Your agents. Your hardware. Your rules.
Hearth is a pre-release platform for running AI agents on servers you own — in your garage, your closet, your spare rack — instead of renting them by the token from someone else’s data center.
We think that’s a better deal for you, and that it’s about to matter a lot more than it does today.
Why roll your own
Cloud AI is convenient right up until it isn’t: your data gets hoovered up by the service owners, the model changes under you, and your entire history gets locked into a contract you don’t control on somebody else’s servers. That’s a fine trade for casual use. It’s a bad trade for the assistant that reads your email, watches your code, and remembers your life.
Agents are supposed to make your life simpler — fewer tabs, fewer chores, fewer things you have to remember to do yourself. They can’t really do that if you don’t trust what they’re connected to, or if the only way to get one is to hand a stranger your data on a subscription. So we’re building the other option: a stack you run yourself, small enough to understand, open enough to fix, and yours the way a tool in your own garage is yours.
We’re not going to pretend this is effortless. Self-hosting takes a little more care than clicking “sign up.” We believe that’s the point, not the problem. We’re building Hearth to show that the time investment pays off in something you can run, on your own, that works how you want it to.
Two ways in
Everything we learn, we publish as we go: the build log, the benchmarks, the configs. That’s the on-ramp if you’d rather roll your own, on hardware you already have. It’s how we got here, and we’d love the company.
If you’d rather skip ahead, the builds below are the easy button. They are hardware we’ve put together and tested ourselves, set up with the same stack we run — language models, speech-to-text and text-to-speech, all on the box, so you can talk to it — with validated, measured numbers.
The builds
Hearth is pre-release. Reserving costs nothing and commits you to nothing; it tells us which build to make first and who to write to when it’s ready.
Hobby
$2,000
A palm-sized box running your own agents, voice included.
- Memory
- 32 GB, or 16 GB as a lower-cost option
- Language models
- up to 32k context, ~15 tokens/s (measured on 16 GB)
- Voice
- speech-to-text and text-to-speech, on the box
No image generation on this tier yet — there are models that will run here, we just haven't built it. In the meantime, nano-banana-2 is free and worth a look.
Reserve a Hobby kit →Developer
$6,000
A mid-tower workstation that runs big models with long memory.
- Memory
- 64 GB system, plus 2 × 32 GB GPUs
- Language models
- up to 256k context; ~80 tokens/s on a fast 35B model, ~20 on a dense 27B, slower as conversations get long
- Voice
- speech-to-text and text-to-speech
- Pictures & video
- image generation, and 5-second video clips (Wan2.2)
Production
Coming later
Rack-scale, for teams. We know what it looks like; we'll build it once the first two are in people's hands.
Tell us you're interestedFollow along
Want to watch it come together? Join the list and we’ll write when there’s something worth reading. Nothing else, no spam, unsubscribe whenever you want.
Read the build log
We’re writing about all of this as we go — the ideas, the history, the occasional warning label — over on the blog. Start with “A Journey into Agentic AI”, our five-part series on what it means to interview, warn, remember, evolve, and finally trust an agent you built yourself.