Teaching Claude to Query My Payment Database

I asked Claude: "Which provider is performing best and why?"

It didn't guess. It didn't hallucinate. It called a tool, queried my live PostgreSQL database, and returned this:

That's not a chatbot. That's an analyst with direct database access.

What is MCP?

Model Context Protocol (MCP) is an open standard that lets AI agents call external tools in a structured way. Instead of embedding data in a prompt, you expose tools — functions with descriptions — that the AI can call when it needs information.

The key insight: Claude doesn't know what's in your database. But if you give it a tool called GetPaymentSummary with a description that says "Get volume and status breakdown of recent payments", Claude knows when to call it — and what the result means.

"The description attribute is architecture. How you describe a tool determines when Claude calls it, what arguments it passes, and how it interprets the result."

Building PayCore.Mcp in .NET 10

The MCP server is a standard ASP.NET Core project with one NuGet package:

dotnet add package ModelContextProtocol.AspNetCore

// Program.cs — the entire setup
builder.Services.AddMcpServer()
    .WithHttpTransport()
    .WithTools<PayCoreTools>();

app.MapMcp("/mcp");

Three lines of setup. The rest is your tool implementations.

[McpServerToolType]
public class PayCoreTools
{
    [McpServerTool]
    [Description("Get a summary of payment volume and status breakdown. "
                + "Use this to answer: 'What is the total payment volume?', "
                + "'How many payments succeeded vs failed?'")]
    public async Task<string> GetPaymentSummary(
        [Description("Number of recent payments to include (1-500). Default 100.")]
        int count = 100)
    {
        var payments = await _intents.GetRecentAsync(count, null, null);
        var summary = new {
            TotalPayments    = payments.Count,
            ByStatus         = payments.GroupBy(p => p.Status)
                                       .ToDictionary(g => g.Key, g => g.Count()),
            VolumeByCurrency = payments
                .Where(p => p.Status == "captured")
                .GroupBy(p => p.Currency)
                .ToDictionary(g => g.Key, g => g.Sum(p => p.Amount))
        };
        return JsonSerializer.Serialize(summary, _jsonOptions);
    }
}

The four tools

Connecting Claude.ai: the ngrok journey

Claude.ai is a cloud service. Your development machine is not. Connecting them required ngrok — a tunneling tool that creates a public URL pointing at your local server.

The journey had three obstacles:

# Obstacle 1: Wrong URL
app.MapMcp()    # registers at '/' — Claude can't find it
app.MapMcp("/mcp")  # explicit path — works ✅

# Obstacle 2: CORS
# Claude.ai is a different origin — blocked by default
app.UseCors(p => p.AllowAnyOrigin().AllowAnyHeader().AllowAnyMethod());

# Obstacle 3: ngrok version too old
ngrok update
ngrok http https://localhost:7070 --host-header="localhost:7070"
# → https://contusion-affirm-steadily.ngrok-free.dev/mcp ✅

What Claude found

Once connected, the first query was: "What is the total payment volume in PayCore?"

That last line is what separates an AI tool from an AI analyst. Claude didn't just answer the question — it identified the next question worth asking.

The bigger implication

Every .NET developer building a business system now has a straightforward path to making that system conversational. Add ModelContextProtocol.AspNetCore, write a few tool methods with good descriptions, expose an /mcp endpoint, and any Claude-connected interface can query your system in natural language.

This isn't the future of software. It's available today, in .NET 10, with a NuGet package and three lines of Program.cs.