AI coding agents build UI code with no built-in visual verification loop

When Anthropic or OpenAI ships a model update, engineering teams have no way to know if their AI coding agent still follows the same instructions and produces the same UI output it did before. Developers discover regressions only after burning hours on broken outputs or catching hallucinated 'task complete' claims post-merge. This CLI captures a baseline of agent behavior (instruction-following plus visual UI snapshots) and flags drift automatically whenever the underlying model changes.

AI coding agents like Claude Code routinely break their own prior behavior after model updates or prompt changes, and developers only discover the regression when the generated UI looks wrong at runtime. This MCP server records a behavioral baseline of agent-produced UI state on the first passing run, then on every subsequent run diffs both the DOM output and a visual screenshot snapshot against that baseline, surfacing what changed before the developer wastes a debugging session.