Mikupad

Mikupad is a browser-based LLM frontend packaged as a single HTML file, for writers and people who want control over generated text. It connects to a model backend of your choice and supports both direct text continuation and chat with instruct models. It's open source under AGPL-3.0.

The interface runs locally in your browser. Its compiled edition can work fully offline with a local backend. Supported connections include llama.cpp, koboldcpp, AI Horde and OpenAI-compatible APIs; using a remote service sends generation requests to that service. An optional local Node.js server provides access to the session database across your network or remotely. A hosted browser interface is also available through GitHub Pages.

For longer writing projects, persistent context lets you keep background material attached to a prompt. Memory adds text at the beginning, while an Author's Note places guidance near the end at an adjustable depth. World Info inserts relevant material when keywords appear. Mikupad automatically saves and restores prompts, and session import and export support backups or sharing.

You can undo and redo generated continuations, inspect token probabilities where the backend supports them, and regenerate text from a chosen token. Logit bias lets you influence which tokens the model is likely to produce. Chat templates handle message boundaries for instruct models and support the Chat Completions API on compatible backends.