Reading Bukowski's *Ham on Rye* again, I kept thinking about systems that fail you. Our enterprise platform had the same problem: documentation was an archaeological dig in Confluence and READMEs. LLM agents promised autonomy but required constant context pasting. The solution was building an MCP ecosystem to give agents structured, dynamic access to our tools.

The core is Model Context Protocol from Anthropic. Instead of overwhelming prompts, it exposes a discovery layer. Tools, resources, and prompts are described in a schema the LLM queries dynamically. The challenge was generating accurate manifests from our messy, undocumented codebase. I built `@mxconsulting/mcp-gen` to parse JSDoc annotations, handling multiple function declaration patterns and validating parameters to prevent runtime errors.

This wasn't just a code problem; it was an infrastructure one. On Kubernetes, manifests live as ConfigMaps. A mutating admission webhook injects them into pods, calculating checksums as pod annotations. A custom recontroller then watches for ConfigMap changes, updating only affected deployments to trigger rollouts. The MCP server loads these manifests, exposing everything over HTTP for agents to discover and use autonomously.