Skip to content

About

Built where the work happens 

Vaultr is a small, source-open product for legal teams that need AI without a transmission requirement. This page says who builds it, why it exists in this shape, and what stage it is honestly at — the same claims discipline as the rest of the site, turned on ourselves.

Singapore · AGPL-3.0 · assisted evaluation

01

Why this product exists

Legal work runs on confidentiality, and the first generation of legal AI was built as if that were a compliance detail. Tools arrived that asked firms to send their matters to somebody else's cloud and trust a policy document in return. For a profession whose raw material is privileged, that was the wrong default — not an unacceptable risk to negotiate down, but the wrong default.

Vaultr started from the opposite end: the boundary first. What happens on the device stays on the device, and anything that would cross the boundary is a named, explicit mode — never a fallback. Everything else, from the evidence ledger to the published boundary test, follows from that decision.

“The room stays closed” is not a slogan about paranoia. It is a description of the default processing mode.

02

How it is built

In the open. The repository is public under AGPL-3.0 — licence, attribution, history and all — because a security team that can read the code is worth more to us than any paragraph this page could write. Dependency lockfiles are committed, automated checks run on the supported flow, and every claim on this site traces to an artifact you can open.

03

Who builds it

Vaultr is built by a small team operating from Singapore, working directly with the firms evaluating the product — no account-management layer between the people building it and the people testing it. Support during evaluation is replies from the builders; the access form reaches exactly those people.

04

What stage it is at, honestly

Assisted evaluation, by request. No public download, no self-service sign-up — deliberately, because what is being evaluated is fit against your bundle and your intended mode, not a demo. macOS on Apple silicon is the tested target; Windows and Linux are planned, not released. No independent audit exists yet; the security brief's not-done list says so in those words.

Release stage
Assisted evaluation with a strict allocation of local deployments
Tested platform
macOS on Apple silicon
Licence
AGPL-3.0, public repository
Independent audit
None yet — stated on the security brief
05

The practical fit

Hardware, in plain language

16 GB unified memory minimum for small local models; 32 GB or more for long bundles. More memory helps with longer bundles and larger local models; it is not a performance guarantee. The current tested target is macOS on Apple silicon.

Who this is — and isn't — for

For firms that prioritise an inspectable boundary, source-linked work and human checkpoints over maximum cloud breadth. It is not for a team looking for instant self-serve access, a universal operating-system promise or an AI system that makes the final call.

Windows and Linux
Planned, not released. No compatibility date is promised.
Independent boundary reproduction
The next evidence milestone. No completion date is claimed.
Production integrations
Evaluated with firms case by case; no integration availability is claimed today.

Evaluation data and exported artefacts stay under the retention rules of the mode you choose. In Local Mode, your device and its backups are the boundary; an external endpoint brings that host's retention into scope. There is no Vaultr-operated matter-data backup service.

06

Reach us

Evaluation, questions and security reports all land with the same small group of people.