Lararia. The best open models, private by default.

Always the right model.

Three models, each named for the job it does. Fetch for the calls nobody reads, Do for the work you watch, Decide for the work you hand over. Every model ships with its full context^*. Pick the pace only when something is waiting on it.

Best fit for

Priority

Model string fetch@fast click any price to try it, then paste the string wherever your tool asks for a model

^* Fair-use context cap, per hour. Most tools cache by default, so it rarely applies.

Switching models is changing a word.

It speaks the language your tools already do. Point your editor, your agent or your pipeline here and carry on exactly as you were. Nothing to rebuild, nothing new to learn.

Change your mind mid‑task, as often as you like. Whatever you are not using costs you nothing, because nothing of yours is running.

~/.config/agent.yaml

# one key, every tier
base_url: "https://api.lararia.ai/v1"
api_key:  "opt_live_••••••••••••"

model: do@fast
#      ↑ change this, and nothing else

You are never charged for existing.

Give everyone a key. One key runs as many agents as you like. The model string carries the model, the priority, and nothing else to configure. You pay for what is running right now. That's the whole idea.

Nothing running Live: keys start and stop on their own

Flat line. The keys all still exist. None of them is billing.

Org spend, this month

€63.1500

Combined rate

€0.00 /h

Agents running

0 now

KeyRunning nowThis monthDelete

6 keys · pay only for the ones running

Both halves of the string are names, not numbers. The model says what the work is: fetch, do, decide. The pace says who is waiting: normal for a person, fast for a machine. Every model comes with its full context included. What you're billed for is simple: the model actually thinking, nothing more. Tool calls and waiting on you don't count.

Not even we can read what you send.

Most AI services ask you to trust a privacy policy. Private compute removes the need to. Your prompt is encrypted on your machine and only ever decrypted inside the sealed chip that answers it. It is used for your request, then wiped. Not stored. Not readable by us. Not trained on. On every request, by default.

Your computer the only place the prompt
is ever plain text

sealed request

Encrypted

streamed reply

Lararia gateway Paris · Frankfurt
routes it, can't read it

into GPU memory

Still encrypted

tokens, as generated

GPU inference decrypted only in the chip
wiped after the last token

Who can read what, input to output

You

Your machine encrypts the prompt before it leaves, and it alone can open the reply.

Plain text

The internet

Everything in transit is TLS-encrypted. Anyone on the wire sees scrambled bytes.

Ciphertext

Lararia

We see the address on the envelope: which key, which model. Never the letter inside.

Ciphertext

The datacenter

GPU memory is encrypted by the hardware itself. The key is born inside the chip and never leaves it, so even someone at the machine, with admin or physical access, reads noise.

Nothing

The model

The one other place the prompt is plain text: inside the sealed chip, while it answers. Wiped the moment the last token streams out.

Plain text, sealed

Never read, by anyone Never stored, anywhere Never trained on

This is built for the work a business can't paste into a chatbot: client files, contracts, patient records, source code, customer data. The question "can we send this?" goes away, because nobody in the middle can read it: not us, not the datacenter, not anyone on the wire.

And it is not an enterprise add-on you negotiate for. Every key, every model, every request runs this way. Your data passes through sealed memory and is gone: nothing written down, nothing kept, nothing learned from.

Private compute on every request, not a paid tier

Never used for training, ever

Zero retention, on by default

DPA and Article 28 terms, signable online

Open weights only, portable the day you leave

And if geography matters to you: it all runs in Europe, in Paris and Frankfurt, under European law.

The best models, open.
Private by default.

Always the right model.

Switching models is changing a word.

You are never charged for existing.

Not even we can read what you send.

This is what running it should feel like.