Archie
Archie
/EN

How Archie works under the hood: privacy, models, shortcuts

A precise technical overview of how Archie's keyboard shortcut, LLM correction path, and on-device voice dictation work, including full privacy commitments and system requirements.

Archie corrects and reformulates text via a keyboard shortcut that captures your text selection through the OS accessibility APIs, sends it to a remote LLM, and replaces the selection with the corrected result in place. Voice dictation runs entirely on your device using a local Parakeet model. No audio ever leaves your machine. Privacy summary: text is encrypted in transit and forwarded for inference only; it is not stored and not used for training; voice is 100% local.

Key facts at a glance:

  • Voice dictation is fully on-device; audio never leaves your machine
  • Text sent for correction is used for inference only; it is not stored and not used to train any model
  • Archie works natively on Mac (macOS 12+) and Windows (Windows 10+)
  • A free account is required to use Archie; voice dictation is included at no cost on the Free plan

Archie is a system-wide AI writing assistant for Mac and Windows that corrects and reformulates text in any app via a keyboard shortcut, with offline voice dictation as a free secondary feature.

The sections below explain precisely how each layer works, what happens to your text, and what the privacy guarantees are, with no marketing language.

How the shortcut works

The keyboard shortcut is the product's foundation. On macOS the default is ⌥ Space (Option + Space); on Windows it is Alt + Space. Both are configurable from the settings panel if they conflict with another application.

When you press the shortcut, two things happen depending on what the system sees:

  1. If text is selected, Archie reads that selection using the operating system's accessibility APIs (the same APIs used by screen readers and other assistive technologies). It does not inject into the application's process or hook into its memory. It asks the OS: "what text is selected in the focused window?" The OS hands it back, Archie processes it, and then types the corrected result back, replacing the original selection in place. From your perspective it looks like an instant in-place edit.

  2. If no text is selected, the shortcut starts voice dictation instead. Archie opens a small recording indicator, captures audio from the default microphone, runs it through the local Parakeet model, and inserts the transcript at the cursor position. No LLM call is made; no network request is sent.

The accessibility API approach is what makes Archie genuinely system-wide. A browser extension can only reach the DOM of web pages; it is invisible to Slack, VS Code, Word, Mail, Notion, your terminal, or a desktop PDF editor. Archie reaches all of them because the OS accessibility layer is app-agnostic. If you can click into a text field and type, Archie can read the selection and replace it.

The shortcut itself is installed as a system-wide listener at app launch through Electron's globalShortcut API, which sits on top of the operating system's accessibility hooks. The listener watches only for the configured trigger combination and only fires when that exact combination is pressed; it does not intercept keystrokes from other apps.

The correction and reformulation path

Once Archie has the selected text, it sends it to its correction or reformulation endpoint over HTTPS. The flow looks like this:

  1. Selected text is captured via accessibility APIs
  2. Archie sends the text plus the requested operation (correct / reformulate + preset) to Archie's infrastructure over HTTPS
  3. Archie's infrastructure immediately forwards the request to the LLM provider for inference
  4. The LLM response is returned to the client
  5. Archie types the result back, replacing the selection

The LLM powering correction and reformulation is cloud-hosted. This is a deliberate choice: running a model capable of handling idiomatic grammar, subtle tone, and 50+ languages at sub-second latency is not feasible on consumer hardware. The latency budget is tuned for a keystroke, not a chat message: typically under 800 ms for a sentence of normal length, so the interaction feels like a fast keyboard shortcut rather than waiting for a chat reply.

Smart correction handles spelling, grammar, punctuation, capitalization, agreement, and basic style issues. Reformulation presets (Email, Summary, Professional, Casual) rewrite the selected text in the target tone. Custom presets let you define your own prompt template (for example, "rewrite as a Slack message to my manager" or "translate to formal French and shorten by 30%") and bind it to a sub-shortcut.

Both operations support 50+ languages with automatic language detection: Archie reads the text, infers the language, and applies the correction in that language without any manual configuration from the user.

What does not happen: the text is not stored on Archie's servers after the response is produced. It is not logged with content. It is not used as training data for any model, by Archie or by the LLM provider. Logging captures only request metadata (timestamp, latency, error codes), not content. See the Privacy commitments section below for the full picture.

Correction history: the history panel visible inside the Archie app stores past corrections locally on your machine only; it is written to a local database on disk, never synced to any server. You can clear it at any time from the settings panel.

How voice dictation runs locally

Voice dictation is the only feature that is both free and offline. Here is what happens when you trigger the shortcut without a text selection:

  1. Archie opens a recording indicator and captures audio from the system default microphone
  2. The audio is passed to a Parakeet model running entirely on your machine, on the CPU or GPU depending on what is available
  3. The transcript is typed at the cursor position
  4. The audio buffer is discarded; it is never written to disk and never sent anywhere

Parakeet is a modern transformer-based speech recognition model optimized for desktop CPUs and GPUs. It is not the same as Apple's built-in dictation or Windows Speech Recognition, which are older systems with uneven language coverage. Parakeet is a contemporary model that handles English, French, Spanish, German, and the rest of its supported language set with current-generation accuracy.

