FluidUse
A Swift library for local computer-use on Apple Silicon: it reads a form in a running app or browser through the Accessibility API and types answers into each field using a small on-device model.
FluidUse reads a form in a running Mac app or browser through the macOS Accessibility API, asks a small on-device model what belongs in each field, and types the answer into the real app, taking about 1 millisecond per decision on the Apple Neural Engine with nothing leaving the machine. Its first model, CUA-S1-FORMS, is a 706K-parameter form-filling specialist converted to Core ML; a WebFormDriver component does the equivalent for an embedded WKWebView, and a DocumentEntities component turns a PDF or a text file of label-value lines into the profile used to fill a form.
Underneath forms, FluidUse is built on laya, an open decision model: a 322-million-parameter mmBERT encoder plus a decision head that answers typed choice, score, or boolean questions about a block of text with calibrated probabilities in a single pass, without generating tokens. A manager component runs precompiled Core ML buckets sized for different input lengths on the Neural Engine, reporting a few milliseconds per short question on a recent Apple chip, and the project reports its Core ML buckets matching the accuracy of the original PyTorch reference across laya's published evaluation suites while running many times faster.