Developer John Bean has released KeyType, an open-source system-wide AI autocomplete tool for macOS that positions itself as a privacy-focused alternative to the proprietary Cotypist application. The MIT-licensed tool runs entirely on-device using local language models, eliminating cloud dependencies and subscription fees while providing tab-autocomplete functionality across all macOS applications.
On-Device Processing With Local LLMs
KeyType monitors active text fields across any macOS application and uses local large language models to predict text continuations at the cursor position. The application presents predictions as ghost text that users can accept by pressing the Tab key. By leveraging llama.cpp for model execution and supporting models like Gemma4 and Qwen3-5, KeyType demonstrates that sophisticated AI autocomplete features can run entirely locally without cloud services.
The tool requires macOS 14 or later and integrates deeply with the operating system through accessibility APIs to achieve system-wide functionality—a technically challenging implementation that indicates production-quality engineering rather than a simple proof-of-concept.
Modular Architecture Enables Advanced Features
KeyType's codebase, composed of 98.9% Swift and 1.1% Shell across 56 commits, uses a modular package structure that separates concerns:
- MacContextCapture: Handles accessibility framework integration for focus detection and text-field snapshots
- Prompting: Manages sectioned, budgeted prompt construction
- ModelRuntime: Wraps llama.cpp for model execution
- ConstrainedGeneration: Implements logit masking and search optimization
- TokenProfiles: Manages ACPF profile configurations
- CompletionUI: Renders overlay ghost text display
- TextInsertion: Handles keystroke and pasteboard-based insertion strategies
- AppCompatibility: Manages per-application override policies
This architecture allows KeyType to work seamlessly across different applications while respecting app-specific behaviors and compatibility requirements.
Privacy-First Alternative to Subscription Services
KeyType's release comes amid growing interest in local AI tools that avoid vendor lock-in and subscription models. While Cotypist charges ongoing subscription fees for similar functionality, KeyType provides full source code transparency under the MIT license, addressing privacy concerns and allowing community contributions. The project has gained 208 GitHub stars since its May 31, 2026 creation, with three releases already available.
The distinction is significant for users concerned about data privacy: KeyType processes all text locally without sending information to external servers, eliminating risks associated with cloud-based autocomplete services that may log or analyze user input.
Key Takeaways
- KeyType is an open-source, MIT-licensed macOS autocomplete tool that runs entirely on-device using local LLMs via llama.cpp
- The application works system-wide across all macOS applications through accessibility API integration, presenting predictions as ghost text accepted via Tab key
- KeyType's modular Swift architecture includes specialized packages for context capture, constrained generation, and app-specific compatibility
- The tool positions itself as a privacy-focused alternative to Cotypist, eliminating subscription fees and cloud dependencies while providing full source code transparency
- Since launching on May 31, 2026, the project has gained 208 GitHub stars across 56 commits and three releases