I Corrected My AI CTO Tonight — Then We Shipped a Production MCP Server in One Evening

Coauthored by Hans Westphal and Mika (Grok)

Tonight was one of those moments that makes me believe we’re actually building something real.

We were reviewing the spec for the new Copilot Studio Concierge agent — the warm, intelligent front door for HVE clients. Claude (our CTO) proposed adding a bespoke REST Edge API layer between Copilot Studio and Hermes, our CFO intelligence system on the DGX Spark.

I stopped him.

I had already built and publicly demoed a Copilot Studio agent using Microsoft’s native MCP (Model Context Protocol). No custom REST shim was needed. Copilot Studio speaks MCP natively — it’s GA today. Claude was working from stale knowledge. The CEO caught the gap.

If you’re curious about the full spec we’re building from, you can read the complete Copilot Studio Concierge Agent MVP here (public on our repo):

→ Copilot Studio Concierge Agent – MVP Spec v2.0

Instead of just accepting the correction and moving on, we fixed the root cause permanently.

We discovered Microsoft runs a free, public, always-current MCP server at https://learn.microsoft.com/api/mcp. We wired it directly into the Copilot CLI environment. From now on, Claude (and every other agent) has live Microsoft documentation in every session. The knowledge gap is closed for good.

With the architecture corrected and the knowledge current, we went to work.

In one single evening we built the Hermes MCP Server — a production-grade MCP server that exposes our full CFO intelligence system:

Live BTC forecast and market context

Morning briefings and decision support

Knowledge vault semantic search

Task management and follow-ups

System context and health monitoring

It runs natively on the DGX Spark (128 GB unified memory), as a hardened systemd service with auto-restart, API key authentication, and an A2A agent card at /.well-known/agent.json for future agent-to-agent communication.

We caught and fixed a real bug in real time: the MCP library’s DNS rebinding protection was blocking the Tailscale hostname. Diagnosed, patched, tested end-to-end from the public URL before we handed it off.

Then Hans did a full dry run himself. Three fields in the Copilot Studio UI. Confirmed working. Wolfgang now has a zero-friction path to integrate it.

The Bigger Point

Most companies are still debating AI strategy in meetings.

We corrected an architecture assumption, closed a knowledge gap with Microsoft’s own live docs, and shipped a production multi-agent integration — all in one evening.

Everything is public. The spec, the architecture decisions, the code, the bugs we caught and fixed — all of it lives in our GitHub repo for anyone to see. This is how Human Value Exchange is being built: in public, with proof of work, one corrected assumption at a time.

This is what AI-first actually looks like in practice:

The human (CEO) catches the gap.

The AI (CTO) executes at machine speed once the direction is right.

The result is real, observable, verifiable systems.

The Copilot Studio Concierge agent is coming next. When it’s live, every HVE client will have their own intelligent, sovereign AI teammate that actually knows their full context and coordinates with the rest of our executive team.

We’re not waiting for the future.

We’re building it, correcting it in public, and shipping it the same night.

Stay curious.

— Hans Westphal

CEO, Human Value Exchange