weekends of work (most of these ideas would require lots of poking around unfamiliar APIs and documentation, not just immediately writing the thing I want) can now be done in a half-hour or less—or, at least, I can find out if something's going to be unworkable or too fiddly to screw with after all and should be completely and permanently abandoned, in minutes rather than hours.

[mikemcquaid]

Sandbox and separate user instead of VM but sandvault is good for this and does auto install (full disclosure: have made some PRs to it).

https://github.com/webcoyote/sandvault

[rao-v]

I threw something like this together w a simple browser front end, mostly because I like running mid to large open models but can’t trust them to not go insane. Will share at some point soon

[_doctor_love]

Look forward to seeing it!

[RossBencina]

> Free idea: ...

I have been thinking about this too. Is it not as simple as installing Claude in the VM and connecting via an SSH terminal, or if you want a GUI use VSCode with the Remote SSH extension, which will give you the file browser UI etc. Presumably you can extensions in the VSCode Claude/whatever chat extensions in the VM too.

[Willamin]

> as simple as installing Claude in the VM and connecting via an SSH terminal

I've done exactly this, and it works pretty well!

1. I setup a VM in UTM (but this could be any kind of containerization thing). I don't even bother with a non-root account in there (the agent has free rein to install packages, write files, etc). 2. I SSH into the container. 3. I install Claude or whatever there. 4. I setup git things in a way where I can push/pull to move code between the container and my host machine.

Upsides: the agent is isolated from the rest of my host system, only being able to read/write what I've explicitly handed to it. Downsides: the agent is isolated from the rest of my host system, so it's more limited in capability.

[_doctor_love]

UTM is undefeated for me, especially in combination with Vagrant. What I'm hoping for is a polished turn-key solution version of all this.

tart is also an option I like a lot, but it's macOS only.