In a Jetbrains IDE, for example, you check a devcontainer.json file into your repository. This file describes how to build a Docker image (or points to a Dockerfile you already have). When you open up a project, the IDE builds the Docker image, automatically installs a language-server backend into it, and launches a remote frontend connected to that container (which may run on the same or a different machine from where the frontend runs).

If you do anything with an AI agent, that thing happens inside the remote container where the project code files are. If you compile anything, or run anything, that happens in the container too. The project directory itself is synced back to your local system but your home directory (and all its credentials) are off-limits to things inside the container.

It's actually easier to do this than to not, since it provides reusable developer tooling that can be shared among all team members, and gives you consistent dependency versions used for local compilation/profiling/debugging/whatever.

DevContainers are supported by a number of IDEs including VSCode.

You should be using them for non-vibe projects. You should DEFINITELY be using them for vibe projects.

The downside for me and the main reason I do use vms less than I did a few months ago is that I need my agentic coding tools to use development tools a lot. And those tools need a lot of resources. And I have those resources on my laptop. Which is a nice mac book pro with plenty of RAM and 16 CPUs. I can run vms on this thing without issues of course. But tools just run a lot faster when I run them outside those VMs. And agentic coding tools run builds all the time. We're talking some really non trivial time savings here. Watching qemu build a thing for 10 minutes that I know should build in 45 seconds is painful. Especially if it happens over and over again.

The trick is doing sandboxing without performance impact. And very soon you'll also want to be able to run local models. I've been toying with the latest qwen and gemma models on my laptop. I haven't gotten around to doing coding with those just yet. But apparently they aren't completely horrible at it. That won't work on most cloud based vms. Unless you get a really big and expensive one. You could actually make that work if you only use them for a few minutes.

You don't give your GH keys, email credentials and ssh keys to a coworker. They have their own accounts with scoped permissions. Need them to read an email? Forward it. Need them to work on a repo? Add them as a contributor and enforce the same branch policies you would for any human.

There are still risks, but they're similar to delegating work to humans, so it's up to you how much access and trust to give.

All that said, no way in hell I’m giving either access to production databases or environments.

No, 'safety oriented' lab has a clause like that which can't be revoked historically. Anthropic, like the majority of 'don't be evil' firms is apart of the great masquerade.

That ability to transparently start a container and connect it to the SSH pipr is useful for isolation methods for coding agents involving containers and I imagine it would work equally well for things like Firecracker VMs. It's made my experiment working with an "immutable OS" (Universal Blue based) much more ergonomic. Also, it's the only way I've found to let Zed run remotely inside a container without having the container run a ssh server.

> The actual development happens on a rented server

Why not Hyper-V or libvirt/KVM? VM escapes aren't a thing in real life (or VMs from hyperscalers wouldn't exist), so why deal with additional cost, latency, and third-party trust when you could just run it yourself?

Almost back to point A then. If the server is compromised in some way, they can use (not copy) OP's keys and use them to clone repos/inject code/etc..

Treat it as a colleague, making PRs that you review.

All you need is a separate limited user account on your computer. Multi-user Unix-y systems were designed for this kind of thing for decades.

My entire development environment is literally just "sudo".

claude-cli executes whatever is in the BROWSER env variable to open your browser at a current URL, so I pointed it at a simple shell script that writes the URL to a named pipe which is mounted into the container. The sandbox tool outside of the container is reading from that named pipe. When it receives a URL to open, it pops up a confirmation dialog with info about the URL. If you accept, it opens it in your host browser.

The second step is, the callback URL after you sign in on the claude website wants to connect back to a port on localhost to complete the sign in. If the sandbox is being run with host networking mode, this just works fine as claude cli has already opened that port so it's listening on the host network. However if it is not running in host networking mode, the sandbox tool figures out what port it need to listen on from looking at the URL, listens to it, and when it is hit, it just podman exec's curl inside the container to complete the callback.

When they try to open a browser, they also print the URL to the console. Open that in your browser and go through an authentication flow; it'll end forwarding you to a localhost URL like http://127.0.0.1:8080/authorization-code/callback?code=XXXX&... which will fail.

Copy that callback URL, connect to your VM/docker container, and curl it.

The curl stage requires the agent make a call to auth.whatever-vendor.com so if it fails at this stage, check your VM/container network settings. And make sure you quoted the curl right so the & wasn't misinterpreted.

It'll then save a file at ~/.codex/auth.json or ~/.claude.json or similar, so you won't need to log in again. The secret in this file will periodically rotate, so you need to mount it read-write not read-only.