Privacy and safety

Freehand sends audio or text to the services you choose for transcription, optional cleanup, and on-demand speech. Each service may run on this computer, a private server, or a hosted provider. The server’s own privacy and retention policy still applies.

Check every stage

Local transcription does not make cleanup or speech generation local. Review the selected connection for each task before sending sensitive audio or text.

What goes where?

When you use…	What leaves Freehand	Destination
Microphone or audio-file transcription	Recorded or selected-file audio	Selected transcription endpoint
Optional cleanup	Completed transcript	Selected cleanup endpoint
Speak, Preview, or Listen	Your text, a sample phrase, or chosen transcript	Selected speech endpoint
Authenticated requests	Configured API key	Selected capability endpoint
Update checks	A release-metadata request; no recordings or transcripts	GitHub release service
Managed runtime/model installation	Explicit binary or selected-model download requests	Official pinned release and model hosts

Local retention has separate limits for audio, current results and history, settings, and runtime output.

Managed local recognition

Managed runtime listeners accept connections from this computer only. They are not exposed to your LAN, but other local software may still access a loopback service.

Part of the workflow	Keep it on this computer
Voice or audio-file recognition	Select the managed transcription Connection
Optional cleanup	Select managed llama.cpp with S1-mini, another local cleanup service, or leave cleanup off
Text to speech	Select managed NeMo with MagpieTTS, another local speech service, or leave speech off

• Downloads: runtime binaries and selected model weights stay in Freehand’s application-data directory until removed. Installation needs internet access; browsing the catalog does not run inference.
• Saved manual connections: remain available with their keys. Those keys are never sent to the managed runtime.
• Runtime failure: never silently redirects captured audio to a remote server. Recover local setup or explicitly select a manual connection for new work.
• Removal: runtime/model removal leaves source recordings, manual connections, and other applications’ model caches intact. Weights are installation data, separate from captured audio and transcript history.

What happens during selected-model GPU warm-up?

Starting a selected GPU runtime includes warm-up, including start-at-launch. llama.cpp and NeMo use built-in warm-up. CUDA whisper.cpp receives one second of synthetic silence on this computer, using only its selected loaded model; Freehand discards the response.

This captures no microphone audio, creates no history entry, invokes no other catalog models, and contacts no remote inference server. Connection checks remain metadata-only.

Audio

Audio source	Freehand’s lifetime
Microphone capture	Active transcription request only; released and temporary audio deleted after success, failure, or cancellation
Selected audio file	Active request only; original file is never modified or deleted
Generated speech	Memory until cleared, replaced, a recording begins, or Freehand exits; saving a file is explicit

Audio is never retained in transcript history. Your chosen inference server may retain audio or text under its own policy.

Transcripts and history

Text	Where it lives	When it clears
Latest dictation result	Memory; available with history off	Next recording, Clear, or exit
Current file result	Memory; available with history off	Clear/replace the file, another transcription, or exit
Unsent speech-composer draft	Current window session; survives task/settings navigation	Window reload or exit
Optional transcript history	Memory; off by default, at most 20 entries and 2 MiB	Removal, disabled/cleared history, retention limits, or exit

Composer text is never written to browser storage or the settings database. History contains raw and cleaned text with limited non-secret run details. It contains no audio, credentials, request headers, full paths, or destination identity. Transcription details disappear with their history entry; an open details window does not keep a separate copy.

Keeping both versions after successful cleanup requires enabled history and space within its limits. Raw fallback after failed cleanup and failed-insertion recovery work with history off.

File results always need explicit Copy. Voice results use focus-safe insertion or manual copy according to your settings. See session history to enable, compare, and remove retained results.

Safe text insertion

Start dictation with the intended app and text field focused. Keep the destination focused until delivery finishes; Freehand never brings it to the foreground for you.

Platform	What must remain focused
Windows	Original app, window, and focused control
macOS	Same app and window; moving to another field in that window delivers to the currently focused field

Release physical shortcut modifiers before delivery. Windows waits briefly for held Ctrl, Alt, Shift, or Windows keys; if they stay held, use Copy.

Secure Input cannot identify every password field

macOS Secure Input blocks delivery while active. Custom secure fields that do not enable it may not be detected. Avoid dictating sensitive text into an uncertain destination.

If the destination changes or safe delivery is unavailable, the result stays available for explicit Copy. Check for partially inserted text before pasting to avoid duplicates.

Clipboard-paste insertion is disabled. Automatic delivery does not silently replace your clipboard.

