Model Salt is a tool that adds externally generated randomness to open ended prompts for AI models. Its model_salt tool can be called before a model begins a task, providing randomized choices that help vary the model’s response across repeated runs. The randomness setting ranges from 0 to 2, with 1 as the default. Model Salt is available for several coding assistants, including Claude Code, Codex, Gemini CLI, and Kimi Code. Setup uses terminal commands to add the tool and, for some assistants, a skill that guides the model on when to call it. Users can check whether it is connected through each assistant’s MCP interface or command. For image and other asset generation in ChatGPT or Gemini, the site also describes a workflow in which users paste a message into a chat; the chat presents randomized choices and asks for confirmation before generating the asset.