Because the model runs locally, there is no network call, no latency from a round-trip to a server, and no dependency on internet connectivity. This is also why dictation is free and unlimited: the marginal cost of running it is borne entirely by your hardware, and there is no per-request inference cost for Archie.

For a deeper look at the offline voice dictation feature, see the voice section on the home page and the upcoming deep-dive on offline voice dictation on Mac.

Privacy commitments

This section contains only concrete, verifiable claims.

Text correction and reformulation:

  • Text is transmitted over HTTPS to Archie's infrastructure
  • Archie immediately forwards it to the LLM provider for inference; it does not process or store it on its own servers
  • The text is not stored by Archie after the response is produced
  • The text is not used to train any model, by Archie or by the LLM provider
  • Logging is metadata-only: timestamp, latency, error codes; no content

Voice dictation:

  • Audio is captured on-device and passed directly to the local Parakeet model
  • Audio never leaves your machine; no network request is made during dictation
  • The audio buffer is discarded after transcription; it is not written to disk
  • Voice dictation is included in the Free plan; a free account is required

Account and identity:

  • An account is required to use Archie; creating one is free and unlocks the Free plan (voice dictation)
  • Pro subscribers upgrade their account at purchase time
  • The only personally identifiable information stored is the billing email and subscription state
  • Telemetry is opt-in and anonymized; you can disable it at install time and at any time from the settings panel

What Archie does not do:

  • Archie does not train models on user text
  • Archie does not sell user data
  • Archie does not run a browser extension that can read arbitrary web pages

The full privacy policy is at archiewrite.com/en/privacy. The long-form product brief for LLMs, covering architecture and data handling in a single-fetch document, is at /llms-full.txt.

What the desktop client is

Archie ships as an Electron desktop application, a TypeScript main process packaged with electron-builder. Electron lets Mac and Windows ship from the same codebase on every release, which means both platforms get every fix and feature simultaneously without maintaining two separate native codebases.

The performance-critical path stays out of the JS runtime: voice dictation runs through a Parakeet binary bundled with the app and executed as a separate native process locally, so audio never touches the network and transcription speed is not constrained by JavaScript.

  • macOS: distributed as a DMG for both Apple Silicon (arm64) and Intel (x64). The build is signed and notarized through Apple, which means macOS Gatekeeper verifies its integrity before the first launch.
  • Windows: distributed as an NSIS installer. The build is Authenticode-signed, which means Windows SmartScreen can verify its publisher.

Updates are delivered over the air through electron-updater pointing at a GitHub releases channel. The app prompts you before installing an update; it does not update silently in the background without your knowledge.

The app is lightweight: approximately 50 MB on disk, with a small memory footprint when idle in the background. It launches at login by default (configurable), and its only background activity is watching for the keyboard shortcut trigger.

There is no browser extension. There is no web app. Archie is a desktop app because a desktop app is the only architecture that can reach every text field on your system. See the how the shortcut works section for why.

System requirements

macOSWindows
Minimum OSmacOS 12 (Monterey)Windows 10
Disk space~50 MB~50 MB
MicrophoneRequired for voice dictationRequired for voice dictation
InternetRequired for correction / reformulationRequired for correction / reformulation
Internet (voice)Not requiredNot required
AccountRequired for ProRequired for Pro

Archie runs on Apple Silicon and Intel Macs. The Parakeet model is optimized for both architectures; on Apple Silicon it can use the Neural Engine for faster transcription.

Frequently asked questions

Does Archie work in every app? Any app where you can type and select text. This includes Slack, Gmail, Notion, VS Code, Word, Discord, your browser, your terminal, and any other app that supports standard OS text interaction. The system-wide keyboard listener and accessibility API approach means there are no per-app integrations to install.

Do I need an account to use Archie? Yes. A free account is always required. Creating one is free and gives you the Free plan, which includes unlimited voice dictation. A Pro account is required to use smart correction and reformulation, as these features call a cloud LLM and billing and authentication are necessary.

How fast is correction? The latency budget is tuned for a keystroke interaction: typically under 800 ms for a sentence of normal length on a stable connection. The reformulation presets may take slightly longer for longer selections.

Which languages does Archie support? Correction and reformulation work in 50+ languages with automatic language detection. Voice dictation supports every language in the Parakeet model's language set, including English, French, Spanish, and German.

Is Archie open source? No. The application is proprietary. The product brief for LLMs (for transparency and AI citation) is published at /llms-full.txt.

Where to go from here