What Is Dottie Desktop? The AI Agent That Controls Your Mac

What Is Dottie Desktop? The AI Agent That Controls Your Mac

Last updated on

Dottie Desktop is a free, open-source AI agent for Mac. You speak or type a task; it executes on macOS through 134 built-in system tools — files, apps, Mail, Calendar, Terminal, screen OCR, and more. It is not a chat window that only answers. It is an agent that changes something on the machine.

Requirements: Apple Silicon (M1 or later) and macOS 15.4+. Cumulative downloads are past 1,000 (1,069 as of 2026-08-27). Source is on GitHub under MIT. Coming from a launcher graph: best Alfred workflows for Mac.

TL;DR

Chatbot vs agent (the only definition that matters)

A chatbot returns text. An agent decides which tools to call and runs them.

If you say “move every screenshot on my Desktop into Archive,” a chatbot prints steps. Dottie creates the folder, finds the files, and moves them — after you grant the relevant permissions and confirm destructive steps when required.

That split is why “AI assistant for Mac” search results feel noisy. Most hits are chat apps in a native shell. Dottie is built for the other half: after the model finishes, something on the Mac is different that you did not do by hand. Longer taxonomy in AI Agent for Mac.

How it works (short architecture)

Three local servers on your Mac:

ServerRolePort
Gateway (Node.js + Hono)Orchestrates chat, tool calls, streaming1317
Audio (Python + MLX)Speech recognition and TTS1315
LLM (local inference)On-device models when you choose local1316

Everything listens on localhost. Nothing is exposed to the network by default. Responses stream over Server-Sent Events. Rich cards (weather, calendar, timers, confirmations) render in the conversation via DotUIKit — not only plain text.

Tool calling follows the usual pattern: the model receives your request, picks tools, the Gateway runs them against macOS APIs, results stream back. Docs for the full tool list: docs.dottie.ai.

Voice-first on macOS

Three modes cover different desk setups.

Push-to-Talk — Hold a hotkey (default Option+Space, configurable), speak, release. Words go to the focused field or into chat, depending on mode. STT runs locally (parakeet-tdt via MLX) when you use the on-device path — audio need not leave the Mac.

Chat mode — Mic in the Dottie window. Voice becomes an agent message with full tool access: “email Sarah about the deadline” can compose and send through Mail.app after confirmation rules you set.

Wake word — “Hey Dottie” from anywhere. On-device phrase detection via Apple’s Speech framework; a floating overlay offers Send, Edit, or Dismiss. Hands-free when you are cooking, standing, or mid-hardware work.

TTS uses Kokoro-82M with speed control and optional auto-speak. Realtime mode supports barge-in (interrupt mid-reply by speaking). macOS Services: right-click selected text → Speak with Dottie. Full walkthrough: voice control guide.

What the 134 tools actually cover

Tools map to macOS capability categories. Permissions default off; you enable only what you need.

  • Communication: Mail (read/search/send), Messages, Contacts
  • Time: Calendar (EventKit), Reminders
  • Knowledge: Notes, Safari tabs/titles, file search, Photos
  • Media: Music playback, screenshots + vision/OCR
  • System: App open/focus/close, Terminal/shell, System Settings-adjacent actions
  • Context: Weather, stocks, sports, maps, web search, image generation

Security model: category permissions + confirmation dialogs for destructive actions. Dottie is sandboxed to what you permit. It is not a silent root shell with a friendly UI.

Three model paths (not “local only”)

Pick per priority: privacy, cost, or raw capability.

PathWhenNotes
Local (MLX / Ollama)Private or offline workOpen-source models on Apple Silicon; default small chat model for tool calling; STT/TTS on-device
BYOKYou already pay OpenAI/Anthropic/xAI/CerebrasKeys stay yours; requests go provider-direct
Dottie Pro ($20/mo)Keyless managed accessSame email as in-app register unlocks Pro

Local and cloud can coexist — use local for routine automation, cloud for hard reasoning. Deeper local setup: run local AI models on Mac and Ollama vs LM Studio.

Extended thinking: prefix /think when you want slower multi-step reasoning before tool calls.

