Questions, answered plainly.

Software that turns a computer with an NVIDIA GPU, or an Apple Silicon Mac, into local AI infrastructure. It checks the machine over, works out whether a model will fit safely in the GPU’s memory, sets up everything needed to run it, and gives you an OpenAI-compatible API on that machine.

Three sorts of people, which is why there are three plans.

Home: someone with a gaming PC or a workstation who wants to run models locally without first becoming an infrastructure engineer.

Home Pro: the same person a year later, with a second machine, a server in a cupboard, and a laptop that wants to reach both.

Business: an organisation with GPU hardware it would like to run as shared infrastructure, with members, roles and an audit trail. That tier is being shaped with its first customers. See pricing.

Four, and we keep them apart on purpose.

Available means it is in the published build. Install it and you have it.

In progress means under way right now, and not in the published build yet.

Planned and Exploring are where the ideas live before they are built.

The roadmap labels every capability, so you can always check where something stands rather than taking a marketing page's word for it.

No. Kubernetes is under the bonnet, but day-to-day use is a handful of commands and an ordinary API. You never write YAML, and you do not need to know what a container is. If you do know Kubernetes, it is your cluster and Nodeau does not hide it from you.

Not at this time. The installer and the published distribution are public. The platform source stays with us for now.

Three commands. On Linux the machine needs a working NVIDIA driver already. On an Apple Silicon Mac you are ready to go straight away.

curl -fsSL https://get.nodeau.ai/install.sh | bash
nodeau install
nodeau quickstart

The install guide walks through each one, and the documentation has the full flow for Linux and macOS.

Roughly ten minutes of setup, plus the model download. The starter model is about 2.6 GB, so on a slow connection that is the long part.

Nodeau Home is free, needs no account and no card, and stays that way: one machine, one GPU, local inference, the model catalogue, the dashboard and the fit checks. Paid plans are about scale and collaboration, so several machines, batch work and teams. Remote management is in the published build: open the fleet console from any browser and see and operate the machines you own. Seeing your machines is free on every plan; changing one from there is what a paid plan buys. Model replication is planned, not shipped. See pricing.

Yes. nodeau uninstall shows you the plan and asks before it changes anything. Nodeau keeps track of what it installed versus what was already there, so things that predate it stay exactly where they are. Your downloaded models, your GPU driver and any Kubernetes cluster it did not create are all left alone.

If Nodeau adopted a Kubernetes cluster you already had, it removes its own objects and leaves the cluster itself alone. It did not install that, so it leaves it for you. If anything looks left behind, send us nodeau support bundle and we will sort it out.

NVIDIA cards on Linux, several per machine, each judged on its own merits. Also the integrated GPU of an Apple Silicon Mac through Metal. Nodeau has been run end to end on an RTX 5070 Ti, an RTX 3080, an RTX 5060 Ti and an RTX 2080 under Ubuntu 24.04, and on an M3 Pro under macOS.

On other NVIDIA cards it still works. Where it has a direct measurement for your configuration it uses it. Otherwise it works out a cautious estimate from the model’s own shape, holds extra memory back, and tells you the number is an estimate. What it will not do is quietly present a guess as a measurement, or promise that every card on the market is known to be fine.

No, deliberately. A working driver is a prerequisite. Installing or changing a GPU driver affects whether your machine boots to a desktop, and Nodeau will not make that decision for you. It checks the driver, and stops with an explanation if it is not working.

Nodeau stops before changing anything and tells you the driver is not working. Fix the driver first through your distribution, then run nodeau install again.

Not today. Native Windows is not planned. Support through WSL is on the roadmap and is currently Planned. The GPU path through WSL is genuinely different and still untested, so the installer says so plainly rather than half-working.

Not yet. It is detected and refused rather than attempted. It is on the roadmap.

Not today. It is listed under Exploring on the roadmap, because it needs a different runtime and memory model rather than a configuration flag.

Yes, on Apple Silicon, where it is qualified. There is some scope worth knowing before you start. Nodeau runs models on the Mac's own GPU through Metal, and installing needs no password and no driver.

What a Mac does not do: it runs standalone and cannot join a fleet, because it is a native execution plane rather than a Kubernetes node. Nodeau leaves your laptop as a laptop rather than installing a cluster on it. Batch inference is Linux-only and refuses with a clear error. And because Apple Silicon shares one pool of memory between everything, whether a model fits depends on what else is running: the same request can be admitted today and correctly refused tomorrow, and Nodeau tells you the numbers when it refuses.

Intel Macs are unsupported.

