Automous Assistant
PREVIEW · v0.14
Ask anything.
Show me anything.
One AI assistant for questions, camera vision, and translation.
Assistant is off
Tap Live to start and hear a greeting and tell the assistant what you need. Realtime voice stays connected while the session is active. English is the default; ask to change languages.
Audio output
Speaker mode controls app audio, not the phone’s earpiece/loudspeaker routing. Adjust device media volume for more loudness. Say “speaker on”, “mute”, “speak louder”, or “volume 50 percent”.
Automous Assistant · Vision
Camera ready for questions · tap to request focus
Opening camera…
Point at text and tap Read / Translate Camera. The current frame is shared with the Live assistant when connected; no photo is saved.
When Live is connected, the same assistant voice will read and translate what the camera sees. You can ask follow-up questions in Live.
Changelog
Version history for this demo. Expand a release to see what changed.
v0.14 · Temporary visual memory and quieter camera status
- Keep up to 12 compressed camera frames in browser RAM for the current Live session, then clear them when Live ends.
- Take a new frame for each spoken request while Camera and Live are active, retaining earlier frames as conversation context.
- Hide detailed realtime and vision errors from the screen; log technical diagnostics to the browser console.
- Teach the assistant to pronounce Automous as Auh-toe-mus.
- Images are not saved to the phone gallery; OpenAI data retention is governed separately.
v0.8 · Automatic camera context for every spoken turn
- When Camera and Live are both active, each completed spoken request automatically attaches a fresh in-memory camera frame before the AI response is requested.
- Removed keyword-only camera triggering, which missed requests such as “What does it say?”
- AI is explicitly told the camera image has already been supplied; it should not attempt to take a screenshot.
- Camera image status and errors are shown in the camera view; no photo is saved to the gallery.
- Kept quiet listening, English default, unified voice, and manual camera action as a fallback.
- Vision still depends on realtime transcription events, mobile permissions, and network access; this version needs testing on a phone.
v0.7 · Voice-triggered vision and quiet listening
- When Live and Camera are active, each spoken request is processed with a fresh temporary camera frame for visual context.
- Turn detection no longer generates a response before the camera frame is attached; only one response is requested per completed spoken turn.
- Added quiet listening instructions: no unsolicited check-ins, no response to silence, and no invented unreadable text.
- Improved camera readiness feedback and preserved the single Marin voice and English default.
- Image captures are temporary and not saved to the gallery; camera permissions and browser focus limitations remain.
v0.6 · Automous Assistant and full-screen vision
- Rebranded Automous Translator to Automous Assistant in the UI, assistant prompt, and install metadata.
- Camera tab opens a full-screen live preview in one tap, with an immediate permission request if needed.
- Added camera switching, higher-resolution requests, continuous autofocus where available, and best-effort tap-to-focus.
- Added temporary in-memory camera frames for visual questions and a one-tap Ask about view control.
- Added voice-triggered frame sharing during active Live sessions where input transcription is available.
- Kept unified Marin voice, speaker controls, English defaults, and prior release history.
- Camera focus controls vary by browser/device; frames are transmitted to the AI service but not saved to the gallery.
v0.5 · Simplified Live and unified camera voice
- Removed redundant Talk navigation; Live is the only voice-session entry point.
- Camera frames can be sent into the active Live session for spoken reading and translation in the same assistant voice.
- Aligned fallback speech voice with the Live voice (Marin) where supported.
- Added clear camera and Live connection feedback; camera requests require a deliberate tap.
- Camera access remains permission-based; no image is saved by this app.
v0.4 · Realtime English-first Live
- Replaced four-second audio chunks with WebRTC real-time speech using gpt-realtime-2.1.
- Live begins with an English greeting and asks what you want to do.
- Assistant and translation languages are separate; translation requires an explicit instruction.
- Removed the implicit Spanish target; English is the default.
- Added a single voice playback stream to prevent overlapping generated responses.
- Kept Speaker toggle, volume slider, camera translation, and version history.
- Note: phone speaker routing remains controlled by the operating system.
v0.3 · Speaker controls and release history
- Added Speaker on/off button for AI voice playback.
- Added adjustable in-app playback volume, remembered on this device.
- Added spoken commands for mute, unmute, louder, quieter, and volume percentage.
- Added this changelog for tracking releases.
- Documented mobile browser speaker-routing limitations.
v0.2 · Automous Assistant
- Renamed Verba to Automous Assistant.
- Introduced conversational Talk assistant and short-turn microphone sessions.
- Added voice-command routing and interpreter mode.
- Fixed mirrored camera preview and added Translate Live without saving a photo.
- Removed app access-code requirement.
v0.1 · Initial prototype
- First Cloudflare Pages demo with OpenAI-backed translation, speech, camera, and explanations.
AI can make mistakes. Verify critical details. Microphone and camera require your permission. Camera frames and recorded speech are sent for processing only when needed. Realtime audio is sent to OpenAI during Live sessions. This demo does not persist conversation history on its own server. Live automatically stops after 15 minutes. No access code: anyone with this URL may use your OpenAI credits. Keep it private and monitor spending. Browser restrictions may limit hands-free behavior, especially with the screen locked.