Privacy

Privacy

Trie performs speech recognition and text understanding on your Mac; dictation passes through no server. This page describes what the current version (0.5.0) processes, how long it is kept, when it connects to the network, and the known limits. Effective and last updated: 2026-10-07.

Key facts

On-device processing
Recognition and understanding run locally and call no cloud service. The model service uses offline settings and code-level network restrictions; a missing model produces an error rather than a download.
When it connects
In two cases only: installation and model download, and a single request for the version file when you choose “Check for Updates” in the menu bar. There is no other network activity.
When the microphone captures
Trie starts in standby with the microphone closed; once active, the selected input device captures continuously; audio enters a 30-second ring buffer in local memory and is overwritten by what follows. Start and stop only select the segment sent to recognition. With a Bluetooth headset the device stays open, but silence is written outside dictation. Switching back to standby or quitting Trie stops capture; tapping the trigger key in standby activates Trie and starts that dictation.
Use
Your audio, transcribed text and usage habits serve only the dictation you request: they are not uploaded to us or to third parties, not used to train or improve models, not sold, and not used for cross-context behavioral advertising.

Data and retention

Audio
Held in memory only, in two places: the 30-second buffer above, which is not written in standby and is zeroed on a normal exit; and, during a dictation, the segment itself (up to 3 minutes), kept in the model service’s memory for whole-segment recognition and released when the dictation is finalized, cancelled or interrupted. No recording files are created.
Transcribed text
Used to paste into the active field. For recovery after a failed paste, the latest result is kept in memory until the next result replaces it or Trie quits. It is not written to logs and no history file is created.
Previous clipboard contents
Before automatic pasting, your existing clipboard contents are held in memory; after pasting they are restored if the clipboard has not been changed by another action, then released.
Glossary and settings
Local files you maintain (~/.config/trie/glossary.txt and the configuration file), used only on this Mac and deleted by you; settings contain no dictated content.
Logs
~/Library/Logs/Trie/: event names, timings, counts, device settings and diagnostic paths (which may contain your user name); no audio or transcription content; never uploaded automatically; kept on this Mac and deleted by you.
Diagnostics export
Created on your Desktop only when you choose “Export Diagnostics…”, filtered through an allowlist, with no dictation content, full paths or glossary entries; whether to send it to us is your decision.

Permissions and control

Microphone
Captures speech. Without it, dictation is not possible.
Accessibility
Reads the necessary properties of the foreground application, window and input control to determine where to paste, and performs the paste; it does not read the text in the field. Without it, text cannot be pasted automatically.
Input Monitoring
Observes modifier-key changes and other key-down and mouse-down events to distinguish a solo right-⌘ tap from combined actions; it does not record typed text or a keystroke history. Without it, the right ⌘ cannot trigger dictation.
Stopping and clearing
Choosing “Standby” in the menu bar stops capture and returns headset buttons to the system; Trie returns to standby every time it starts; quitting Trie stops capture and zeroes the buffer. Logs, glossary, settings and diagnostics files are deleted by you; all three permissions can be revoked in System Settings at any time.

Network and service providers

Installation
The install script downloads the application from trie.ai (hosted by Cloudflare), installs uv from astral.sh and fetches Python dependencies from PyPI through uv, and downloads the models from huggingface.co or a mirror you configure.
Check for Updates
Only when you click: one request to https://trie.ai/download/latest.json, compared with the installed version. The request carries no identifier, uploads nothing, and is never automatic.
What providers see
These providers receive the IP address and request information needed to serve the request and handle it under their own privacy policies. The requests carry no dictation audio or transcribed text. We deploy no analytics scripts on the website, have not enabled the host’s traffic analytics, and keep no separate access logs.
Email
Messages sent to [email protected] are used only to respond to you and for no other purpose.

Known limits

Shared memory
The audio buffer lives in a POSIX shared-memory segment that processes of the same user on this Mac can read while Trie runs. If the process is force-quit or crashes while the system keeps running, the last 30 seconds of audio remain in the segment until Trie next starts and clears it; the segment does not survive a reboot.
Clipboard
Text reaches the target application through the system clipboard. Clipboard-history utilities may record it, and with Universal Clipboard enabled it may sync to your other Apple devices. Restoring the previous clipboard cannot recall copies other tools have already kept.
Target applications
Trie does not simulate Return, click Send or call a submit interface. The target application may still treat a paste as a submission (a terminal, for example) and may save or sync text under its own settings. If you switched windows after stopping, or the focus is in a password field, the text is only copied to the clipboard.
Developer evaluation
The developer may manually run a separate tool that submits independently prepared recordings to a cloud reference model for comparison. It is not invoked by the product and does not touch your device.

Responsibility, rights and contact

This notice is maintained by the Trie team; the responsible party’s name and location will be added in a later version. Basis for processing: on-device processing is initiated and controlled by you; website and download requests serve the service you request and its security; email is used to respond to you.

The current version has no account system and local dictation needs no sign-in. Should accounts, licensing or settings sync be introduced, we will explain the data processed, the purposes and your choices before enabling them, and obtain consent where the law requires.

Subject to applicable law you may request access, correction, deletion or a copy of your personal information and, where applicable, withdraw consent, restrict or object to processing; contact [email protected], or complain to the competent supervisory authority. Content stored only on your device cannot be accessed or deleted by us remotely. Trie is not directed at children. This notice does not limit your rights under applicable law.

We update this notice when our data practices change and revise the date at the top; material changes are announced on the website before they take effect.

Back to home