Mantlecore - Managed cloud for open-source AI agents

I'm Yaz, the co-founder of Mantlecore.

This started with a pattern we kept noticing. Open-source AI agents like Agent Zero and Hermes are genuinely incredible, but almost every friend we showed them to said the same thing: "This looks amazing, I'd love to use it..." and then got stuck before they ever gave the agent its first task.

- For non-technical folks, going beyond running something locally means servers, Docker, networking, SSL, security, updates, a wall of infrastructure before any actual value shows up.
- For technical people, deployment is just the start of the problem. Agents accumulate memory, files, config, and integrations over time. One experiment or one bad update can wipe out something you've spent weeks building.

So people either don't start, or they stop experimenting because the cost of breaking things is too high. Our goal with Mantlecore is to remove the whole layer. One-click deploy into a kernel-level isolated VM, with automatic and manual snapshots so you can update, tinker, and break things fearlessly. because you can always roll back in seconds.

Bring your own API keys or OAuth subscription for you favorite LLM provider or use our managed LLM service, we handle everything and it works right away.

The real mission is bigger than hosting, though. We want more people actually using open-source agents, spending their energy on use cases and exploration instead of babysitting infrastructure. If we make that easier, the whole ecosystem gets more interesting.

Would genuinely love your feedback, your favorite agents, and the things you wish just worked out of the box?

the kernel-level isolation per agent is genuinely a smart call, basically lets you treat each one like a sandbox you can actually trust to mess around in.

Hi everyone, I’m Vitalii, the other co-founder of Mantlecore.

I wanted to share a little about the part of Mantlecore I’ve spent most of my time working on: isolation, persistence, and the infrastructure that users hopefully never have to worry about.

From the beginning, we wanted to combine two things that do not naturally fit together: a serverless experience and a stateful agent runtime.

Serverless means you click a button, your agent starts almost immediately, and you never have to think about servers or machines.

Stateful means your agent keeps accumulating memory, files, configuration, integrations, and work that may become more valuable over time. Losing that state is not an option.

Normally, infrastructure is optimized for one or the other. Trying to achieve both shaped almost every technical decision we made.

It also took us a few attempts. We rebuilt major parts of the architecture three times, questioned several of our early assumptions, and spent more time than I would like to admit watching agents start, stop, break, recover, and move between machines. It was challenging, but honestly, it was also a lot of fun.

Because AI agents can execute arbitrary code, we wanted every agent to have a real hardware-virtualization security boundary, rather than relying only on container isolation. The challenge was making that boundary lightweight enough that launching an agent still feels instant instead of feeling like you are booting a traditional server.

We also designed the agent’s state to exist independently from the machine running it. Its files and data are not trapped on a particular host. This means we can snapshot an agent, roll it back, restore it elsewhere, survive infrastructure updates, and even move a running agent between physical machines with roughly half a second of downtime.

None of these ideas are completely new on their own. The difficult and interesting part was bringing them together into a runtime where compute is disposable, state is durable, and the user never has to worry about any of it.

For us, that is the real goal of Mantlecore: all the engineering complexity stays behind the scenes, while someone can launch an agent with one click and focus entirely on what they want it to do.

This foundation is also what gives us room to keep improving efficiency and reliability without disrupting the user’s data, agent, or experience.

We’re still early, and there is a lot more we want to build, but I’m proud of how far the system has come and excited to finally put it in people’s hands.

this looks great honestly, especially the one-click deploy part, that's such a pain point right now. one thing that would be really useful though is some kind of scheduling or cron-like feature so you can kick off agents on a recurring basis without having to manually trigger them every time. would make it way more practical for research workflows

Deployed Agent Zero in like 30 seconds, which honestly sold me on its own. The kernel isolation gives me way more confidence to actually experiment without worrying about breaking other agents.

The kernel-level isolation per agent is a really thoughtful call, it's the kind of detail most platforms skip and then regret. Love that snapshots make it feel safe to actually experiment instead of babying everything.

the kernel-level VM isolation is a really sharp call—way more confidence-inspiring than just slapping agents in containers and hoping for the best

The kernel-level isolation is a nice touch, gave me confidence to mess with Agent Zero without worrying about breaking anything. Snapshots actually kicked in fast too.