Skip to content

Reference

Privacy & offline

Understand what stays local, what uses a network, and how to work offline.

On this page

OpenWhisper records, transcribes, and stores on your computer. There is no telemetry, analytics, or crash reporting. The only connection made without you turning something on is the update check; everything else starts from a feature you enable, a download you approve, or an action you take.

Understand each connection#

FeatureWhat uses a network
Update checkOn by default: asks GitHub for the latest release at launch and every 4 hours, sending only the app version. Turn it off in Settings → App → General → Updates.
Local speech modelThe one-time model and runtime downloads. Installed models load from the local cache without checking online, and Hugging Face Hub telemetry is turned off.
API speech engineYour recording goes to the selected server: OpenAI, Groq, or a server you run. A server on localhost keeps audio on this computer.
Remote engineAudio goes to your paired host for transcription, plus dictionary hints for a Whisper host.
AI cleanupTranscript text and cleanup instructions (dictionary terms, learned rules) go to your cleanup model. A server on this computer keeps it local.
Command Mode and transformsSelected text and your instruction go to the cleanup model, even when standard cleanup is off.
Coding agentsClaude Code, Codex, or OpenCode, when you choose them for cleanup, meetings, walkthrough write-ups, or actions, run as themselves and talk to their own services.
Meeting AI insightsThe meeting engine processes transcript text, your brief, and any enabled recall. AI insights never upload audio.
Fast judgmentsTranscript excerpts go to TypeSafe (directly or through OpenRouter), or to your own text model if you choose Judge: Your text model.
OpenAI speaker labelsWith explicit consent, the other side's meeting audio is uploaded after the meeting.
Actions and hooksWebhooks, text filters, and Open a link actions send your words to addresses you configured.
Host record storageRecords transfer to the paired host when you choose Host or Both.
API & MCPListens on 127.0.0.1 only (plus Tailscale if you turn that on), behind an access token. MCP Events sends events to addresses a connected agent chooses.

Choosing a local speech engine doesn't make the other features local. Review each one you enable. The app's network and privacy reference lists every destination, port, and default.

When you choose a hosted provider (OpenAI, OpenRouter, Groq, OpenCode Go or Zen, TypeSafe) or an installed coding agent, your requests go straight to that service under its own terms and privacy policy. To keep everything on your computer, use a local speech engine and a local text server.

App awareness and nearby text#

Settings → Personalize → Styles controls app awareness. App detection uses local window information; app names can be saved with History and used by Stats.

Read text near the cursor is a separate, off-by-default Windows option. Nearby text goes only to your cleanup model, never to the speech engine or a remote host, and isn't saved in History. Password fields, terminals, OpenWhisper's own windows, and apps on Never read from are excluded.

Command Mode and transforms deliberately read the text you select. Their selection and result can be saved to History.

Prepare for offline use#

  1. Choose a local speech engine, or the API engine with a speech server on this computer.
  2. Download its model and every required runtime in Settings → Models & storage → Downloads while you're online.
  3. Use a text server on this computer (Ollama, LM Studio, vLLM, llama.cpp, LocalAI) for cleanup, profiles, Command Mode, transforms, and meetings, or turn those off.
  4. For meetings, either turn Fast judgments off or set Judge: Your text model on a local server. Use Off or On-device speaker labels; start one short meeting online first so the on-device speaker model is cached.
  5. Turn off the update check, and set When a model is missing from this computer to Never connect.
  6. Disconnect and make a short test recording.

A remote engine still needs a reachable host, even without internet.

Control model downloads#

Settings → Models & storage → Downloads → When a model is missing from this computer chooses whether to ask (the default), allow, or never connect.

For source launches, setting HF_HUB_OFFLINE=1 before launch blocks model downloads regardless of settings. Neither option is a global network switch: cloud transcription, network text providers, and the update check have their own settings.

Know where credentials and records live#

API keys added in Settings are stored in Windows Credential Manager, macOS Keychain, or the Linux Secret Service, never in a plaintext settings file. When macOS asks to use your Keychain, choose Always Allow; Allow covers one read, so macOS asks again at every launch. If the Linux keyring doesn't respond, OpenWhisper opens without saved keys and says so. Source setups can also use environment variables or a .env file; protect those yourself.

History, transcripts, recordings, walkthrough captures, dictionary entries, snippets, Scratchpad notes, and Stats are local application data. With a remote engine, Where records are kept can move or copy records to the host; see Remote engines.

Exports contain transcript content and recording references, not audio. Share them according to what they contain.

Review downloaded components#

Downloads shows each component's publisher, source, version, license, and integrity check. Every file is pinned to a size and SHA-256; that confirms you got the expected file, not that the model or runtime is secure. Keep the app and its runtimes updated, and review important transcripts for errors.

To report a security issue privately, see the security policy.

Checked against OpenWhisper 2.8.97.App documentation
Still stuck?

Get help from the community, or tell us what went wrong.