Built for the Gemini Live Agent Challenge 2026 · Creative Storyteller Category

SAGA is a cinematic story universe engine where prose, illustrations, narration, ambient score, live voice direction, and persistent world memory flow together in one manuscript. It is designed as an agent, not a chatbot: Gemini Live listens, reasons as a co-author, and can autonomously trigger the next story movement with GENERATING: directions.

Live demo: https://saga-frontend-172547633566.us-central1.run.app
GitHub: https://github.com/Shreyp087/SAGA

Most AI storytelling tools still act like text boxes. SAGA treats story creation as a living system:

• See: inline illustrations with consistent character visual profiles
• Hear: narration and ambient score appear inside the same manuscript flow
• Speak: Gemini Live acts as a voice co-author, not just transcription
• Remember: Firestore + Qdrant keep the story world persistent between visits

Mermaid source lives in docs/architecture/SAGA-Architecture.md.

graph TB
    User["User\n(Browser)"] -->|Voice/Text| Frontend["Next.js Frontend\n(Cloud Run)"]
    Frontend <-->|WebSocket| Backend["FastAPI Backend\n(Cloud Run)"]
    Frontend <-->|WebSocket| LiveProxy["Gemini Live Proxy\n(FastAPI /ws/live)"]

    Backend --> GeminiFlash["Gemini 2.0 Flash\n(Story Engine)"]
    Backend --> Imagen["Imagen 4\n(Illustrations)"]
    Backend --> Veo["Veo 2\n(Cinematic Clips)"]
    Backend --> TTS["Gemini TTS\n(Narration)"]
    Backend --> Lyria["Lyria 2\n(Ambient Music)"]

    LiveProxy --> GeminiLive["Gemini Live API\n(Voice Co-Author)"]

    Backend --> Firestore["Firestore\n(Persistent World)"]
    Backend --> GCS["Cloud Storage\n(Media Files)"]
    Backend --> SecretMgr["Secret Manager\n(API Keys)"]
    Backend --> Qdrant["Qdrant Cloud\n(Vector Memory)"]

    GeminiFlash -.->|ADK Orchestration| ADK["Google ADK\n(Agent Framework)"]

• Interleaved manuscript: text, images, audio, video, and score appear in one stream instead of separate tabs.
• Gemini Live as an agent: the co-author asks clarifying questions, then autonomously triggers new story generation.
• Persistent world return: close the browser, return later, and SAGA welcomes you back with story-specific characters and locations.
• Character Visual Bible: generated character profiles are injected into every illustration prompt for visual consistency.
• 3D world globe: locations and story connections appear in a live world atlas as the manuscript evolves.
• Story Bible export: a formatted PDF chronicle with manuscript sections, character archive pages, and story metadata.

• Python 3.12+
• Node.js 22+
• Docker Desktop + Docker Compose
• Google Cloud SDK (gcloud)
• Gemini API key from https://aistudio.google.com
• Optional: Qdrant Cloud URL + API key for remote vector memory

git clone https://github.com/Shreyp087/SAGA.git
cd SAGA
make setup

cp backend/.env.example backend/.env
cp frontend/.env.local.example frontend/.env.local

Fill in at least:

• backend/.env: GOOGLE_API_KEY, GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_REGION, GCS_BUCKET_NAME, QDRANT_URL
• frontend/.env.local: NEXT_PUBLIC_BACKEND_URL, NEXT_PUBLIC_WS_URL

make quickstart

Expected local URLs:

• Frontend: http://localhost:3000
• Backend API: http://localhost:8000
• Backend health: http://localhost:8000/health/
• Qdrant (local compose): http://localhost:6333

Landing page -> Begin Your Saga -> /story
            -> WebSocket connects
            -> text streams
            -> inline image appears
            -> narration/score attach
            -> world globe updates

gcloud auth login
gcloud config set project YOUR_PROJECT_ID
./scripts/deploy.sh YOUR_PROJECT_ID us-central1

That script:

1. Enables required Google Cloud APIs
2. Syncs secrets into Secret Manager
3. Builds and pushes backend/frontend images to Artifact Registry
4. Deploys saga-backend to Cloud Run
5. Deploys saga-frontend to Cloud Run using the discovered backend URL

cd infrastructure/terraform
terraform init
terraform apply -auto-approve \
  -var="project_id=YOUR_PROJECT_ID" \
  -var="gemini_api_key=YOUR_GEMINI_KEY" \
  -var="qdrant_url=https://YOUR-QDRANT-CLUSTER"

Relevant files:

• scripts/deploy.sh
• infrastructure/terraform/main.tf
• infrastructure/terraform/variables.tf
• infrastructure/terraform/outputs.tf

These are planning estimates for hackathon-style usage, not billing guarantees.

• Demo video placeholder: DEMO_URL
• Full script: docs/DEMO-VIDEO-SCRIPT.md

• backend/app/agents/story_agent.py: interleaved multimodal story orchestration
• backend/app/agents/saga_adk_agent.py: explicit ADK agent surface and tool registration
• backend/app/api/live.py: Gemini Live proxy and GENERATING: trigger path
• backend/app/api/websocket.py: story streaming, resume, and media restoration
• backend/app/services/firestore_service.py: persistent worlds and cinematic welcome-back generation
• frontend/src/components/story/StoryCanvas.tsx: inline multimodal manuscript renderer
• frontend/src/components/story/WorldMap.tsx: live 3D world atlas

SAGA was built specifically for the Gemini Live Agent Challenge 2026 in the Creative Storyteller category. The repo includes:

• ADK agent definition
• Cloud Run deployment automation
• Terraform IaC
• architecture diagram
• demo script
• Devpost submission draft
• blog post draft
• judge-facing deployment proof

• GitHub repository: https://github.com/Shreyp087/SAGA
• Demo video: https://youtu.be/mdONC55NxEU
• Blog Link: https://dev.to/shreyp087/how-i-built-saga-a-living-multimodal-story-engine-with-5-google-ai-models-102b

#GeminiLiveAgentChallenge