Caelum Stack is a desktop assistant you drive with your voice. Wake it with a word, ask in ordinary words, and watch it find your documents, tidy your folders, look things up and answer out loud — without your hands ever leaving the desk.
Runs on your machine. Windows first — macOS and Android to follow.
Listening
Find last year's electricity bills and put them all in one folder.
Found nine, January to December. They're in Documents / Bills — want them renamed by month?
Everything else is a text field you type into. This is a name you call. It stays quiet until it hears you, wakes for three minutes, and gives you the room back when you're done.
An always-on detector listens locally for your assistant's name. Nothing is recorded, nothing leaves the machine, and it only opens its ears once the word actually lands.
Choose one from a short list when you set it up. The list is fixed for v1 for a reason worth knowing: each name is recognised entirely on your machine, and that takes a small model of its own. Whichever you pick is the only word that wakes it, so the rest of the room can carry on talking.
Once it hears you, it stays awake for a full conversation. No word before every sentence, and it goes quiet again by itself.
The microphone is suppressed while the assistant is speaking, so it never answers itself or loops on its own reply.
Switch between automatic wake-word listening and a manual microphone button whenever the room is loud or shared.
This is not a chat window that gives you instructions to follow. Ask out loud, in ordinary words, and it goes and does the thing itself.
Ask for the document you only half remember. It searches your folders by name, by date, or by what is written inside — then opens it, moves it or renames it for you.
Rename two hundred photos by the date they were taken, sort a download folder, convert a batch of files. The things you would do one at a time, done at once.
Work, the house, the association you help run — each folder gets its own space and its own conversation, so nothing bleeds between them.
Every change can be taken back. Step back one action or throw the whole afternoon away — nothing it touches becomes permanent until you say so.
Opening hours, an address, a phone number, what the news is saying. When an answer needs the internet, it goes and gets it instead of guessing.
Ask for something that takes five steps and it writes the list, works through it, and shows you where it has got to — rather than making you ask five times.
Nothing on your computer changes without you seeing it first. Approve each step while you are learning to trust it, and let it run once you do.
Tell it once where the invoices go, or what to call you, and stop repeating yourself every time you sit down.
Caelum Stack probes your hardware and tells you honestly what it can and cannot run, then downloads and serves whatever you choose. If nothing on the list fits your machine, that is a real answer too — and a cloud provider is one dropdown away.
| Model | Size | Fit | Speed |
|---|---|---|---|
| Qwen3 · 8B | Q4_K_M · 4.7 GB | perfect | 42 tok/s |
| Mistral · 7B | Q5_K_M · 5.1 GB | perfect | 36 tok/s |
| Llama 3.1 · 8B | Q4_K_M · 4.9 GB | good | 31 tok/s |
| Gemma 2 · 27B | Q4_K_M · 16.6 GB | marginal | 9 tok/s |
RAM, VRAM, GPU backend, cores and free storage are probed before a single model is suggested — so 'too tight' is a fact, not a guess.
Search Hugging Face, pick a quantisation, watch the download stream in, and serve it locally the moment it lands.
Every downloaded model is actually run to see whether it can call tools, and graded on what it did — not on what its metadata claims.
Tell it to release an idle model after a minute, five, fifteen — or never — and your RAM comes back when you walk away.
The wake word, the transcription, the speech and the search all run as local services on loopback. Nothing is recorded and no audio is ever uploaded. Where the conversation itself goes is then a choice you make, model by model.
Caelum Stack will not pretend that every machine can run a good model — that depends on your hardware, and it will tell you the truth about yours rather than sell you a fantasy. What it does promise is the choice. If your machine can run a model, you never have to hand your work to anyone. If it cannot, nothing about the product is withheld from you.
Serve it from the Models page and the whole loop — voice in, reasoning, answer out — stays on your hardware. No key, no provider account, no request over the wire.
Connect Anthropic, OpenAI, Google or anyone else and you get the same voice, the same tools, the same interface. Only the conversation leaves — your audio was already transcribed on your machine.
Switching between the two is one click, whenever you like.
Being straight with you: two things reach the network whichever model you pick — the update check, and the catalogue Caelum Stack reads to know which models exist. Neither one carries your conversations.
The app ships localised, the transcription detects what you actually said, and the assistant answers in a natural voice you pick.
Fully localised UI, switchable at any time.
Transcription detects the spoken language for you.
Natural female voices in English and French.
Caelum Stack is being tested by a small invited group right now, and we are not opening new seats yet. This list is who we tell first when we do — and who hears the day v1 ships.
Two emails, at most: one when the next round opens, one when v1 ships. No newsletter, never shared, unsubscribe in a click.
A small, invited group on Windows.
Wider testing, stability and polish across the voice pipeline.
The first public release, on Windows. macOS and Android follow.