MCP tool servers¶

How to provision the MCP server that M5 · MCP Tools connects to — deploy it, read its URL + key, and (optionally) register it in the Foundry tool catalog so agents reference it by connection id.

M5 · MCP Tools connects an agent to a remote Model Context Protocol (MCP) server. The lab itself assumes that server already exists — it only reads MCP_SERVER_URL / MCP_SERVER_LABEL from .env and wires them to an agent. This page covers the missing half: how to stand the server up.

You have two practical options:

Option	Effort	When to use
A · Deploy the reference Azure Functions MCP server	~10 min, needs Azure	You want the exact 37-tool "project tracker" used by the reference lab.
B · Run any MCP server you already have	minimal	You have another MCP endpoint (hosted or local) and just want M5 to call it.

Just reading along? You can skip this.

M5 prints its Expected output in prose, so you can read the whole lab without a server. Provision one only when you want to run the cells.

Option A — Deploy the reference Azure Functions MCP server¶

The reference series (08-agents/08-05-contoso-pmo-mcp) ships a multi-tool MCP server built on the Azure Functions MCP extension. Azure Functions exposes MCP tools over a built-in SSE endpoint at /runtime/webhooks/mcp/sse, protected by a system key — which is exactly the URL shape M5 expects (https://<host>/runtime/webhooks/mcp/sse?code=<key>).

1. Prerequisites¶

The toolchain from Setup (az login, an Azure subscription).
Azure Functions Core Tools v4 — install.
The reference server source. Clone the upstream reference repo and change into the Contoso PMO MCP sample:

git clone https://github.com/microsoft/foundry-samples.git
cd foundry-samples/08-agents/08-05-contoso-pmo-mcp

(Folder names in the upstream repo move occasionally — if the path differs, search the repo for the 08-05-contoso-pmo-mcp sample.)

2. Create the Function App and deploy¶

Create a Flex Consumption Function App (and its storage account), then publish the code:

RG=foundry-workshop-rg
LOCATION=eastus
STORAGE=fwmcp$RANDOM
APP=foundry-workshop-mcp-$RANDOM

az group create -n $RG -l $LOCATION
az storage account create -n $STORAGE -g $RG -l $LOCATION --sku Standard_LRS
az functionapp create -n $APP -g $RG \
  --storage-account $STORAGE \
  --flexconsumption-location $LOCATION \
  --runtime python --runtime-version 3.12 --functions-version 4

# from the sample directory:
func azure functionapp publish $APP

Prefer azd?

If the sample includes an azure.yaml, azd up provisions and deploys in one step. Either path leaves you with a running Function App.

3. Read the server URL + key¶

The MCP SSE endpoint is guarded by the mcp_extension system key. Fetch it and assemble the full URL:

HOST=$(az functionapp show -n $APP -g $RG --query defaultHostName -o tsv)
KEY=$(az functionapp keys list -n $APP -g $RG \
  --query "systemKeys.mcp_extension" -o tsv)

echo "MCP_SERVER_URL=https://$HOST/runtime/webhooks/mcp/sse?code=$KEY"

Treat the key like a password

The ?code=<key> query string is the credential for the server. Keep it in .env (git-ignored) — never paste it into a notebook cell or commit it.

4. Set the M5 variables¶

Add these to your .env (see Setup → Configure):

MCP_SERVER_URL=https://<host>/runtime/webhooks/mcp/sse?code=<mcp_extension-key>
MCP_SERVER_LABEL=project_tracker

MCP_SERVER_LABEL is just a short name that namespaces the server's tools inside an agent, so multiple MCP servers can coexist. M5 §1–§5 needs only these two values.

Option B — Use any MCP server you already have¶

M5 is server-agnostic: any reachable MCP endpoint works. Point the same two variables at it —

MCP_SERVER_URL=https://<your-host>/<your-mcp-endpoint>     # include any ?code= / token
MCP_SERVER_LABEL=project_tracker

— and re-run the lab. The exact tool names in the output will differ (M5 happens to call list_overdue_tasks), but the SDK flow — attach tool → ask → watch mcp_call items — is identical. The closely related M5b · Work IQ lab follows this same pattern against the Microsoft 365 Work IQ MCP server.

Register it in the tool catalog (optional — M5 §6)¶

M5's final section references the server through the Foundry tool catalog instead of baking the key into each agent. Register the server once as a project connection, then agents reference it by connection id — rotate the key in the connection and every agent picks up the change.

In the Microsoft Foundry portal, open your project → Management center → Connected resources (the same connections surface used for Azure AI Search and Application Insights).
Add a Custom / MCP connection: paste the MCP_SERVER_URL (with its key) as the target and give it a name, e.g. project-tracker-mcp. The credential now lives in the connection, not in agent code.
Record the connection name in .env:

MCP_CONNECTION=project-tracker-mcp

M5 §6 then resolves it with project_client.connections.get(MCP_CONNECTION) and passes project_connection_id instead of an inline key.

Approvals for write-capable servers

M5 uses require_approval="never" because the project tracker is read-mostly. If your server has tools that write or spend money, register it the same way but set require_approval="always" in the agent so each call surfaces for a human to approve.

→ Back to M5 · MCP Tools, or review Control plane & governance for the enterprise deployment story.