A native macOS AI agent workspace. Run multiple independent AI sessions in tabs, each with its own conversation context, tool access, and the ability to spawn parallel sub-agents for complex tasks.
- Native desktop interface for long-running agent work instead of a single browser chat
- Parallel sub-agents and local tool access in the same session model
- Public release surface for a production-shaped macOS AI product
Download the latest DMG from Releases.
The public repository currently ships release artifacts and product documentation rather than the full application source tree.
- Multi-session tabs — each tab runs an independent AI agent conversation
- Sub-agents — sessions can spawn child agents that work in parallel and report back
- Local tools — agents interact with your machine through plugins running in isolated subprocesses
- Speech-to-text — built-in voice input with configurable keyboard/mouse triggers
- Streaming output — real-time token streaming with activity indicators
- Background execution — sessions keep running when the app is minimized
Download the latest DMG from Releases, open it, and drag Mi Work to /Applications.
Auto-updates are built in via Sparkle — the app checks for new versions automatically.
Requirements: macOS 15 (Sequoia) or later.
- Launch Mi Work from
/Applications. - On first launch, open Settings and connect an Anthropic API key.
- Create a new tab to start an isolated work session.
- Enable the tools you want that session to access.
- Use sub-agents when the task benefits from parallel execution.
Settings supports two setup paths:
- Sign In with Anthropic — OAuth flow opens your browser, creates an API key automatically
- Manual entry — paste an existing API key from console.anthropic.com
The key is stored securely in macOS Keychain.
- Native AppKit — no Electron, no web views, built with Swift and AppKit
- Swift concurrency — actors for isolation, structured concurrency for all async work
- Plugin system — external processes communicate via JSON-RPC over Unix sockets
- Dual-stream UI — content stream (conversation) and activity stream (status bar) rendered independently at up to 120fps
Proprietary. All rights reserved. See LICENSE.