How Sharp works

From command to action

1. You type or speak

   Up to 500 characters of English. Speech is turned into text on the phone by Whistle, the built-in speech model.

2. Sharp narrows it to a group

   Time, People, Device, Apps or Create.

3. The router picks the action

   Needle 3, the built-in router model, chooses one action in that group and fills in its values, such as a time or a contact.

4. Values are checked against your words

   A time or name you didn’t say stays empty for you to fill in. Sharp doesn’t guess.

5. Act, card, or “I’m not sure”

   A low-risk action can run when Sharp is sure. Otherwise you get a card. If nothing fits well enough, Sharp says “I’m not sure what to do”, suggests the closest actions and keeps your words to edit. “Why this?” on a card says which group matched and which values came from your words.

The 20 actions

Every command becomes one of these 20 actions. Messages, emails, calls, calendar events and rides always show a card first, and so do alarms and timers you set. The rest can run straight away when Sharp is sure of the command. Turn on Always confirm in Settings to put every action on a card.

“Can run straight away” means only when Sharp is confident and every value came from your words. Even these actions often show a card instead.

Cards and hand-offs

When an action belongs in another app, Sharp opens that app on the right screen with your text, search or destination filled in, and the card names the step left to you, such as “you tap Send”. Sharp passes only what that action needs. Hand-offs can’t be undone from Sharp, because the other app owns them.

A card is a sentence built from your own values. Every field can be edited in place, and the button names the action: Add event, Open message, Call, Open WhatsApp. Undo is offered for the flashlight, the volume and calendar events Sharp adds directly; music offers Pause instead.

Sharp has hand-off routes for WhatsApp, Telegram, Signal, Spotify, YouTube, YouTube Music, Amazon, Flipkart, Zomato, Swiggy, Uber, Ola, Rapido and Google Keep, plus your phone’s own messages, email, maps, camera and clock apps. If the app you name isn’t installed, Sharp says so and offers its Play Store page instead of picking another app. Only two routes have been tested, on an Android emulator; see hand-offs in Inputs and limits.

Ways in

You can open Sharp from its icon, from the “Assistant” tile in Quick Settings, or, if you make it your default digital assistant, with a long-press of home or power. The long-press opens a sheet over the app you’re using, already listening, with your first four pins.

To talk, hold the mic and let go to stop, or tap it once: it stops after a short silence, a second tap, or 30 seconds. Sliding off the mic or tapping Cancel discards the audio. Where Android won’t show the “make default” dialog, the row in Sharp’s Settings opens Android’s “Digital assistant app” screen instead.

In the sheet, an action that opens another app closes the sheet. A text answer stays about 6 seconds, and a result with Undo stays at least 6 seconds. Speak results, off by default, reads a short confirmation aloud with your phone’s text-to-speech engine.

Models

Sharp comes with two built-in models: Needle 3, which understands your commands, and Whistle, which turns your voice into text. Settings → Models lets you download others. Gemma 4 E2B, about 2.6 GB, can also answer a plain question in text, and a second speech engine is available. Each download starts only when you tap it.

Before a download, Sharp shows the size, licence, suggested memory and the host, Hugging Face or GitHub, and warns you on mobile data or a metered connection. The host receives the request like any web download. You can also import a model file you already have; Sharp checks that it is the right file before keeping it.

Gemma 4 E2B suggests 6 GB of memory, and Sharp warns that it may be slow or fail on phones with less. Its answers are written by a language model on your phone, are labelled “Answered on this phone”, can be wrong, and never trigger an action. For speech you can also pick your phone’s own recognizer, on Android 12 or later where the phone offers on-device recognition.

Permissions

Sharp asks for a permission only when a command first needs it, after a one-line reason, and doesn’t ask again unless you start it from Settings. Each permission has a route that works without it: if you say no, typing or another app takes over that step, as the table shows.

Sharp also uses the internet only for model downloads you start. It has no permission to send text messages, read your call log, use the camera or read your location. Full list in Sharp’s privacy policy

Inputs and limits

Sharp matches your words to one action and shows you what it understood, but it can misread a command, mishear a name or not understand a phrasing at all. These are the limits we know of. The card is there so you can catch a mistake before anything happens, so read it before you confirm.

• English, one command, 500 characters. Sharp understands English only, one command at a time, up to 500 characters. “Set a timer and text Mom” is two commands.
• Some phrasings aren’t understood. Some short or unusual phrasings aren’t understood, such as “standup at 9” or “louder”. Sharp then says it isn’t sure, suggests close matches and leaves your words to edit. Naming the action helps: “add standup at 9”, “volume up”.
• It can pick the wrong action. Sharp can misread a command, for example taking “alarm please” as a request to show your alarms. That is why messages, emails, calls, events and rides always wait on a card, and why the card says in plain words what it will do.
• It asks more often than it acts. Even low-risk actions often show a card when Sharp isn’t confident enough. Always confirm in Settings makes every action wait for a card.
• Values it can’t work out are left empty. If Sharp can’t tell who you mean or what time, it leaves that field empty on the card for you. Check repeat days on alarms: “every weekday” may not be filled in.
• Voice can mishear. Speech is turned into text on the phone, and it can mishear, especially names. The card shows the words it heard.
• Hand-offs depend on the other app. Only the YouTube and YouTube Music hand-offs have been tested, on an Android emulator. The other 14 routes, including WhatsApp, Telegram, Signal, Spotify, Amazon, Flipkart, Zomato, Swiggy, Uber, Ola, Rapido, Google Keep and the Play Store, have been checked on paper but not run yet. Hand-offs use each app’s public links, and an app update can change them, so check that the app opened on the right screen.
• What Android doesn’t allow. Sharp can’t switch Wi-Fi, Bluetooth or NFC on or off; it opens the panel. It has no wake word, doesn’t read your screen, and doesn’t pay, buy or book for you.
• x86_64 devices. On the few Android devices with an x86_64 processor, such as some Chromebooks, the built-in models can’t run. You need the optional Gemma 4 E2B download to type commands, and another speech engine to talk.
• Gemma 4 E2B answers. Gemma 4 E2B is an optional download of about 2.6 GB that suggests 6 GB of memory. Its text answers are written by a language model on your phone and can be wrong, and an answer never triggers an action.
• Without a permission. With no calendar permission, the Calendar app opens with the event filled in; with no phone permission, the dialer opens. A person named in an event goes in the title, and no invitation is sent.
• Pins and history. Up to 8 pins. History keeps your last 500 commands.
• Android version. Android 10 or later, on Android phones. The option to use your phone’s own speech recognizer appears only on Android 12 or later, on phones that offer on-device recognition.