Cotabby logo
Cotabby

Security model

what Cotabby reads, where it goes, and why

System-wide autocomplete needs access to the field where you type. Cotabby keeps generation local, does not include telemetry, and publishes its source so that behavior can be verified.

By Jacob Fu · Updated · 7 minute read

security summary

  • Active-field text is used transiently to generate the next suggestion.
  • Optional screen and clipboard context is not stored by Cotabby.
  • Model inference works offline after any required model download.
  • Source code is available under AGPL-3.0 for independent inspection.

suggestion data flow

  1. 1. You type. Cotabby detects a supported editable field and reads nearby text exposed by macOS Accessibility.
  2. 2. Context is assembled. The active field is combined with enabled nearby-screen or clipboard context.
  3. 3. The local engine runs. Apple's Foundation Models runtime or llama.cpp generates a continuation on your Mac.
  4. 4. Ghost text appears. The suggestion remains separate from your document until you accept it.
  5. 5. You decide. Tab accepts one word, backtick accepts the full continuation, and Esc dismisses it.

macOS permissions explained

Accessibility
Reads the active editable field, positions ghost text, and inserts words you explicitly accept.
Input Monitoring
Detects acceptance and dismissal shortcuts while another application is active.
Screen Recording
Optionally reads text visible near the cursor to make a suggestion relevant to the surrounding email, document, or conversation. Cotabby does not record video.

Permissions can be revoked at any time under System Settings, then Privacy & Security. Removing a permission disables the capability that depends on it.

network access

Cotabby does not need network access to generate suggestions. Network requests are limited to checking for application updates and downloading a local model when you explicitly choose one. Model files may be hosted by Hugging Face, whose server receives the ordinary metadata associated with that download request.

The Apple Intelligence engine uses the model runtime built into macOS and does not require Cotabby to download a GGUF model. After setup, both engines can produce autocomplete offline.

what is stored

Cotabby stores application settings, keyboard bindings, disabled-app preferences, custom instructions, macro definitions, and downloaded model files on your Mac. It does not maintain a cloud account or a server-side history of your writing.

Clipboard and nearby screen text are transient inputs to a suggestion; Cotabby does not save them as a history. The definitive policy is the Cotabby privacy policy.

how to verify the claims

Cotabby's repository includes the application source and release history. Security researchers and users can inspect permission use, network code, model integration, and changes between releases on GitHub.

Report a suspected vulnerability privately to hello@cotabby.app. Do not include sensitive text beyond what is necessary to reproduce the issue. General bugs can use the feedback form.

comparing privacy claims

“On-device” should be evaluated separately from source availability, telemetry, update checks, model downloads, and the breadth of macOS permissions. Cotabby documents each of those dimensions rather than treating local inference as the whole security story.

If you are comparing another autocomplete product, the Cotabby vs Cotypist guide separates their shared local-processing model from their ownership, pricing, and feature differences.

try Cotabby on your Mac

Free, open source, and local. Use Apple Intelligence or a GGUF model, then accept the next word with Tab.

View source