Claude Code plugin evals: real model calls, trust warnings, and CI gating explained

ClaudeDevs · x · 2026-09-12

Anthropic's docs detail how claude plugin eval works: every run and judge grader is a real model call counted against your plan or API bill, so pilot with --runs 1 first; plugins' hooks and MCP servers run as you, so only evaluate plugins you trust. Each case combines a realistic prompt with graders (regex, tool-call checks, or model-judged rubrics). Use evals to measure how reliably your plugin steers Claude, catch regressions, compare against a no-plugin baseline, and gate CI on scores. Try it with claude update.

Related event: Claude Code Launches Plugin Eval Tool to Quantify Plugin Value(4 posts)→

Original post →

More from coding & agent

coding & agent channel →