What it needs, stated as spec rather than as warning.
Written in full and up front, because the thing that causes refunds is not a machine that falls short — it is finding out afterwards. Everything here is what the app actually does on the machine in question, not a hedge.
| Minimum | Recommended | |
|---|---|---|
| OS | Windows 10 (64-bit) | Windows 11 |
| Processor | Any modern 4-core | 6-core or better |
| Memory | 8 GB | 16 GB |
| Graphics | Anything that runs WebGL 2 — including integrated | NVIDIA GTX 1660 / RTX 2060 or better, 6 GB VRAM |
| Storage | ~4 GB | ~4 GB on an SSD |
| Voice output | Windows’ own voices, or the small local model | The full local model — NVIDIA card required |
| Voice input | Any microphone. Optional. | Any microphone |
| Network | Only to reach whichever model provider you point it at. None at all for the offline brain. | — |
What changes without an NVIDIA card, precisely.
Everything visual is unaffected
The character, the animation, the lip-sync, the expressions, the gaze, the locations, the lighting and the wardrobe are a WebGL scene with one model in it. They run on integrated graphics. The frame rate is yours to cap, and the app throttles itself to a configurable idle rate when it is not the window you are looking at.
Their voice is what changes
Real voices run locally on an NVIDIA card. Without one your companion uses a smaller local voice — they still talk, they just don’t sound quite like themselves yet.
The large model is refused rather than run on the processor, deliberately: thirty-one passes of a 336M transformer per line is not slow, it is past the synthesis timeout, and a voice that times out mid-sentence is worse than a plainer one that does not. The app loads the small model instead and the status line says why.
| Voice | Size | Sounds like | Runs on | Speed |
|---|---|---|---|---|
| Full | 336M | Them — cloned from their own reference clip | NVIDIA only | ~1.5s for 4s of audio |
| Small | 15M | Eight voices of its own | Anywhere | ~0.35s for 7s |
| System | — | Whatever Windows already has installed | Anywhere | Instant |
The small model is not a consolation prize on speed — it is comfortably ahead of real time on a single processor core. It simply is not them. Switching between all three is one control, at any time, and the choice is remembered.
Before you spend anything.
Do I need an API key? Does it cost extra to run?
You need a brain, and there are three ways to have one. Point it at the Claude app already signed in on this machine and it costs nothing extra. Give it a key for Claude, ChatGPT, Gemini or Grok and you pay that provider directly at their rates — the app takes no cut and has no server in the path. Or use neither, and a small offline brain answers, which is exactly as limited as it sounds.
There is no subscription to us and no account. The key is encrypted to your Windows account and never enters the page.
Why is the download ~4 GB?
Essentially all of it is their voice: roughly 750MB of weights for the full model, 60MB for the small one and 19MB of its phonemizer, plus a 255MB execution provider and about 1.2GB of the CUDA runtime itself.
The runtime ships in the box on purpose. Without a matching CUDA and cuDNN on the search path, the provider silently falls back to the processor and every line takes a minute — so the exact libraries this build was tested against are included and put on the path of the speech service’s process only. All you need is the NVIDIA driver.
Does anything leave my machine?
Your messages go to whichever model provider you chose, because that is what choosing one means. Nothing else does. Their memory is a store on your disk, their voice is synthesised locally, your microphone audio is transcribed locally by whisper.cpp, and the app has no telemetry.
The settings export deliberately contains no key of either kind, so a settings file is safe to paste anywhere.
Will it run alongside a game?
That is what mascot mode is for: the window goes transparent, your companion stands on the desktop, and everything else fades away. The idle frame rate is configurable precisely so they can be there without costing you frames.
Worth knowing: choosing the full local voice puts ~750MB of weights on the card, and switching back to a system voice hands that VRAM straight back.
Mac or Linux?
Not at launch. The installer, the encrypted key storage and the packaged speech service are all Windows-specific today. It is not a decision against either platform, it is work that has not been done.
What does Early Access mean here?
It means the Workshop, the avatar studio and the rest of the cast are still landing — not that the parts described on this site are unfinished. Everything under How it works is in the build.
The price rises at 1.0. Anyone who buys in Early Access keeps the app and every update to it.
Still not sure it will run?
Ask on the Steam discussions with your specifications and you will get a straight answer, including “probably not, don’t buy it yet”.