Tasker technical documentation

Tasker is a task manager and day planner for iPhone, iPad, Mac and Apple Watch. Under the app sits TaskKit, a Swift package that holds the whole engine: models, storage, sync, planning, natural-language capture, the AI tool layer and the agent platform. The app is a thin SwiftUI shell over it.

These pages explain how that engine works, for developers and for anyone who wants to check the privacy claims instead of taking them on trust. They describe TaskKit 0.2, the engine inside Tasker 1.1. TaskKit ships inside the app; these pages are an engineering write-up, not an install guide for a public SDK.

Principles

• On-device first. Planning, capture and the tool layer run locally. Nothing leaves the device unless you turn on a network option.
• No servers. There is no Tasker backend. Sync goes through your own iCloud private database; sharing with other people uses files or a direct Wi-Fi link.
• Rules before models. Deterministic code answers first. A language model is optional and only fills gaps the rules leave.
• A plan is always produced. Every AI provider can fail, go offline or be switched off, and the built-in planner still answers.
• One substrate for every AI surface. Siri, Shortcuts, the in-app assistant, MCP clients and agents all call the same tool catalogue, so they behave the same way.
• Assistants are accountable. Every change an assistant makes passes one policy gate, asks before anything risky, and leaves a receipt.

The pieces

Module	What it holds	Links UI?
TaskCore	Models, the TaskStore facade, sync, planning, import and export, crypto, the AI layer and the agent platform	No, headless
TaskUI	SwiftUI views built on TaskCore: lists, the task map, location picker, pages, the assistants hub and approval cards	Yes
TaskIntents	App Intents for Siri, Shortcuts and Spotlight	Yes
TaskAdapters	Ready-made OS adapters for notifications and geofencing	Yes
task-mcp	An MCP server executable over stdio, with an optional HTTP binding	No

Where to start

• Architecture: layers, seams and the public API.
• Data and sync: models, schema versioning and the merge rules.
• Security: what is encrypted, where keys live, and the threat model.
• Planning: the rule planner, the provider chain and time blocking.
• Capture and tools: the capture grammar and the tool catalogue.
• Agent platform: assistants, the trust ladder, approvals and receipts.
• MCP server: connecting Claude Desktop or any MCP client to your tasks.
• Backups and export: the .unkt format.
• Roadmap: what is designed or in progress but not shipped.

Want the short version first? The Tasker showcase has a live capture parser and day planner you can try in the browser.