Setting up, talking, the buddy, the canvas, personalities, memory, agents, apps, coding, models, updates and privacy. Five minutes to read; search narrows the list.
Nothing matches. Try a shorter word.
Download the .dmg, drag Pixie to Applications, then right-click → Open the first time. (Windows .exe and Linux .AppImage previews are on the Download page; on Windows turn on Settings → Privacy & security → Speech → Online speech recognition so Pixie can hear you. What works where.) The :) appears in the menu bar and the welcome window starts: your name, permissions, a model key, a mic test, and the shortcuts. The natural voice downloads in the background while you go.
| Permission | Used for |
|---|---|
| Microphone, Speech | Hearing you, on-device |
| Screen Recording | Seeing the screen when you ask |
| Accessibility | Finding controls precisely, clicking when you allow it |
Screen Recording needs a relaunch after you allow it. If a switch is on but Pixie says "not allowed", use Reset in Settings → Permissions.
Settings → Models: pick a provider, paste its key. Models load live from the provider, newest first, with context size and price. Mark one for talking (T), one for agents (A), and optionally one fallback that answers if the main model fails.
| Provider | Key |
|---|---|
| Claude | ANTHROPIC_API_KEY |
| OpenAI | OPENAI_API_KEY |
| DeepSeek | DEEPSEEK |
| Do this | What happens |
|---|---|
| Hold ⌃⌥, talk, let go | Pixie listens, thinks and answers out loud |
| Tap ⌃⌥ | A box to type instead |
| Hold fn⌃ | Dictate clean text into any field |
| Hover the notch | Chat and Agents tabs |
| esc | Stop everything, including a live agent |
Click the :) in the menu bar | Type a question; the menu has Show Jobs, Show Canvas, Settings, updates and Quit |
Change the keys in Settings → General → Shortcuts, or just ask: “make dictation ctrl shift”. On Windows the keys are CtrlAlt to talk and CtrlWin to dictate.
| Say | What happens |
|---|---|
| “quit”, “exit”, “close Pixie”, “turn yourself off” | Says bye and quits |
| “stop working”, “stop everything”, “cancel all” | Stops every running job and live agent |
| “never mind”, “stop”, “be quiet” | Stops talking (like esc) |
| “new chat”, “start over” | Starts a fresh conversation; the last one stays in Chats |
| “what can you do?”, “which model are you on?”, “what's my shortcut?” | Answers from how it's set up right now |
| “talk slower”, “use the Emma voice”, “turn the volume to 80%”, “stop reading answers out loud”, “don't click for me”, “switch to Claude”, “turn memory off”, “call me Sam” | Changes the setting at once; several in one go is fine |
| “switch to Eon”, “be Xtrem”, “turn on spicy”, “draw more”, “only draw when I ask” | Changes the personality or how often it draws |
| “open the voice settings”, “check for updates” | Opens that page, or checks GitHub |
The short ones (quit, stop working, never mind, new chat, switch to Eon) work instantly without asking the model. API keys, permissions and Google sign-in can't be changed by voice: Pixie opens the page for you.
The round :) rides just off your pointer. It points at real controls by name, draws circles, underlines and selections, and clicks or types when you ask. Settings → General: Always on, Idle (only while Pixie is busy) or Off; and whether it may click for you.
Today's chat carries on all day: after a quiet hour or a restart, Pixie still knows what you were talking about (and how long it's been). It ends when you press New chat in the notch, say “new chat”, or the day changes: at midnight, or the hour you pick in Settings → Memory. The last 12 turns go to the model in full and older ones as a short summary, so long days stay quick. Every chat is kept in the Jobs window.
Long-term memory is a set of short notes about you: who you are, how you like answers, your people, projects and setup. Say “remember that…” or “forget…”. On Automatic, Pixie also reads a chat once it's been quiet for a few minutes and keeps anything lasting; Only when I ask keeps just what you tell it; Off stops reading and writing notes. The most useful notes go with every question and Pixie searches the rest when they matter. In Settings → Memory you can add, edit (click a note), search and forget notes, or forget everything. Notes live in memory.json on your computer; passwords, keys and card numbers are refused.
Background agents work on their own (research, documents, files, code) and keep a live note. Live agents do the task on your screen and explain each step. Both show in the notch's Agents tab; past jobs and chats are in the Jobs window. Agents never overwrite your files and can't touch .ssh, .aws, keychains or .env files.
| Agent | Install |
|---|---|
| Codex CLI | npm i -g @openai/codex |
| Claude Code | npm i -g @anthropic-ai/claude-code |
| opencode | npm i -g opencode-ai |
Pick the default agent, a model per agent, sandbox and permission modes in Settings → Coding.
