Description

Ask from your Pebble, keep the conversation going on Android, and choose the model, search tools and speaking voice independently. The watch shows a compact reply; the phone keeps the full answer, formatting and source links.

Requires Android 8 or later, the Dick Tracy Watch companion, and your usual paired Pebble phone app. Get the companion and setup details: https://dr.eamer.dev/dicktracy/

ON THE WATCH
Press Select to dictate a question or follow-up. It sends when Pebble dictation ends; choose Send: Review first to check the transcript before sending. Up/Down scroll, Back cancels, and holding Select marks the full reply to open on your phone. The menu includes New conversation, larger text, Tools On/Off, Watch voice, Voice test and Auto-read. Sending mode and text size persist. The backlight stays on while the app has focus.

ANDROID CHAT AND HISTORY
Type questions without a watch. Read selectable Markdown answers with clickable citations, a Sources list and a label showing which tools ran. Reopen saved replies to read or listen, continue an earlier conversation, start a new chat, or edit a failed question and retry. Up to 200 question/reply records are saved; Delete all clears history while keeping your provider keys and settings. The phone screen stays awake while the app is visible.

ANSWER SOURCES
OpenAI, Anthropic, xAI, Gemini, OpenRouter, Runware, Hugging Face, Mistral, Cohere, Ollama Cloud and Kimi, plus a custom HTTPS endpoint compatible with Chat Completions. Enter a model ID or choose a preset. Each provider keeps its own key, model and tool selection. Anthropic also has a workspace-ID field. Check saved setup sends a short test with tools off, without adding it to History.

ON-DEVICE MODELS
Gemini Nano uses Android AICore on supported phones and requires the companion to stay on screen. Gemma offers a resumable 2.59 GB Gemma 4 E2B download with size/SHA-256 checks, or import of a portable GPU/CPU .litertlm model. Downloads default to unmetered networks; local model files can be removed separately from chats. Models need compatible hardware and enough storage/RAM. Local inference never silently falls back to a hosted model. Optional search and hosted speech still send data to their selected services.

SEARCH AND TOOLS
Configure these in Android Settings → Tools, then toggle the saved selection from the watch. Tools start off. Availability depends on the chosen model and account.
• OpenAI: web search/page reading and hosted calculations.
• Anthropic: web search, link reading and hosted calculations.
• xAI: web search/page reading, X search and hosted calculations.
• Gemini: Google Search, URL context, hosted calculations and places/maps on supported models.
• OpenRouter: web search, link reading and current time.
• Mistral: web search and hosted calculations.
• Ollama Cloud and Kimi: standalone search and link fetching.

Any answer source, including local models, can instead use OpenAI, Ollama Cloud or Kimi for search. That service retrieves evidence; your selected model writes the answer. The watch shows a source count and the phone opens the links. Search budget offers 1–3 calls where supported; it is not a spending cap. Calculations run at the provider, and maps use a place you name rather than phone location.

VOICES AND PLAYBACK
Use any answer provider with ElevenLabs, OpenAI, Runware, Mistral or installed Android offline voices. ElevenLabs includes voice presets and your own account's voice ID. Listening settings have Slower/Normal/Faster speeds, Test voice and a separate saved speech selection. Tap Listen beside any saved answer for phone playback.

Experimental watch speech uses the phone to synthesize the full answer and sends buffered audio sections to compatible Pebble Time 2 or Pebble 2 Duo firmware. No voice model is installed on the watch. Auto-read starts off; enable it on phone or watch to speak new watch-requested replies automatically. Phone questions do not trigger watch playback. Select or Back stops audio, and Listen moves playback to the phone. Watch mute is respected; pauses and limited speaker quality remain possible.

PRIVACY, COSTS AND LIMITS
Keys and conversations use AES-256-GCM with Android Keystore and are excluded from backup. Hosted answers receive the question and up to ten recent exchanges from the same conversation/provider/model. Hosted speech receives the reading text. Pebble's own dictation service handles speech recognition and when recording ends.

The watch's text reply is limited to 900 UTF-8 bytes; the full answer remains on Android and is used for speech. Hosted requests use your provider accounts and may cost money; an app subscription may not include API credit. These details cover watch 1.8.0 with Android 0.9.0-dev (build 11).

By Luke Steuber — https://lukesteuber.com