Playground

Everything below uses browser-local Generation: no Gateway, API route, server action, API key, or inference endpoint. Choose Chrome's browser-managed Gemini Nano when available, or explicitly activate the Gemma fallback. There are no automatic downloads and no silent runtime switching.

Try it

• Send Draft a support reply for Ana about a delayed order (ana@acme.com). in either chat.
• Use Protect these Detection categories to choose which model-backed categories are redacted. Email, phone, URL, and other deterministic detectors remain always-on.
• The inspector shows the exact protected history and current content delivered to the browser runtime, never the private mapping or original input.
• The response is restored locally, so the UI shows the original email.
• When Gemma is cached, Clear cached model removes its browser cache; it does not affect Chrome's separately managed Gemini Nano.
• New chat waits for stop and runtime cleanup, then clears that tab's history, privacy session, error, and inspection before creating a new session.

Each force-mounted tab owns a separate private conversation, privacy session, history, observer, and mapping. A generation gate allows one chat at a time: the owning chat can stop; the other chat's submit and reset controls remain disabled until cleanup releases the gate. Both examples use their real public Vercel AI SDK and TanStack adapters. The docs-internal browser runtime is not a public local-pii API.

Runtime disclosure

• Gemini Nano is browser-managed and requires compatible desktop Chrome, hardware, and Prompt API availability. Browser-managed resources are not an artifact download controlled by this page.
• Gemma is onnx-community/gemma-3-270m-it-ONNX, revision 2dbbfdb1b59bd034eb959428c6a7da9dd7ea27f0, q4f16 WebGPU. Its six pinned model files total 293,284,073 bytes. The first activation can also fetch 23,614,439 uncompressed bytes of versioned ONNX Runtime Web support files, for approximately 316,898,512 static artifact bytes (~317 MB before transfer compression). Origins are https://huggingface.co, https://*.cdn.hf.co, and https://cdn.jsdelivr.net; review the model card and Gemma license terms before activation.
• Explicit artifact downloads contain only static model and runtime resources, with no user content. Prompts, restored values, private mappings, and inspection state do not go to an application endpoint.
• Small browser-local Generation models are useful for this demonstration, but quality and reasoning vary by model and language. Translation does not imply equal quality in English, Portuguese, and German.