An internal MCP server — and the docs clients actually use

Problem

AI coding agents and internal tooling had no structured, safe way to query or operate on the platform’s real capabilities. Every integration meant either hand-rolling ad hoc scripts against internal APIs or trusting an agent to guess at behavior from scattered documentation — slow, error-prone, and especially risky for external clients integrating with a production messaging platform.

Solution

I built an internal MCP (Model Context Protocol) server that exposes the platform’s real operations as typed, agent-consumable tools, so AI agents — Claude Code, GitHub Copilot, and in-house tooling — can interact with the system directly instead of improvising against incomplete context. I paired it with official integration documentation built in Astro + Starlight and delivered to clients, so the same clarity that makes the server useful to an AI agent also makes it useful to a human integrator reading the docs for the first time.

Impact

• Became the standard way both internal AI tooling and client integration teams onboard onto the platform’s real capabilities.
• Reinforced a principle that runs through all of this work: adopting AI strategically doesn’t mean chasing every new feature, it means building the actual infrastructure — a real protocol server, real documentation — that makes AI-assisted work reliable instead of improvisational.
• A direct, current example of the “AI-native workflow” the rest of this portfolio’s engineering practice leans on, from the Angular modernization to the vulnerability-mitigation rollout.