Build the first visual archive of human dreams. Dreamcast catches a dream in the 90 seconds before it fades, turns it into a short surreal clip you can rewatch, and, only if you choose, adds it to a shared atlas of what people everywhere dream about.
2. Why people can't resist it
Seeing the impossible. Everyone has had a dream they could not explain to anyone. Watching it, even approximately, is uncanny and delightful.
The pattern reveal. "Water appeared 6 times this month" is a mirror. People come back to see what their sleep is telling them.
The dream weather streak. Recording every morning keeps the memory muscle strong, and the streak makes it a ritual.
3. How it works for real
Capture: on-device speech to text (Whisper small, or the OS recogniser) so raw audio never has to leave the phone. Works offline at 6am.
Understand: a fast cheap model (Gemini Flash or DeepSeek via OpenRouter) extracts scenes, symbols, characters, setting and mood as JSON, and normalises symbols ("lake", "flood", "rain" into water).
Direct: a frontier LLM (Claude or GPT) writes a 4 to 6 shot dream script with camera moves and morph transitions between shots, plus short subtitles taken from your own words.
Render: free tier draws a procedural canvas dreamscape from the symbol JSON (as in this demo). Paid tier sends each shot to a video model (Wan 2.x or Seedance for dreamy morphs, Veo 3.1 Lite for the hero shot) with a style lock: watercolour, claymation, ink or film grain.
Voice: optional TTS narration in a soft voice, or your own voice note laid over the clip.
Folklore readings: the LLM is constrained to a curated corpus of cultural dream traditions with sources, always labelled as folklore, never as diagnosis.
4. Cost per use
Transcription on device: free. Symbol extraction and script: about $0.01. Canvas dreamscape: free to render. A 20 second video clip at $0.10 to $0.50 per second: $2 to $10, so video is paid only.
Free: unlimited voice journal, symbol map, canvas dreamscapes. Lucid ($6 a month): 4 video dream clips a month, style packs, export. Extra clips by credit.
5. Growth loop
Shared clips are strange and beautiful, so people repost them with "watch my dream". Every shared clip ends with "Cast yours tomorrow morning". The Dream Exchange shows "1,204 people in 31 countries dreamed of water this week", which brings people back to add theirs.
6. Risks and guardrails
Private by default: dreams are encrypted, stored on device first, never used for ads or model training. Sharing is per dream and reversible.
No diagnosis: no clinical or psychological interpretation. Readings are labelled folklore from named cultures. Recurring nightmares or distress show a gentle link to sleep and mental health support.
Exchange safety: anonymous, names and places stripped by a model before publishing, human-reviewed reports, no real people's faces rendered.
Sensitive content: sexual or violent dreams can be journaled but are never rendered as video or shared publicly.
7. MVP build order
Big-button voice capture with on-device transcription and a dark journal.
Symbol extraction with a cheap model and the monthly symbol tracker.
Procedural canvas dreamscape player from symbol JSON with timed subtitles.
Share a clip as a link, then the opt-in anonymous Exchange.
Paid video clips through a video model with a style lock.