AI coding agents can call tools and generate code quickly, but I argue that speed without standards awareness creates new failure modes. In this session, I'll show how an MCP can act as a governance surface, where policies, constraints, and grounded engineering guidance are introduced before code implementation begins. Using PolicyNIM, an open-source preflight layer, we'll give coding agents grounded, citeable engineering guidance.

A small suite of Markdown files becomes structured, AI code generation guidance by retrieving relevant policy evidence, reranking results, synthesizing suggestions with citations, and failing closed when the evidence is too weak to trust. We'll walk through two MCP tools and the CLI, transport choices, and auth boundaries.

This talk is about practical lessons from integrating the MCP with AI coding agents like Codex and Claude Code, along with an evaluation workflow for comparing retrieval quality and measuring impact.

You'll leave with a concrete design pattern for building MCP servers that expose evidence-backed constraints, guidance, and safer defaults for agentic software workflows.