What I learned building an open-source kit that turns your AI coding assistant into a DevOps agent
Full disclosure: I work at CloudBees (product marketing). This is not an official CloudBees product. It's a side project I built on my own because I wanted to test a theory: what happens when your AI coding assistant can see your entire delivery stack, not just your code?
Short answer: the conversations change pretty dramatically.
I open-sourced a starter kit that connects Claude Code to CloudBees Unify via MCP and gives the agent 7 skills: pipeline overview, build triage, security scan, release readiness, feature flags, CI health, and Jira ticket filing. Each skill is just a markdown file describing a workflow. The agent reads it, calls the MCP tools, and returns a structured answer. Fork it, swap skills, add your own.
A few things I learned building it:
- The pattern matters more than the tools. MCP + markdown skills + a data plane that normalizes across CI systems — that's the interesting part. You could rewire this to other platforms.
- Read-only by default is non-negotiable. The kit ships with write access off. You have to explicitly opt in to let the agent change anything. A colleague flagged supply chain risks during review, so we also pinned every dependency to a specific version. If you're building something similar, do this from day one.
- Context across tools > context within one tool. When the agent can see across Jenkins, GitHub Actions, and your security scanner at the same time, it can answer questions no single dashboard can. Like "are we ready to release?" across 4 components on 3 different CI systems.
I built the entire demo environment (Jenkins pipelines, repos, dummy data, the kit itself) almost entirely through Claude Code. That was its own learning experience.