Microsoft’s 1997 paperclip now runs local LLMs

What it does

Clippy is an Electron desktop app that runs large language models locally via llama.cpp and the GGUF format. It bundles one-click installers for a handful of popular models—Gemma3, Llama 3.2, Phi-4, and Qwen3—and presents them through a chat interface deliberately styled after Microsoft Office 97. Everything stays on your machine; the only network traffic is an optional update check, and the app automatically negotiates with Metal, CUDA, or Vulkan to find the fastest local backend.

The interesting bit

Beneath the retro skin, the app serves as a reference implementation for @electron/llm, showing how to ship local inference in an Electron app without making the user touch a terminal. The actual payload is the collision of 1997 UI kits with 2025 model weights, which the author openly labels software art—or satire, if you find the aesthetic grating.

Key highlights

• Runs GGUF models fully offline after initial setup; the only network call is an optional update check.
• Auto-detects the best local inference backend (Metal, CUDA, Vulkan) via node-llama-cpp.
• Bundles one-click installers for Gemma3, Llama 3.2, Phi-4, and Qwen3.
• Supports custom models, prompts, and parameters for users who bring their own weights.
• Explicitly disclaims any Microsoft affiliation, existing purely as an artistic homage to Kevan Atteberry’s original character.

Caveats

• The author openly warns that many other local-LLM chat apps are likely better, so this is not a project chasing feature parity.
• It is unapologetically software art, meaning the retro interface is non-negotiable regardless of your aesthetic tolerance.

Verdict

Grab this if you are building Electron apps with local LLMs and need a working reference for @electron/llm, or if you genuinely miss the Windows 98 aesthetic. If you want a polished, modern chatbot, the author explicitly suggests you look elsewhere.