Experiment opening

Custom MCP Server

Test the complete path from a small Python MCP server to tools that an agent can discover and call inside Copilot Studio.

Status: Field test
Focus: Tool discovery & orchestration
Environment: FastMCP · Copilot Studio

Hypothesis

Can a useful tool surface stay small and legible?

The experiment began with a narrow question: could Copilot Studio discover and use a custom set of tools without wrapping every capability in a separate conversational action?

The test deliberately kept the domain simple so the protocol, discovery behaviour, and governance implications stayed visible.

Build

Three tools, one explicit contract.

A Python server built with FastMCP exposed three fictional project-management tools: project status, owner assignment, and open-risk listing.

The server used the current Streamable HTTP transport at the /mcp endpoint. The older SSE transport was deliberately avoided.

Connection

Make the local experiment reachable, then observable.

A temporary Cloudflare tunnel made the local server reachable for the proof of concept. Copilot Studio then registered the MCP endpoint through a custom connector.

The connector required a connection even with authentication set to None for the experiment. Tool discovery was validated in a new agent session, because an existing conversation does not automatically refresh its tool inventory.

System flow

A small path from intent to tool result.

The experiment keeps every handoff explicit, making discovery and execution easier to inspect.

01
Intent
A user asks for a project operation in natural language.
- Copilot Studio agent
- Conversation context
02
Discovery
The agent selects a tool from the MCP contract.
- Custom connector
- Tool schema
03
Execution
FastMCP receives the call and runs the selected Python function.
- Streamable HTTP
- Fictional data
04
Response
The structured result returns to the agent for a contextual answer.
- Tool result
- Agent response

Key decisions

Keep the protocol test honest.

Use Streamable HTTP

Expose the server through the current MCP transport at a single endpoint.

It keeps the experiment aligned with the current protocol direction instead of building around deprecated SSE behaviour.

Keep data fictional

Test discovery and orchestration without connecting a real business system.

That separates protocol learning from security, privacy, and data-integration work that belongs in a production design.

Treat the connector as governed

Observe how the custom connector appears inside the Power Platform control plane.

MCP does not bypass platform governance; the connector remains visible and classifiable through data-loss-prevention policy.

Observations

What the field test demonstrated.

01
Copilot Studio discovered the custom tool set and could invoke the expected function from a natural-language request.
02
A new agent session was necessary before an updated tool inventory became visible.
03
The custom connector remained part of the existing Power Platform governance surface.

Production path

What must change before real data arrives.

01
Replace the temporary tunnel with a stable, monitored host.
02
Add OAuth 2.0 and validate authorization at the tool boundary.
03
Define connector allowlisting, data-loss-prevention classification, and operational ownership before real data is introduced.