Alfred

A personal Android voice assistant that talks to your n8n workflows.

Alfred replaces your phone's assistant shortcut with a lightweight overlay — long-press the side button, speak, and what you said is delivered to a self-hosted n8n webhook where an AI agent (or any workflow) does the actual work. Built with Kotlin and Jetpack Compose for a Samsung Galaxy S24 Ultra, but should behave on any Android 14+ device.

Modes

Mode	How it works
Task	Speak a task; capture auto-stops when you pause. Sent as type: "task", response arrives as a notification.
Note	Open-ended dictation with a manual stop button. Sent as type: "note".
Convo	Back-and-forth conversation — your question is sent as type: "convo" and the workflow's reply is spoken aloud via TTS, then Alfred listens again.

The overlay remembers your last-used mode. Task and Note sends run through a foreground service, so you can switch apps immediately — the result lands as an expandable notification and is recorded in a history screen (tap the notification or open it from Settings).

Setup

1. Install the APK from Releases (sideload; Android 14+).
2. Open Alfred → gear icon → set your n8n webhook URL and auth (Bearer token, Basic, or a custom header — matching n8n's Header Auth credential). Credentials are encrypted at rest via Google Tink with an Android Keystore master key.
3. Test connection to confirm the workflow answers.
4. Set Alfred as your assistant: Settings → Apps → Default apps → Digital assistant app (Samsung: Settings → Advanced features → Side button → Press and hold → Digital assistant app), or via adb:

   adb shell cmd role add-role-holder android.app.role.ASSISTANT com.yshah.alfred

5. In Alfred's Settings, request the battery optimization exemption so background delivery survives aggressive OEM power management.

The webhook contract

Alfred POSTs JSON to your webhook:

{
  "type": "task | note | convo | ping",
  "text": "the transcribed speech",
  "timestamp": "2026-07-01T21:00:00Z",
  "sessionId": "uuid-per-capture"
}

Route on type in your workflow (ping is the Settings test-connection probe). The response is parsed leniently — any of these work:

• A JSON object or single-element array containing a responseText, output, text, response, message, or reply string field (n8n AI Agent nodes commonly produce [{"output": "..."}])
• n8n's streaming NDJSON event format (the final item event's content is used)
• Plain text, or an empty body (any HTTP 2xx counts as delivered)

Convo mode uses a short (~20s) timeout since you're actively waiting for a spoken reply; Task/Note allow up to 5 minutes in the background.

Building

./gradlew :app:assembleDebug     # debug APK
./gradlew :app:assembleRelease   # minified, signed release (needs your own keystore.properties)

Release signing reads keystore.properties (see app/build.gradle.kts); without it the release build is unsigned. See CLAUDE.md for toolchain notes and hard-won platform gotchas (Samsung side-button behavior, broken on-device speech recognition, n8n response-shape quirks).

Known limitations

• Camera mode was prototyped and deferred — n8n received multipart fields as separate binary files; needs the part-encoding issue resolved first.
• Convo multi-turn continuity depends on your workflow maintaining session state (correlate on sessionId); the first turn is solid.
• Speech-to-text uses the system's Google recognizer, which may process audio off-device — the on-device-only recognizer proved unreliable on the S24 Ultra.

License

Personal project, provided as-is with no license granted for redistribution — but feel free to read and learn from it.