BetaDraft AI is now in private preview

Intelligence that feels instant. 

We are engineering beautiful, locally-run AI models and context engines designed to anticipate your needs without compromising your privacy.

On-device inference
Zero telemetry
Sub-40ms context
am-core-8b /on-device /context graph /zero telemetry /38ms first token /4-bit quantised /no cloud /open weights /am-core-8b /on-device /context graph /zero telemetry /38ms first token /4-bit quantised /no cloud /open weights /

The tech stack

Three systems, one local-first brain.

Every layer is built in-house — from the drafting surface down to the retrieval kernel that runs entirely on your machine.

Beta

Draft AI

A writing surface that thinks one paragraph ahead. Draft predicts intent from your workspace context and renders suggestions before the cursor stops moving.

Latency
38ms
Context
128k
Runs
Local
Research

Local Models

Distilled 3B–8B checkpoints quantised for consumer silicon. No cloud round-trip, no data leaving the device.

am-mini-3b92%
am-core-8b64%
am-vision31%
In development

Context Engine

A continuously indexed graph of your files, messages and decisions — retrieved in milliseconds, encrypted at rest, never synced.

Vector graphIncremental indexAES-256Zero sync
Private

Nothing leaves the device.

Instant

Perceptibly zero latency.

Open

Inspectable weights.

Read how we benchmark local inference

Live engine

Watch it think — offline.

One binary. No API key, no account, no network. The engine boots your model, attaches your private index and answers in under a second.

812msCold boot to first token
0 bytesSent to any server
12,481Context nodes indexed
aestheticmade — am-core-8b — zsh
$
ask anything — it never leaves this machine run

Research notes

Working in the open.

Short, honest write-ups from the lab — published as we learn, not after we polish.

Software that respects the machine
and the person using it.