A Gemini-native chat UI that keeps your data in the browser

What it does

AMC WebUI is a React 18 single-page app that talks to Google’s Gemini API first, OpenAI-compatible endpoints second. Chat history lives in IndexedDB by default; a Dockerized backend exists mainly for key-proxying and CORS relief. The UI covers the usual suspects—multimodal chat, image generation, voice input/output, screen sharing, and real-time streaming—plus a few less usual ones.

The interesting bit

The project ships a browser-based Python runtime via Pyodide, preloaded with numpy and pandas, which auto-installs scipy or scikit-learn on demand. That’s not a party trick; it means code execution happens without a server round-trip. The “Live Artifacts” feature auto-detects code blocks and renders them as interactive HTML previews, including Mermaid and Graphviz diagrams. The README also notes a strict pricing philosophy: cost estimates only appear when the stored metadata can “strictly and precisely” reconstruct official billing—otherwise you get an em-dash.

Key highlights

• Dual API paths: Gemini native mode gets the full feature set; OpenAI-compatible mode is deliberately limited to standard chat completions with isolated keys and model lists.
• Local-first by default: IndexedDB storage, BroadcastChannel tab sync, Web Locks API for write safety, and a full PWA shell with offline app opening.
• File handling with opinions: Browser-side audio compression, ZIP/folder drop parsing, and per-file-type toggles between Gemini Files API and base64 upload.
• Safety controls exposed: Five Google safety filter categories, each with five severity levels—exposed in the UI, not hidden behind defaults.
• Node 26 lock-in: .nvmrc, CI, and Docker all enforce Node 26; engine-strict will reject other versions at install time.

Caveats

• The Docker deployment is explicitly labeled “trusted self-hosted” only; the README warns it lacks auth, rate limiting, and tenant isolation for public exposure.
• OpenAI-compatible mode does not inherit Gemini’s tool chain—deep search, code execution, and image generation stay gated behind Gemini native mode.
• The README truncates mid-sentence during the MCP stdio section, leaving that feature’s scope unclear.

Verdict

Worth a look if you want a Gemini-centric UI with genuine local-first discipline and don’t mind the Node 26 constraint. Skip it if you need a hardened multi-user deployment out of the box, or if your workflow depends on OpenAI-compatible tool calling.