Hoshi is in open alpha. There's no waitlist and no invite code: sign in, and a machine comes up for you.
We spent a long time building this behind a door and onboarding teams one at a time. That was the right call while the core loop was still moving under us. It isn't any more — machines, checkouts, and agentic sessions have been real and running daily for months. So the door is open.
Here's an honest account of what you'll find.
What works today
Everything below ships now. None of it is roadmap.
- A machine per person. A dedicated cloud computer provisioned at first sign-in — its own filesystem, its own shell, its own processes. Isolated, always on, never shared, and it keeps your files between sessions.
- An agent that lives on it. Not a service you call. It installs things that stay installed, runs servers that stay running, and picks up where it left off.
- Sessions that render interface. When the agent needs a decision it shows you a form, a choice, a table, or a chart instead of asking you to read a paragraph and reply in prose.
- Boilerplates and checkouts. An organization publishes how it builds once; every member pulls a working copy onto their own machine, provisioned automatically.
- Unattended work. Cron schedules, multi-step workflows, and webhook triggers. Start something at 6pm; read the result at 9am.
- Real documents. Word, PowerPoint, Excel, and PDF produced as actual files on your machine, previewable in the browser.
- Every surface. The web app, a native desktop app, mobile, and a full terminal UI over SSH — all driving the same machine.
- Your own models. Bring your provider keys; pick the model and agent per session. Keys resolve on the machine and never reach the browser.
What's rough
Alpha is a real word and we'd rather you hear this from us.
The docs trail the product. We ship faster than we write. If something isn't documented, it may still exist — ask, and we'll answer in Discord before the doc lands.
The API isn't fully documented. Personal access tokens, the machine API, sessions over SSE, schedules, and connectors are all there and usable. The complete reference is not written yet.
Pricing isn't settled. There's nothing to pay today. We haven't decided what Hoshi costs, and we're not going to pretend otherwise with a page full of tiers we haven't committed to. When that changes you'll hear it here and in Discord well before it affects you.
No uptime promise. Machines get restarted. Releases land during the day. Don't put anything on Hoshi during alpha that would ruin your week if it were interrupted.
You will find bugs. People are doing things with their machines that we didn't anticipate, which is the point, and it's also how we find the edges.
What we're asking
Two things, and both are genuinely useful to us.
Tell us when it breaks. A one-line report in Discord with what you were doing beats a perfectly-written issue three weeks later. Most reports get triaged the same day; a lot get fixed the same week.
Tell us what you're trying to do. The most valuable thing we get from early users isn't bug reports — it's watching someone try to use a machine for something we never designed for. Several of the features listed above started as somebody describing a workflow we hadn't imagined.
Early members of a small community have outsized influence over what gets built. That's not a marketing line; it's just arithmetic when the community is this size.
Start here
Sign in and your machine comes up in moments. Then:
- Read getting started — it's short.
- Pull a project or start in your personal space.
- Open a session and ask for something that needs a computer, not just an answer.
- Put something recurring on a schedule and see it finish without you.
If you want the argument for why any of this is shaped the way it is, that's in why your agent needs a computer of its own.
See you in Discord.