NEO 2.0

Say "Neo", or press Option+Command, and NEO operates the Mac the way a person would. It reads the accessibility tree, clicks and types, opens apps, runs shell commands, edits files, browses, sends mail and sets reminders. It asks before doing anything it cannot undo.

Intent accuracy: 0.90 (Held-out, vs 0.44 zero-shot)
Tool selection: 0.85 (Tool head)
Routing: ~20 ms (One encoder pass, M3 Pro)
Tools: 52 (Registered)

Why it feels fast

Most assistants send every word to a large model and wait. NEO routes the request first, so a simple command can run without a language model at all.

A volume command hits a regex fast path straight to the tool: 0 LLM calls, about 0 ms.
A factual question goes to Laya locally, then streams from Groq: about 150 ms plus the first token.
A complex instruction such as replying to an email goes to Laya, then the agent loop, which ships only the tools it needs, typically 3 to 7 of 52.

The router

The decision layer is Laya, a 421M parameter local model. One encoder pass, roughly 20 ms on an M3 Pro, feeds five small calibrated heads trained on about 16k utterances: NEO's own templates plus 12k real crowd-sourced phrasings from CLINC150 and MASSIVE mapped onto NEO labels. It decides what kind of request this is, which single tool performs it, whether the user wants an answer or an action, whether it is destructive, and whether it needs the screen. Held-out accuracy on real phrasings is 0.90 on the intent head, where Laya zero-shot alone scores 0.44; the tool head scores 0.85. When the intent head is unsure, it falls back to Laya's zero-shot answers, still locally. Only then does a language model get involved, and only the one the task needs.

What it can do

Every capability is a plain Python function registered with a decorator, which keeps adding a new one to a few lines.

See: the accessibility tree as text, plus screenshots as a vision fallback for canvas and Electron apps.
Act: click, drag, scroll, type, hotkeys, and accessibility press and set-value.
Apps and mail: open and quit apps, AppleScript, Finder reveal, mail, calendar, notes and reminders.
Web and browser: search, fetch, and its own Chromium with a persistent profile so logins stick.
Files and system: shell, read, write, edit, trash, volume, brightness, timers and display sleep.
Memory: facts about you, a journal of what was said and conversation summaries, searched by keyword and meaning together and given to every model on every route.
Routines: teach it a spoken phrase such as "I want to study" that opens an exact set of apps and sites.

Where it runs

The brains are Gemini and Groq free tiers, or Gemma 4 on your own GPU. The decision layer, wake word, speech to text and text to speech all run on device.