Hi HN, I built an open-source (MIT) AI agent for macOS that talks directly to Mail, Calendar, Reminders, Notes, and Safari through AppleScript — no OAuth flows, no API gateways, no passwords to store.
The premise is simple: your Mac apps are already signed in. Why should an AI assistant make you re-authenticate everything through a separate integration layer? Why do I need to create an Azure app to get an OAuth token for my Office 365 emails? Why do I need a Google PubSub setup?
I needed something that is very simple to setup for my family and friends and can do almost the same things they do daily on my Mac. And they do not need to enter passwords or other authentication. In the end it is another harness but one that we love to use.
On Telegram you can dictate things like "My mom sent me her vacation dates by email — add them to my calendar" and it chains the right tools: searches Mail, extracts the dates, creates the Calendar event.
What makes it different:
- *No credentials needed for Apple apps* — AppleScript talks to them directly, macOS Keychain handles auth
- *Runs fully offline* — use local models via Pico AI Server, or cloud providers (Anthropic, OpenAI, OpenRouter)
- *Desktop app with setup wizard* — not a CLI-only project; download a .dmg, follow the steps, done
- *Telegram as remote control* — text or voice-message the agent from your phone
- *Skills ecosystem* — extensible via drop-in SKILL.md files (compatible with AgentSkills/ClawHub)
- *Memory between sessions* — remembers your preferences and patterns in a local YAML file you can read/edit/delete
The agent is a ReAct loop in Python. Tools are auto-discovered from type hints. The desktop shell is Tauri (Svelte + Rust) with the Python agent bundled as a sidecar binary via PyInstaller.
I spent a lot of time on AppleScript performance — bulk-fetching with `whose` clauses instead of iterating messages one by one (each Apple Event has ~0.1-0.5s overhead, so naively reading a mailbox would take minutes).
All data stays on your Mac under `~/.macbot/`. Your prompts go to your chosen LLM provider; nothing else leaves the machine.
GitHub: https://github.com/spamsch/son-of-simon
