Claude Mods: Flipping the Harness Inside Out

Dan McAteer argues Claude Mods turn Claude Code's harness into a programmable Attention Interface that users shape themselves. Mods are TypeScript or JavaScript plugins that hook into events like prompts, tool calls and interface renders, and Anthropic already ships six built-in ones, including the off-by-default You Should Know side agent. He contrasts this with OpenAI's ChatGPT Intelligent UI, where the model picks the interface, and Cursor's PR Routing & Approval, where teams write policy inside a fixed flow, and flags mod downsides: noise, roughly 5–6k input-equivalent tokens per supervisor check, and no sandboxing.