Overview of the Release
KeyType is an open-source macOS utility that adds system-wide tab autocomplete powered by local LLMs. The project appeared on GitHub trending lists in recent weeks as an MIT-licensed replacement for the closed-source Cotypist application. It monitors the focused text field in any app, runs inference on-device, and renders inline ghost text that users accept with the Tab key.
How the Architecture Works
The repository organizes logic across several SwiftPM packages rather than a single monolith. MacContextCapture relies on Accessibility APIs to read the current AX focused element, caret position, and surrounding text snapshot. Prompting builds a budgeted, sectioned context string that stays under token limits. ModelRuntime wraps llama.cpp for loading quantized models such as qwen3-5 or gemma variants.
ConstrainedGeneration applies logit masking and trie-based admissibility checks so completions remain valid within the current text field rules. TokenProfiles loads ACPF files that define per-app token restrictions. CompletionUI draws the ghost text overlay without injecting into the target application. TextInsertion falls back between pasteboard and simulated keystrokes depending on the frontmost process.
These choices keep the app lightweight while supporting per-domain overrides stored in AppCompatibility. The design trades some latency for privacy since no text leaves the machine.
Building and Running the Project
Developers targeting macOS 14 or later need a current Xcode installation. After cloning
The release DMG contains a ready-to-run binary for users who prefer not to compile. Onboarding inside the app requests Accessibility permissions and lets users select a local model file. No additional cloud accounts or API keys are required.
Practical Trade-offs
On-device inference removes data exfiltration risks but shifts compute cost to the user's hardware. Performance varies with model size and Apple Silicon generation. Some applications block Accessibility reads or reject simulated input, so the AppCompatibility layer maintains a growing list of exceptions.
Compared with cloud-based autocomplete services, KeyType offers lower latency once the model is loaded yet cannot match the breadth of a hosted 70B-parameter network. The open repository allows contributors to extend prompt templates or add new insertion strategies without reverse-engineering a proprietary binary.
FAQs
Does KeyType require an internet connection? No. All inference runs through the bundled llama.cpp runtime using models stored locally on disk.
Can I train or fine-tune models for KeyType? The current release consumes pre-quantized GGUF files. No training pipeline is included in the repository.
How does it handle secure text fields such as password prompts? The AppCompatibility package contains rules that disable ghost text in password and credit-card fields by default.
---
📖 Related articles
- Agentic Coding: Una Trappola per lo Sviluppo Software?
- Claude Code: da prompt a agente con skills e subagent
- Phantom su GitHub: L'AI co-worker auto-evolvente e sicuro
Need a consultation?
I help companies and startups build software, automate workflows, and integrate AI. Let's talk.
Get in touch