Modelwatch

Brings ModelWatch's continuous drift monitoring into Claude Desktop so you can define behavioral specs, check drift scores, and pull alert history without leaving your editor. You can create new specs with input prompts and expectation rules (refusal checks, format constraints, semantic similarity thresholds), query recent drift reports across your endpoints, and see which scheduled checks fired alerts. Useful when you're debugging a silent model update from OpenAI or Anthropic and want to cross reference what changed in your production behavior baselines. The server talks to ModelWatch's API using your mw_ key, so you need an active account. If you're already running hourly checks on your LLM outputs and want Claude to help triage when drift spikes, this connects the dots.