Yes, both ways, and they were genuinely different problems to solve.

Several GPU machines, one control plane, on mixed NVIDIA hardware. Each machine is measured and judged on its own merits. Part of Home Pro and Business.

Several GPUs inside one machine, where each card is observed, reserved and scheduled on its own. Nodeau picks the physical card, Kubernetes allocates that card, and the telemetry shows which one ran the work. Home Pro puts up to two cards in any one machine to work. On Home the second card is still detected and reported, ready for whenever you want it.

A curated catalogue, chosen to span the card sizes people actually own, from 8 GB up to 24 GB, and pinned to exact published files. The starter model is Qwen3.5-4B (Apache-2.0, about 2.6 GB), which is the compact model for 8 GB cards. Larger ones go up to a flagship that needs a 24 GB card or two smaller cards in one machine. Run nodeau model list to see what your installation can serve, and nodeau model info for what a model will cost on your card.

Running a model that is not in the catalogue is Exploring on the roadmap. The obstacle is knowing what it will cost before it runs, which is the part Nodeau exists to get right.

Because a good answer takes work. A model profile starts life as real measurements on real hardware, so adding a model properly means measuring it. Nodeau can fall back to a cautious estimate worked out from the model’s own shape, which is what lets it run on cards we have never touched, but an estimate is the fallback and not the goal. Nodeau always tells you which one you got, and where the number came from.

It compares what the GPU can actually give a workload, meaning the memory the compute API can address, minus what is already in use outside Nodeau, minus what Nodeau has reserved for other workloads, minus a safety reserve, against what that model needs in that configuration plus a margin. If the result is negative it refuses and shows the arithmetic.

The “what the model needs” number is either measured on hardware like yours, or estimated from the model’s own architecture with an extra margin added. The decision tells you which. That matters: an estimate carries more uncertainty, and you should know when you are relying on one.

Three reasons. Some of the card’s memory is not addressable by the compute API at all. Your desktop is already using some of it. And a workload that fits with nothing spare will fail as soon as anything else allocates. Nodeau plans against what is actually usable rather than the number on the box.

From their original publisher, over HTTPS, at a pinned revision. Nodeau does not redistribute weights. Every download is checked against a known SHA-256 before it is used, and a file that fails is quarantined rather than left where a later run might trust it.

No. The model runs on your GPU, on your machine. There is no Nodeau-hosted GPU involved.

Nodeau does not collect or transmit prompts or model output, and there is no telemetry. Your requests go to a model running on your own hardware over the loopback interface. That is true whether or not you have an account.

One honest caveat: if you point a third-party application at the endpoint, that application can do whatever it likes with your data. Nodeau cannot make promises about software it did not write.

No. The local endpoint binds 127.0.0.1 and nothing else. Not 0.0.0.0, not a LAN address, and there is no setting that quietly changes it. You can check with ss -tlnp | grep nodeau.

Yes. A key is generated locally with a cryptographic random source and stored with owner-only permissions. Requests without it, or with the wrong one, are rejected. Reveal it with nodeau auth show.

Yes, including streaming. Most OpenAI clients work by pointing base_url at http://127.0.0.1:8080/v1.

Not for inference. It needs the network to install, to download container images, and to fetch a model the first time. After that the model is cached locally and inference is entirely local.

No. Nodeau Home installs and runs with no account at all: local inference, the model catalogue, the dashboard, the fit checks. Nothing phones home and nothing is gated behind a sign-in.

An account is for the things that genuinely belong elsewhere. Which installations are yours, which plan is in force, and being able to look at that from a laptop with no GPU in it.

Your account. Sign in and you can see your organisation, the machines linked to it, and the plan each one is running under. You link a machine by running nodeau login on it and approving it in the browser.

It never sees your prompts, your documents or your results. It is an account system, not a way in to your hardware.

Nothing, on your machines. Your plan is a signed entitlement stored on the installation and checked locally, so inference, batch jobs and the local dashboard carry on regardless. A machine that never sees the internet again keeps exactly the plan it was granted.

Machines. Home is one person and one machine with one GPU, free. Home Pro is the same person with several machines they own, and it adds batch inference across them. Remote management ships: you can see and operate your machines from a browser anywhere. Model replication and scheduled batch are planned for it and are not built yet. Business is the next line along: an organisation with several people in it. Members, roles and single sign-on are what it is for, and those are in progress with design partners now. Your own limits on what a fleet may run, and a record of what changed, are already in the published build. Full comparison on pricing.

