AI Stays Local

For developers and power users

A local model is a local process. That makes a few things possible that a hosted API cannot offer: no key to leak, no per-call cost, no request leaving the machine, and a filesystem that is simply there.

TerminalLocal
ask "which of these contracts auto-renew?" --folder ~/Clients/2026
reading 34 documents on this machine
Three agreements renew automatically unless notice is given:
Nordwind GmbH - 60 days Kaplan & Ruiz - 30 days
Delacroix SARL - 90 days
cited: Nordwind-MSA.pdf p.7, Kaplan-Ruiz-SOW.pdf p.2, Delacroix-cadre.pdf p.11
ask "draft the notice for Kaplan" --model large --allow-cloud
this request would leave your computer. review and approve? [y/N]
A sketch of the planned command line. The commands below are a design proposal, not a released interface.

Command line

Run a prompt, query the document index, or pipe a file through a local model from a shell, on the machine that already holds the data.

Private beta

Local API

A loopback HTTP endpoint so your own scripts and editors can reach the same models, with no key to leak and no request leaving the machine.

Planned

Structured output

Constrain a response to a schema you supply, so the output can be consumed by software rather than read by a person.

Planned

Model routing

Declare which class of task goes to which model, and whether a task is permitted to leave the machine at all.

Planned

Supervised execution

Multi-step tasks that stop and ask before any step that writes a file or opens a connection, with the pending action shown in full.

Planned

Engine adapters

Inference engines sit behind an adapter, so the runtime can adopt a better one without the product being rebuilt around it.

Prototype

Where this is honest about itself

None of the above has shipped. The command line exists as a specification and an early prototype; the local API, structured output and supervised execution are designed and not built. There is no package to install today and no endpoint to point a client at.

On engines: the runtime is built so that inference sits behind an adapter rather than being welded to one implementation. Third-party engines are used under their own licences, with their notices preserved, and are never presented as our work. If you care about that distinction - and if you are reading this page, you probably do - it is written down in the application’s notices rather than being a claim on a marketing site.

If the command line and the local API are the parts you care about, say so in the application - there is a Developer option, and it changes what we build first.

Apply as a developer tester
Prelaunch. The desktop application is in development and the private beta has not opened yet.