Mikupad

Mikupad is a browser-based LLM frontend for writers and anyone who wants to edit and guide generated text. It connects to a separate model backend and keeps writing context attached to your prompts.

• Text continuation and chat: Continue a draft directly or chat with instruct models, using templates to handle message boundaries and compatible backends' Chat Completions API.
• Persistent writing context: Memory holds background text, Author's Note places guidance near the end of the context, and World Info adds material when matching keywords appear.
• Generation editing: Undo and redo generated text, inspect token probabilities on supported backends, and regenerate from a selected token.
• Token control: Logit bias lets you adjust how likely the model is to produce particular tokens.
• Saved sessions: Prompts save and restore automatically, while session import and export let you keep backups or share work.

The interface runs in a browser as a single HTML file. It supports llama.cpp, koboldcpp, AI Horde and OpenAI-compatible APIs, so your choice of backend determines where generation runs. An optional local Node.js server provides remote or local-network access to the session database. Mikupad is open source under AGPL-3.0.

Token probability inspection depends on the backend: koboldcpp requires token streaming to be disabled, and oobabooga requires an _HF sampler.