There is a dashboard. nodeau dashboard opens a read-only view of your models, workloads, GPU state and the decisions Nodeau has made, served by Nodeau itself on your own machine. It needs no extra install, it is authenticated, and it is bound to the loopback interface like everything else.

Reaching it from another machine is part of Home Pro and Business, and is on the roadmap.

Yes, with a paid plan. Hand Nodeau a file of requests and it queues the job, waits for a whole GPU to be free, runs it, and writes results back correlated exactly to your inputs. Cancel it and the card is released once the process is genuinely gone rather than when something says it should be.

Batch is part of Home Pro and Business. See pricing.

Nodeau refuses before it starts anything, and shows the arithmetic: what the card can actually give a workload, what is already in use, what the model needs, and by how much you are short. It also suggests what would change the answer, such as a smaller quantisation, a shorter context, or freeing the card.

A refusal is not a crash, and Nodeau's exit codes say so, which matters if you are scripting it.

Usually by changing one line. Point base_url at http://127.0.0.1:8080/v1, use the key Nodeau generated, and chat completions, streaming and model listing work as you would expect. Asking for a model your installation does not serve gets you OpenAI's own model_not_found error rather than a surprise answer from a different model.

It is built for running your own models on your own hardware: a workstation under a desk, a machine in a cupboard, a few boxes a small team shares. That is a real job and Nodeau does it.

Two boundaries worth knowing before you build on it. Nodeau does not fail over on its own. It decides where a workload runs and leaves it there, so if a machine stops, what it was running stops with it. It also assumes the people with access to a machine are trusted, so it is built for teams that already share the hardware rather than for isolating strangers on one box. If you need a multi-tenant service with an uptime guarantee, that is a different kind of product.

Yes, and on purpose with two unlike machines rather than two matching ones. Nodeau was installed from a published release onto a second machine with a different NVIDIA card, and the two now run as one cluster: each watches its own hardware, checks its own copy of the model, and gets judged on its own merits. It has been through reboots, power cuts and a card that was already busy.

We would also love to see Nodeau on more hardware. That is what testers and design partners are for. If that sounds like you, say hello.

No, and it is worth saying plainly, because running on several machines invites the assumption. Nodeau chooses where a workload should run and then deliberately leaves it there. Spotting that a machine has died and rescuing the workload elsewhere is a different job, and it is not built. Assume an endpoint stops when its machine stops.

It should fail safely: Nodeau stops, tells you which stage it stopped at and why, lists what it had already changed, and states what it left alone, meaning your driver, Secure Boot, your bootloader, your partitions and your personal files. Re-running nodeau install is safe.

Run nodeau support bundle. It writes one .tar.gz with no API keys, no credentials, no prompts and no model output, and it uploads nothing, so you attach it yourself. If the CLI will not run at all, send ~/.local/state/nodeau/install.log instead. Email either to founders@nodeau.ai.

Please don’t. If you have to repair Kubernetes or container plumbing by hand, that is a bug we need to see. Send the bundle and let us fix Nodeau instead.

Yes. Each machine reports its own hardware and checks its own copy of the model. When you deploy something, Nodeau considers every machine, works out where it will really fit, and keeps the reasoning for the ones it passed over.

It is part of Home Pro and Business. The single command that adds a machine to an existing installation is what we are working on now. Until it lands, we set it up with design partners directly.

Yes, and that was rather the point of testing on two unlike cards instead of two matching ones. An RTX 3080 and an RTX 2080 in the same cluster are measured and weighed up separately, on what each can really give a workload. Nothing in Nodeau treats “a GPU” as a unit of capacity.

The roadmap sets out exactly how far this has been taken.

The machine side is there: several machines, one control plane, one place that decides where things run. So are your organisation's own limits on what that fleet may run, and a record of what changed. Several people with their own roles, and single sign-on, are In progress as Business capabilities, and design partners are shaping them now.

Yes, today. Your organisation can set its own limits on what its fleet may run, inside what your plan already allows: how many workloads at once, how many graphics cards, how many batch workers, which models, which cards and which machines. Set them with nodeau governance on a machine, or from your account on the web with Home Pro or Business. Lowering a limit never stops anything that is already running. Roles for several people are In progress; the roadmap tracks them.

Audit, yes: your organisation can see what changed and who changed it, including what was refused, kept for a year. Organisation-level access control, with several people each in their own role, is In progress. No dates.

Policy-controlled overflow to cloud capacity, with owned hardware first and cloud when policy or capacity requires it, is Exploring. It is a direction rather than a plan.

If your team has GPU hardware sitting under desks or in a rack, we would like to hear about it. Become a design partner.