Goals, triggers, scheduled jobs, memory

Beyond one-shot commands:

  • Goals — Persistent objectives that bias what the agent prioritizes (“stay on top of email”).
  • Triggers — Conditions that fire actions (“email from boss → notify immediately”).
  • Scheduled jobs — Cron-style tasks (Morning Brief: calendar + weather + unread).
  • Memory — Editable user facts plus automatic observations in a local SQLite-backed graph so the agent does not re-ask your name every session.

Privacy defaults

Nothing leaves the Mac unless you choose a cloud path.

  • Local STT, TTS, and local LLM → zero provider traffic
  • Cloud models → your key, provider ToS, no Dottie middlebox
  • Tool permissions off by default; confirm destructive steps
  • Product analytics on the website are separate from the desktop app’s local-first design

Compared with chat-only apps in best AI assistants for Mac, the privacy story is the combination of on-device speech + optional local models + explicit tool grants — not a marketing slogan alone.

URL schemes for Shortcuts, Alfred, Raycast, shell:

  • dottie://record — start voice recording
  • dottie://chat?text=… — send a message
  • dottie://type?text=… — type at cursor
  • dottie://settings / show / hide / stop

Use Dottie as the AI execution layer inside a larger automation. Practical patterns: automate your Mac with AI.

How Dottie compares (one screen)

ProductRole on Mac
ChatGPT / Claude desktopChat and (at edges) sandboxed agents — see ChatGPT vs Claude on Mac
Claude CoworkDocument/sandbox agent — Cowork vs Dottie
Raycast AILauncher + text AI — Raycast AI vs Dottie
BoltAIMulti-model Mac chat client — BoltAI vs Dottie
ElephasKnowledge + writing assistant (Super Brain, Super Command) — Elephas vs Dottie
Apple Intelligence / SiriSystem features with a hard ceiling — Apple Intelligence vs Dottie, Siri vs Dottie, Apple Intelligence alternatives
DottieNative API agent: voice + 134 tools + local/BYOK/Pro

Get Dottie

Timed walkthrough: Getting Started with Dottie on Mac (first 10 minutes).

  1. Download from the landing page (demo, requirements, and Pro link live there).
  2. Confirm macOS 15.4+ and Apple Silicon.
  3. First launch downloads default on-device models (~2 GB). After that, local path works offline.
  4. Enable only the tool permissions you need.

Source and releases: github.com/stevederico/dottie-desktop-releases. Announcement context: Introducing Dottie Desktop.

Frequently Asked Questions

What is Dottie Desktop?

Dottie Desktop is a free, open-source AI agent for Apple Silicon Macs. You give it voice or text instructions and it executes tasks on macOS through 134 native system tools — files, apps, Mail, Calendar, Terminal, screen reading, and more.

Is Dottie Desktop free?

Yes. The app is free and open source under the MIT License. Local models cost nothing beyond your hardware. Cloud providers need your own API keys. Optional Dottie Pro at $20 per month adds managed model access without key setup; it does not lock core automation.

What Mac do I need for Dottie?

Apple Silicon (M1 or later) and macOS 15.4 or newer. Intel Macs are not supported. First launch downloads roughly 2 GB of default on-device models for speech and local chat.

Can Dottie work without internet?

Yes, on the local path. Chat, speech-to-text, and text-to-speech can run fully on-device via MLX and Ollama. Cloud providers and Dottie Pro need a network connection.

How is Dottie different from ChatGPT or Claude on Mac?

ChatGPT and Claude are primarily chat applications. They answer and, in limited modes, act in sandboxes or hosted environments. Dottie is built to call macOS APIs directly so the result is a change on your Mac — moved files, sent mail, scheduled events — not only instructions.

Is Dottie safe? What can it access?

All tool permission categories default to off. You enable only what you need. Destructive actions ask for confirmation. Local processing keeps data on your machine unless you opt into a cloud model.

Does Dottie have a wake word?

Yes. Say Hey Dottie for hands-free capture with on-device phrase detection. Push-to-talk and in-window chat mic are also available. That is a different product shape from ChatGPT or Claude desktop voice, which start from a click inside the app.

← Back to Blog

More from the blog