Show HN: I built an OS that is pure AI

The WASM sandbox + microkernel architecture is the right foundation for agent isolation, but it surfaces a problem that most agent systems don't address until it's too late: if every agent is ephemeral and generated on demand, your audit story is basically nonexistent. An agent crashes, does something unexpected, produces wrong output — and you have no record of what it was running, what context it had, or why it made the decisions it did.

With traditional software you at least have a binary you can inspect. Here the "program" is a generated Rust module that may never exist again in the same form. That's a genuinely hard problem for any use case where you need to reproduce or explain behavior after the fact.

The community agent store helps for sharing, but it also raises the question: when a downloaded agent does something wrong, who's responsible and how do you trace it? This feels like the unsexy infrastructure problem that'll define whether Pneuma can be trusted for anything beyond personal use.

> Currently it runs as a desktop app on Linux, macOS, and Windows

Does that mean it's not actually an OS?

The WASM sandbox per agent with typed ABI is a really clean isolation model. Most agent runtimes I've seen just throw everything into one process and hope for the best.

Curious about the IPC layer between agents. When agents communicate with each other, is there any access control on what one agent can request from another? Or is it open by default?

In practice the moment you have multiple agents that persist and talk to each other, you need some kind of permission model — otherwise one rogue agent can access everything through a chain of IPC calls. Even if each individual sandbox is secure.

The Rust + WASM combo is solid for this. We use Rust for similar reasons — memory safety matters a lot more when agents run autonomously.

I think this is the way things are ultimately going (software as ephemerally generated based on your preferences), but at the moment, I can't see myself using these generated apps. Maybe it's just an aesthetic thing—have you experimented with personalized theming?

It reminds me of: https://fabian-kuebler.com/posts/markdown-agentic-ui/ to an extent.

I had a conversation with Claude on it. I work as an artist and webdev, so excuse my lack of technical knowledge on the matter: https://claude.ai/share/8b40cf7f-ecc6-452f-a550-3d1e5416b3e5

I don't think generating an app would be the move forward. Rather it should be an agentic application that can be installed/added and the main agent controls it.

Building apps on the fly seems cool, But hits the frustration when they don't work at times, When building as a side project we tolerate it because we are in building phase, But when using we might not tolerate the issues.

Finally the security also plays a major role, So instead of building apps at runtime, We need agents and skills as portable objects.

It's a fascinating paradigm – as though it dropped here out of some future.

However, I'd never use it. Reason: having everything dynamically created JIT means there's a trade-off somewhere else; in this case, it's energy consumption. I'd rather "cache" applications (ie. have them as installable software, which has been built just once) than endlessly recreate software from scratch each time, with all the energy usage that entails.

Seems like a cool idea—I'm always game to try new operating systems! Recently been playing around with Playbit (https://playbit.app/)

Pneuma reminds me a little bit of Yansu AI, a project I saw recently on HN that proactively builds apps with AI (https://yansu.app/)

One critique is that I think it might be more efficient to just make use of existing software for some of these things rather than generate them on the fly (for something for example like a text editor), but I guess that will depend on how costly it is to run AI programs in the future (if it becomes much cheaper, then I could see your concept being taken up by more people)

I see a lot of interest and good questions here. I've just published a blog post with more details for anyone interested https://pneuma.computer/blog/how-pneuma-generates-compiles-a...

> the UX

For this to be practical you really need to build things iteratively and asking the user questions, rather than attempting at one shot half assed ideas

Also most people are better explaining details with voice rather than text.

how are you handling networking in the sandbox? like if i generate something that needs to hit an API. does it get full network access or is there some permission model

Cool, curious though long did it take to build it?

How did you solve the problem of hallucinations?

Sounds expensive to sustain and inefficient

This is really cool. It reminded me of my first thought for an AI app was a No-Interface crm/order system. I was talking to my bro in sales and all he ever does is complain about what a pain computers are. This takes that idea to the next level.

llms were not quite there at the time maybe ill revisit.

That's a very interesting and innovative idea. Congratulations, well tuned!

how does pneuma deal with persistence of user data?

[dead]

[dead]

[dead]