Meet Samantha.

Her Mac. Open apps and sites, search, read the current tab, screenshots, the clipboard, volume, battery, music, notes, reminders with a time ("remind me tomorrow at 9am to call mom"), the calendar, unread mail, timers, the weather, files: find one by name, what you just downloaded, how big a folder is, then move, copy, rename, zip, unzip or trash it, asking first.

Pictures. Draw anything you can name (an image model imagines it, she rebuilds it from squares), paint any photo out of squares or in a style ("paint the mona lisa as a mosaic": mosaic, dots, poster, sketch, stained glass), make a logo (the icon on this page is one she designed), remove a background, upscale, enhance, grayscale, rotate, flip, resize, crop, convert.

Math. Calculate, convert units, tips, primes and factors, roman numerals.

Documents. On a Mac she reads PDFs, Word files and plain text in your home folder, finds the passages that mention what you ask about, and answers questions about them, or about your screen, from the text alone: a number or name that is not in the text is declined. Nothing hidden, nothing outside home.

Video and audio. Say "transcribe the video ~/Desktop/clip.mp4" and she pulls the words out of any video or audio file in your home folder with the same Whisper pipeline her voice mode uses.

Real summaries. "Summarize ~/Desktop/report.pdf", "summarize this page" or "summarize my unread mail" hands the text to the biggest local model for three to five real sentences, not the one-line answer a lookup gives. Anything short enough to just read is read back instead.

Translates, offline. "Translate good morning to French", "how do you say thank you in Japanese", or "translate the page github.com into Spanish": the same local model does it.

Research. On a Mac, say "research the history of the printing press". She reads several sources, and a bigger model on the same Mac writes a short brief with a source after every sentence. A sentence she cannot back up is cut.

Eyes. On a Mac, say "look at my screen and tell me what's wrong with this chart" or "what's in ~/Desktop/cat.png". A small vision model on the Mac looks and answers. It asks first.

Hands on your screen. On a Mac, say "click Sign in", "type hello" or "press return". She reads the screen to find the words you named and clicks them, and she shows you each step and asks before it happens. Press Escape any time and she stops.

Camera. On a Mac, say "what am I holding" or "read this label". One frame from the Mac's own camera goes to the same vision model, and the frame is deleted right after the answer. It asks first.

Run code. On a Mac, say "stats on sales.csv" or "chart sales.csv". The biggest local model writes a short Python script from your request and the file's header, and it only runs in a sandbox: a fresh folder holding a copy of that one file, no network, a time and memory cap. A chart comes back as a real PNG. It asks first.

Long work. On a Mac, say "work on cleaning up my downloads". She plans in rounds: when a step fails she plans around it, still does the parts that can be done, and tells you at the end what she could not. Each new plan comes from your words and her own record, never from what she read. Every write still asks.

Asks which. Say "trash this file" or "crop that image" without naming it and she asks which one, instead of guessing or saying no.

What needs you. On a Mac, say "what needs my attention" for one ranked answer built from your unread mail, today's calendar and due reminders, or "am I free Thursday afternoon" for your calendar's open gaps that day or across the week. Both read only, and both say plainly when there is nothing. Say "good morning" for the same with the date on top, and the Mac app brings it to you by itself the first time it opens each morning.

Talk to her. On a Mac, python3 app/chat.py --voice and just speak. Whisper turns your voice into text on the Mac itself, she answers, and she says it out loud. Add an ElevenLabs key and she speaks in a real voice. Add --face and a window shows her face, idle while you talk and talking while she answers. Set it up. Anything that writes still asks first.

Cut her off. With voice on, anything new you send stops her mid-sentence, and "shh" or "stop" just silences her. In the Mac app your chat is saved after every answer and comes back when you reopen it; Cmd-N files it away and starts a clean one. Nothing leaves the Mac.

Joshua Tree. On a Mac, "Joshua Tree status" reports the OS's version, open pull requests and whether its test machine is running. "Boot Joshua Tree" starts it headless in qemu and "stop Joshua Tree" shuts that one process down; both ask first. "What did Joshua Tree say" reads the last lines of its serial log, and "type ls into Joshua Tree and press enter" sends it real keystrokes, asking first. On a Raspberry Pi she is an app on Joshua Tree itself: ask her to write hello world in C and save it as hello.c, and the code lands on the Pi's SD card. "Screenshot Joshua Tree" saves its screen to a file her eyes can read, and "what is Joshua Tree showing" does both in one go, asking first.

Her own library. On a Mac she keeps the lead of every one of Wikipedia's ~10,000 vital articles, anthropology to astronomy, plus a plain-words fieldbook of every science. When the web has nothing or is down she answers "what is" and "who was" questions from the page of that name, says which page, and declines the rest instead of guessing.

Ask any LLM, still local. Say "ask qwen", "ask llama" or "ask gemma" and a question too hard for her goes to that model on your Mac, found through oMLX or Ollama. Say "ask claude" or "ask gpt" and she uses the biggest one there. If you name one you do not have, she lists what you do. It asks first, needs no account or key, and the answer names the model that gave it.

Screen. On a Mac she reads what is on your screen with the system's own text recognition. It asks first.

Memory. Tell her "remember that my dog is called Biscuit" and she still knows next visit: on this page it lives in your own browser, on your Mac in a file only she reads. Recalling and forgetting are by name only, and remembering asks first. On your Mac, from 4.19, she also reads what she remembers back into her answers when it bears on the question: ask about your dog and she uses what you told her. If you ask, she keeps that file in your own iCloud Drive ("keep my memory in iCloud"), so your other devices can carry it; the phone app keeps its own copy and reads it back too.

Explain and teach. On a Mac, "explain the nimble codebase" walks you through a project from its README and folders, and "teach me about tides" gives a short lesson and two quiz questions. Neither changes anything.

Email. On a Mac, "email mom saying I'll be late" finds her in Contacts, shows you the address and the words, and sends only after you say yes.

Chrome. List your tabs, switch to one, read it as it is rendered (so pages built by JavaScript work), and close one, which asks first.

Time. The time in any big city, today's date, days until a date or a holiday.

Chance. Coin, dice, random numbers, passwords, UUIDs.

Text. SHA-256, base64, word count, reverse, capitals, morse, tidy JSON.

Shortcuts. List your Apple Shortcuts and run any one by name, so every automation you already built is one of her tools. She only runs one when you name it, never on her own.

Promises she keeps. Seven rules live in LAWS.md and are checked against every tool on every push: anything that writes or sends asks first, the tools with hidden side effects never reach a model, her hands stay in the home folder, every function is documented, and this page says the same words the Mac does.

Plugs into other assistants. She speaks MCP, the standard plug for AI tools, both ways: Claude Code or any other assistant can use her tools, and she can call theirs, asking first. A harness keeps the conversation, prints every tool call before it runs, and asks first before anything that writes or sends.

This Mac's vitals. Disk space, uptime, memory, processor load, IP address, Wi-Fi, system info, copy to clipboard, sleep the screen, show a file in Finder. These answer on a real Mac. The stand-in above has no disk to read, and says so.