Settings → Apps switches on Calendar, Reminders, Contacts and Apple Mail; each asks macOS once. Then just ask: “what's on tomorrow?”, “remind me to call Sam at 5”, “draft a reply to Alex”. Drafts open for you to send; Pixie never sends mail.
Google: create an OAuth client of type Desktop app in Google Cloud, enable the Gmail and Calendar APIs, paste the client ID and secret, and sign in. Pixie reads Gmail, saves drafts and reads or adds events; tokens stay on your Mac.
Settings → Connectors gives Pixie tools from other services, grouped by what they're for. Each card says how it signs in and whether it's working.
| Connector | Setup |
|---|---|
| GitHub (GitHub's own server) | A fine-grained personal access token |
| Linear, Jira & Confluence | Sign in in your browser the first time |
| Sentry | A user auth token |
| Notion | An internal integration secret, with pages shared to it |
| Slack (community server; posting stays off) | A user OAuth token (xoxp-) |
| Brave Search | An API key (free plan available) |
| Context7 (library docs) | Nothing |
| A folder | Pick the folder |
A green dot means connected, with the number of tools; What it can do lists them. If one can't connect, the card says why in plain words, with Retry and the server's own message under Details. Add your own server takes any MCP server's command (and environment, one KEY=value per line) without editing mcp.json by hand. Connectors run through Node.js; tokens stay on your computer. Pixie asks before anything that posts, sends or deletes.
Pixie updates itself from GitHub. It checks when it starts and every six hours, downloads a new version in the background, checks its signature, and switches over when you aren't using it. Settings → About shows where it is and has Check now and Restart now; the menu-bar menu has Check for Updates. Turn off Update automatically to decide yourself.
When a picture explains it better, Pixie draws on a real Excalidraw canvas in its own window: boxes, arrows, text, or a Mermaid flowchart laid out for you. Everything stays editable. Name it at the top and save it as .excalidraw, PNG or SVG (to Downloads, never overwriting; Show reveals the file).
How often: Settings → General → Drawing: Only when I ask, When it helps (the default: flows, steps, how things connect) or Often. Or say “draw more” or “only draw when I ask”.
The window comes up in front of you, also over a full-screen app, and goes away when you click outside it. Bring it back with Show Canvas in the menu-bar menu, or ask: “open the canvas”, “open my git canvas”, “list my canvases”. Pixie can read your edits back (“add a database box under the server”), clear a canvas, or delete one or all of them; Delete in the canvas does it too (click twice), and deleted canvases go to the Trash.
| Pick | Face | Style |
|---|---|---|
| Pix (default) | :) | Friendly, warm, supportive: Pixie as it always was |
| Inos | 8) | Intelligent, nerdy, curious: precise terms and the why |
| Xtrem | XD | Chaotic, witty, mischievous: jokes about the situation, never about you. Spicy lets it swear a little |
| Eon | ;) | Calm, concise, precise: the answer, nothing else |
Pick one in Settings → General → Personality, or say “switch to Inos”, “be Eon”, “turn on spicy”. The face changes in the notch, on the buddy and in Settings. A personality changes only how replies sound: same voice, same tools, same care before sending or deleting. Documents and reports that agents write keep the plain style.
The natural voice runs offline (about 370 MB, downloaded during setup); the system voice works meanwhile. To stay light it loads when you start talking (about a second, while you speak) and unloads after five quiet minutes, so Pixie uses about 215 MB when idle. Settings → Sounds lets you change, quieten or replace every sound with your own file.
| Status | Notes | |
|---|---|---|
| macOS 14+ | Everything | Universal .dmg for Apple silicon and Intel |
| Windows 10/11 | Preview | Talking (Windows speech recognition: turn on Online speech recognition), spoken answers with Windows voices, screen reading, pointing, clicking and typing, agents. No natural voice, Apple apps or Chrome bridge yet |
| Linux | Preview | Chat, agents, screenshots, tesseract OCR, xdotool clicks (X11), espeak-ng answers. No voice input or talk shortcut yet |
Every feature and setting, ticked per platform, is in CHECKLIST.md.
Speech recognition, screen OCR and the natural voice run on your Mac. Screenshots and your words go only to the model provider you chose, with your key. Chats, memory notes, canvases and usage are kept locally; you can delete any of them. There is no Pixie server.
"Can't be opened": right-click → Open, or System Settings → Privacy & Security → Open Anyway. Permission not sticking: Settings → Permissions → Reset, allow, relaunch. No answer: check the key in Settings → Models. Pixie remembers something wrong: say “forget that…”, or delete the note in Settings → Memory. The canvas vanished: it hides when you click outside; Show Canvas in the menu-bar menu brings it back. Pixie draws too much: Settings → General → Drawing → Only when I ask. Still stuck? Open an issue.