Saved settings and backups

Platform	Settings and backups	Window geometry
Windows	%LOCALAPPDATA%\Freehand\freehand.db	%APPDATA%\Freehand\window-state.json
macOS	~/Library/Application Support/Freehand/freehand.db	window-state.json in the same directory

The database stores non-secret configuration such as server addresses, model choices, vocabulary, custom instructions, and request headers. API keys remain in the native credential store; history stays in memory.

Settings and backups are private, not encrypted

Windows database access is restricted to your user and SYSTEM. macOS uses per-user file permissions. Neither database is encrypted, and Keychain denial never triggers plaintext credential storage.

Upgrade or recovery action	Retained data
Earlier alpha settings and native credentials	Left untouched, not imported or reused; see the first-launch reset notice
Future schema upgrades	Up to three database backups
Explicit recovery reset	An archive of the replaced current database

Read settings recovery before restoring or removing these files.

Credentials and transport

Protection	What it means
Native credential storage	Saved keys go to Windows Credential Manager or macOS Keychain, never the settings database, and are not displayed again
Inactive connections	Keep their saved keys; deleting a connection removes its key only when no saved connection still uses it
HTTPS by default	HTTP requires explicit permission per connection
No HTTP redirects	Inference and connection checks never follow redirects, even to another path on the same server; configure the final base URL

HTTP sends content without encryption

Allow it only for a trusted local or LAN endpoint. Audio, transcript text, and credentials are unencrypted in transport; do not allow HTTP across an untrusted network.

Freehand filters literal API-key copies from server response details and rejects transcript text containing the key, including realtime captions and finals. Realtime checks span messages and model-specific parsing. This cannot make an untrusted server safe: it already received the key and could misuse or transform it. Connect only to services you trust. See saved connections for credential management.

Connection checks

Trigger	Behavior
Finish initial Voice setup	Requires an explicit connection test; Audio file and Text to speech remain independently usable
Launch after Voice setup, or relevant saved connection changes	Automatically checks the saved STT connection
Automatic or manual connection check	Reads a health route or GET /v1/models
Select a model	Does not invoke it

Checks send no audio, prompts, or synthetic inference jobs and never cycle through discovered models. Startup warm-up belongs to the selected managed runtime, as described above.

Update checks

Update action	Timing or effect
Automatic checks	On by default; about 30 seconds after startup, then daily after a successful check
Failed automatic check	Retries after 15 minutes while Freehand is running
Development build	Does not check automatically
Manual check	About → Check now
Disable automatic checks	Settings → General

Checks read GitHub release metadata without sending recordings or transcripts. When an update is available, the updater can download and checksum-verify the platform asset, then waits for you to restart. macOS uses an app-bundle ZIP; checksum verification does not establish Developer ID trust or notarization. See macOS updates or Windows updates.

Diagnostics

Application logs

Operational logs record status, timing, and failure categories. They exclude audio, transcripts, credentials, private headers, full paths, model IDs, URL paths/queries, and destination-window identity.

Update helper logs

When you restart into an update, Wails separately writes wails-update-<pid>.log in the operating system’s temporary directory.

Included	Excluded	Lifetime
Installation/temporary paths, process IDs, replacement errors; paths may reveal your account name	Recordings, transcripts, inference API keys	Remains until you or the operating system remove it

Review that file before sharing it for support.

Managed runtime output

Process output can contain sensitive content

The runtime may write transcripts, prompts, paths, or other private upstream text. Freehand cannot promise complete redaction. Turning transcript history off does not prevent that text appearing in output, and opening a viewer during screen sharing can expose it.

Output storage	Limit or lifetime
Private rolling tail	Memory only, at most 256 KiB and 1,024 chunks; oldest text drops as it fills
Collection	Continues with the viewer closed or scrolling paused
Clear, next start attempt, runtime removal, or Quit	Discards the private tail
Application logs, events, and crash reports	Never receive the captured output

For llama.cpp, capture includes normal information, warnings, and errors rather than debug logging. The private tail is not saved as a log file.

Viewer	When text is shown
Embedded Runtime output tab	Immediately on opening or switching runtime tabs; retains the selected runtime across workflows, Connections, Local runtime, and History
Standalone Process output window	Requires Show output each opening or runtime switch
Hidden viewer, different tab/runtime, global Settings, or hidden workspace	Clears displayed text and revokes reads; leaves the runtime and private tail running
Reopened embedded tab	Immediately reads the available tail again

Both viewers are read-only, with search, colors, and progress updates. They have no command input or file logging. Copy selection copies only your selected text when you ask; other applications may read it, and it may remain on the clipboard after the viewer closes.

See runtime output controls for viewing, clearing, and copying output.