SkyPhusion: Open-source Solution for a Full-featured Multimodal AI Playground Based on Cloudflare Worker

SkyPhusion is an open-source project of a full-featured multimodal AI playground deployed on a single Cloudflare Worker. It supports voice conversations with 35 chat models, image/video/music generation, RAG retrieval, project management, and web search, demonstrating a new paradigm for building complex AI applications on edge computing platforms. The project is maintained by SkyPhusion, open-sourced on GitHub, and licensed under AGPL v3.

The original author/maintainer is SkyPhusion. The project is open-sourced on GitHub (repository link: https://github.com/SkyPhusion/skyphusion-llm-public) and was released on 2026-06-03. SkyPhusion is a feature-rich multimodal AI playground fully deployed on a single Cloudflare Worker. It enables full-stack AI functions such as chat, voice interaction, multimodal generation, and RAG retrieval without the need for traditional server architecture. Its "all-in-one" architecture leverages Cloudflare's edge AI infrastructure to deliver low-latency and highly available services.

1. Multi-model chat: Supports 35 models from 5 providers (including Workers AI, Anthropic Claude, xAI Grok, etc.), all with streaming output;
2. Voice conversation: Submit-free interaction (Deepgram Flux real-time STT + Aura-2 TTS);
3. Multimodal generation: Images (FLUX 2 series, etc.), videos (Google Veo 3.1, etc.), music (MiniMax Music 2.6);
4. RAG retrieval: File upload (PDF/Excel, etc.), vector embedding (BGE-base), Vectorize storage;
5. Project management: Named projects to organize documents and conversations;
6. Web search: Parallel queries with Tavily + Wikipedia.

1. Unified interface: Drives all modalities via env.AI.run() binding;
2. Scheduling assistant: Adapts to provider APIs like Anthropic Claude and xAI Grok;
3. Streaming transmission: Supports SSE streaming output from 5 providers;
4. AI Gateway: Implements observability, caching, and rate limiting;
5. Storage architecture: D1 (metadata/conversations), R2 (binary products), Vectorize (vector embeddings);
6. Long task processing: Cloudflare Workflows;
7. Security control: Cloudflare Access isolates user data;
8. Client optimization: Video keyframe extraction to reduce bandwidth costs.

Adopts a focus mode layout: single-column centered conversations + floating input box; sliding sidebar (history/projects/document search); searchable model selector; top bar with settings pop-up and account menu; supports attachment upload and voice microphone; capability-aware mode switching (only displays applicable attachment types).

Deployment steps:

1. Clone the repository and configure environment variables;
2. Deploy to Cloudflare using Wrangler;
3. Configure Cloudflare Access authentication;
4. Add API keys (OpenAI/xAI/Tavily, etc.). The project uses AGPL v3 license and encourages community contributions and secondary development.

SkyPhusion demonstrates new possibilities for edge AI:

1. Lowered threshold (running complex AI applications on a single Worker);
2. Multi-model strategy (comparing 35 models to select the best);
3. Cost optimization (unified billing via Cloudflare);
4. Privacy protection (edge processing reduces third-party transmission);
5. Rapid prototyping (complete functional reference implementation). It is an excellent learning case and starting point for multimodal AI application developers.