A hands-on walkthrough of building a read-only Kubernetes debugging AI agent using Claude Code, packaged as a Claude Code Plugin within a custom Marketplace. The agent integrates VictoriaMetrics, VictoriaLogs, and Alertmanager via Skills (rather than MCP) to diagnose Pod failures like CrashLoopBackOff, OOMKilled, and ImagePullBackOff. The post covers the full project structure: marketplace.json, plugin.json, agent frontmatter with allowed/deny-tools enforcing least privilege, a reusable SKILL.md with MetricsQL/LogsQL query templates, and a cross-platform bootstrap script for environment setup.
Table of contents
ContentsOverall structure and planMCP vs Skills: why skillsCreating the MarketplaceMarketplace: the .claude-plugin/marketplace.json filePlugin: the plugins/k8s-tools/.claude-plugin/plugin.json fileThe .claude/settings.json fileThe .gitignore fileCreating the AI AgentThe plugins/k8s-tools/agents/k8s-pod-debugger.md fileFrontmatter: Name and Description fieldsFrontmatter: Permissions and Tools fieldsGet Arseny Zinchenko (setevoy) ’s stories in your inboxBody: the Agent’s System PromptThe full k8s-pod-debugger.md fileSkill: k8s-troubleshooting-flowWhy a Skill, separate from the System Prompt?The plugins/k8s-tools/skills/k8s-troubleshooting-flow/SKILL.md fileCreating the Bootstrap scriptVerifying the Marketplace and PluginThe CLAUDE.md fileThe README.md fileVerifying extraKnownMarketplaces508 Impressions