Command Mode

The chord and the gesture

Dictation is the Fn (globe) key. Command Mode is that same key plus Ctrl. The chord is a strict superset of the dictation key, so Command Mode is literally dictation, plus Ctrl.

The gesture follows whatever your dictation hotkey already does. On the default Fn key that is a toggle: press Fn and Ctrl together, speak, then press Fn again to stop. If your dictation hotkey is a push-to-talk chord instead, the command chord is push-to-talk too. There is no new gesture to learn; it is the one you already have, with one more finger.

Two details worth knowing:

• The chord is rebindable under Settings › Command Mode. A new binding applies the next time you launch Evoglyph, and the picker refuses a chord that would collide with your dictation hotkey. See hotkey configuration.
• A command in flight is blue. The floating indicator, the menu bar icon, and the status dot in the popover all pick up Evoglyph's command color, so you can always tell a command from a dictation at a glance.

Rewrite, write, scratch that

Command Mode answers three questions that plain dictation cannot.

• Rewrite this. With text selected, speak an instruction and Evoglyph rewrites the selection in place. "Make this more concise." "Fix the grammar." "Turn this into bullet points."
• Write this. With nothing selected, speak an instruction and Evoglyph generates new text at your cursor. "Write a friendly reply declining the meeting."
• Do this. Speak a device command and Evoglyph performs a system action on your Mac. That is the catalog below.

There is also "scratch that", which works two different ways. The distinction is worth being precise about, because the defaults differ:

• Through the command chord, "scratch that" is one of the twelve actions below, and it is on by default. Evoglyph classifies it like any other command and removes your last dictation.
• During ordinary dictation, with no chord held, "scratch that" and "press enter" can act as inline spoken shortcuts. These are deterministic phrase matches, not a model, and they are opt-in and off by default. Until you turn one on, the phrase is typed out literally, and Evoglyph offers a one-time prompt to enable it.

What you can ask it to do

These are the twelve actions Command Mode can perform. Evoglyph classifies what you said into exactly one of them, or declines. The example phrases are illustrative, not magic words.

The catalog is bounded on purpose. It is a fixed, built-in set, not a list you author, and it does not grow at runtime. There is no arbitrary shell and no free-form system access. "Open link" accepts only http and https addresses. "Open Settings pane" resolves against a curated allow-list of panes rather than an arbitrary identifier.

Running your own Shortcuts is the extensibility hatch. Anything you can build in Apple's Shortcuts app, you can trigger by voice, and Evoglyph runs it by its stable identifier rather than its display name, so renaming a Shortcut cannot quietly re-grant trust to a different one.

How Evoglyph decides it was a command

The thing that makes "run your Mac by voice" trustworthy is that Evoglyph is conservative by construction. What you said passes three gates, in order, before any action fires.

1. A deterministic veto, with no model involved. Injection-shaped input is barred from ever becoming an action and is routed as ordinary text instead. Plain pattern and token checks catch forged delimiters, role prefixes like system:, and chat or tool control tokens. This is the most dangerous class of false positive, so it is caught before a model ever gets a vote on it.
2. "Is this a command?" A yes-or-no pass with no list of actions in context. That absence is the point: showing a small model the catalog primes it to fire on any look-alike sentence, so "launch the product" and "he runs a food truck" start becoming actions. Asking a plain yes or no first lets the model decline everyday speech and still catch real commands.
3. A strict parse into a typed action. Only on a yes does a second pass produce a strictly validated action: exact keys, correct types, in-range values.

Confirmation, undo, and the record it keeps

Confirmation. Quitting an app always confirms. Running a Shortcut confirms the first time you run that particular Shortcut, then remembers it. A confirmation is a small floating Confirm or Cancel card that does not steal focus from the app you are in. The token behind it is single-fire and expires after 120 seconds; if it expires, the card simply disappears and you say the command again. Only one confirmation is ever pending at a time.

Undo. The reversible actions are reversible. Set volume and adjust volume restore your exact prior level and your prior mute state, and mute and dark mode reverse cleanly too.

The record it keeps. Every command writes a compact record into your local history: which action ran, the parameters it ran with, and how it turned out. A "set volume to 50" stores something like {"v":1,"action":"setVolume","args":{"level":50},"outcome":"executed"}. Being precise about what that does and does not mean:

• It does record the spoken parameters of an action: an app name, a URL, a volume level, a Shortcut name. It is not content-free.
• It never records a command's output. The standard output and error of anything Evoglyph runs on your behalf never reach it.
• It stays on your Mac like the rest of your history, and Clear History in Settings erases it along with everything else.

Refused commands write a row too, so "it keeps not catching that" is something you can point at rather than a feeling you have.

Two models, split by what each is good at

Evoglyph runs several small local models rather than one big one, and Command Mode is the clearest case of why. No single model was best at everything, so the work is split.

• The command text work (rewriting a selection, generating new text) runs on LFM2-2.6B, the same base model behind cleanup, here running adapter-free with no adapter on top. It is strong at that work and was unreliable at classifying actions: it over-triggered.
• The command action work runs on Qwen3-1.7B (mlx-community/Qwen3-1.7B-4bit-DWQ, a 4-bit DWQ build, about 0.98 GB, Apache-2.0, also adapter-free). It is excellent at action classification. Tried as the whole command brain it was wrong for the writing path, echoing knowledge questions back verbatim.

Each model is used only where it wins. Both run on the Apple GPU through MLX, in-process, with no network call at inference time.

Because Command Mode ships on by default, the Qwen3 model streams during onboarding alongside the transcription and cleanup models rather than arriving later. It is pinned to a specific revision and checked against a bundled hash registry before it is used, the same as every other model Evoglyph downloads. For what that adds to the total, see model download.

On-device, and free

Nothing about Command Mode changes Evoglyph's privacy posture. Audio, transcripts, and commands never leave your Mac. There is no cloud round trip to classify what you said, and none to decide whether to act on it. See where your audio goes.

Command Mode is also mostly free against your trial word budget, and precisely so. System actions inject no text and are not dictation, so they charge zero words. The spoken shortcuts during dictation charge zero. Only the text work costs anything, and it only costs the net new words it added: a rewrite that shortens your selection charges nothing, a rewrite that expands it charges the difference, and writing fresh text charges its full output, because there was no input to subtract.