What breaks when you ship 21 AI tools that never touch a server

The developers behind FreeToolHub share their experience building 21 AI-powered tools that run entirely in the browser using WebML. By leveraging @huggingface/transformers and onnxruntime-web, the team avoided backend inference costs and prioritized user privacy. However, the transition to serverless AI presented significant technical hurdles. Key challenges included configuring Cross-Origin-Embedder-Policy headers to enable SharedArrayBuffer without breaking third-party dependencies, managing large model downloads for mobile users, and resolving version conflicts between different ONNX runtime packages. The team also learned that browser-based AI is currently unsuitable for tasks requiring rigid structured output, such as JSON generation. Ultimately, the project demonstrates that while shipping client-side AI eliminates infrastructure costs and enhances privacy, it requires rigorous attention to quantization, caching strategies, and asset management to ensure a reliable user experience across diverse hardware.