Keep your dictation on your Mac.

What happens in Local mode

Speech recognition and optional Polishing are separate choices:

Why it's worth it

• Privacy that isn't a checkbox. Your voice is often the most sensitive thing you type — passwords read aloud, medical notes, half-formed ideas, other people's names. When transcription is local, that audio never becomes someone else's server log, training set, or breach.
• It avoids a network round trip. Local transcription does not wait for an upload or a server response, and it works the same on a plane, a train, or bad hotel Wi-Fi.
• It works offline. Cloud-only dictation stops working without a connection. Local transcription does not, once its model is downloaded.
• No hosted lock-in for the core function. Local mode can use Whisper or Parakeet without depending on a speech provider's pricing or availability. Polishing can remain off or use an endpoint you choose.
• No meter running. Local dictation is free and unmetered — no per-word caps, no monthly subscription for the core function.

What hosted choices send

• Your own provider. Speech sends recorded audio, a language hint, and an optional vocabulary prompt directly to Groq, deAPI, OpenAI, or a compatible custom endpoint. Hosted Polishing sends transcript text, vocabulary terms, and any enabled translation-target or email-layout instructions directly to Google Gemini, Anthropic, OpenRouter, OpenAI, or the compatible endpoint you configure.
• Plainsay Cloud transcription. Encoded audio, clip duration, language, and an optional vocabulary prompt go to api.plainsay.app. Speech may run on a Plainsay-operated transcription node or be forwarded to deAPI.
• Plainsay Cloud Polishing. Transcript text and vocabulary terms go through Plainsay Cloud to OpenRouter, using its configured Gemini model. Since v0.2.33, this request also carries the Auto-translate target and the email-layout choice, as structured fields the server turns into prompt wording.
• Auto-translate and automatic email layout. A Cloud subscription unlocks these styles. Their instructions go to whichever Polishing provider is selected — a BYOK hosted provider, a compatible local endpoint, or, since v0.2.33, Plainsay Cloud. None of them send recorded audio anywhere new.
• Voice Filter is best-effort. It runs on the Mac. If it is unavailable or fails, dictation continues; a selected hosted speech service can then receive the unfiltered recording.

What is stored

Plainsay Cloud keeps the records needed to run an account and subscription: an email address and/or Apple account identifier, session records, Stripe customer and subscription identifiers and status, and per-operation usage metadata such as timestamp, type, audio duration, estimated cost, and Polishing token counts. Its application database deliberately stores neither recorded audio nor transcript text. Stripe receives the account email and handles checkout and payment details; Apple handles Apple sign-in, and the configured email-delivery service receives the address and sign-in message.

API proxy access logging is disabled. Minimal application event logs contain only the method, matched route template, status, and elapsed time; they exclude network addresses, headers, query strings, and request bodies. Website and signed-update delivery can still produce ordinary web-server access records.

On the Mac, History is enabled by default and keeps up to 100 raw and final transcript records for 30 days. You can choose 7, 30, 90, or 365 days, clear the records, or turn History off; switching it off deletes saved transcripts and stops future saves. Audio is staged in an owner-only local folder during processing and then removed. Interrupted recordings can remain for recovery; on each launch, Plainsay removes ones older than 7 days and keeps no more than the newest 20. Clearing History purges those recordings too. History and pending audio are excluded from backups.

Local mode can still contact the network

The selected speech model is downloaded once. By default, Plainsay also contacts api.plainsay.app to check for signed updates; you can disable automatic checks in Settings. Those connections do not upload a dictation. The Mac client sends no analytics or telemetry. Provider API keys and the Plainsay Cloud session token are kept in macOS Keychain.

You don't have to take our word for it

A closed cloud app can promise anything in a privacy policy. Plainsay's entire Mac client is MIT licensed and public — you, or anyone you trust, can inspect each speech, Polishing, storage, and update path. Plainsay Cloud is a separate hosted service, so this page describes that boundary instead of implying the server is covered by the client's MIT license. If that mattered enough to read this page, it's worth reading the source.

github.com/conrader/plainsay — inspect the MIT-licensed Mac client

How Plainsay verifies the speech models it downloads →

The contrast, in one sentence

Tools like Wispr Flow state plainly that “transcription always happens in the cloud” — every word leaves your device by design. Plainsay defaults to Local, while its hosted transcription and Polishing paths remain separate choices you can see and change. See the full comparison →