Skip to main content

Does anyone actually read the docs?

· 4 min read
Alex Wilson
VP Engineering, RTGS.global

Let's be honest; nobody wakes up excited to read forty pages of technical documentation. (If you do, I'm sure psychological help is available!)

We write documentation because we have to. Comprehensive, accurate, but distinctly bland.

And when you're in the middle of building an integration, you don't want to read a technical novel from cover to cover. You want the answer to the question in front of you, so you can get back to writing code.

Naturally, the modern developer workflow has evolved. The answer is usually:

"I'll just ask my AI agent."

The eager intern without the manual​

Using a generic AI agent for specialised financial infrastructure is a bit like hiring an extremely enthusiastic junior developer on their first day, giving them no access to the company documentation, and asking them to build a payment integration.

They'll produce code quickly. They'll sound confident. They'll invent field names that seem perfectly reasonable.

And then your first request will return HTTP 400 Bad Request.

The agent isn't broken. It just doesn't know how RTGS.global works. Without the right context, it's guessing—and AI agents are very good at making guesses sound like facts.

Until now, giving an AI agent that context meant copying and pasting documentation into a chat window. Which is not exactly the future we were promised.

Enter our MCP server​

To solve this, we've added a Model Context Protocol (MCP) server to our developer platform.

MCP is an open standard that gives your AI agent a way to access external tools and information—in this case, live RTGS.global documentation—rather than relying on whatever it already knows (or confidently thinks it knows).

Instead of hunting through reference guides or feeding documentation into a chat window yourself, our MCP server does it for you. When connected to your development environment, your AI agent can query RTGS.global documentation, API schemas, and integration recipes on the fly.

It gives your agent three tools to work with:

  • search_docs — searches guides, integration recipes, and API endpoints.
  • get_endpoint — retrieves Participant API endpoint details.
  • get_recipe — retrieves step-by-step integration recipes.

When you ask your AI agent:

  • "What are the steps to initiate a Link and Settle payment?"
  • "Show me the request body for the Payaway endpoint."
  • "How do I verify an ED25519 digital signature?"

It calls the appropriate tool, retrieves the relevant documentation, and answers from that—not from a confident guess. You get accurate guidance without leaving your editor.

Choose your own adventure​

Sometimes you want to follow a guide. That's why we have a Use-Cases section, with step-by-step recipes for common integration tasks such as creating relationships, retrieving balances, processing webhooks, and initiating payments.

Sometimes you already know what you want to do and would rather ask a question than go hunting through six pages of documentation. That's where the MCP server comes in.

Once connected, try asking your AI agent:

  • "Walk me through initiating a Link and Settle payment, including the signing and webhook steps."
  • "I need to process an RTGS.global webhook payload. What should I verify before decoding it?"
  • "How do I retrieve the available balances for my network accounts?"
  • "Show me the steps for creating a relationship with another participant."
  • "Explain the FX Quote flow and the events I should expect."

Same documentation. Fewer tabs. Less scrolling. Ideally, fewer opportunities to wonder whether data_base64 is meant to be decoded before or after you've verified the signature.

Plug in your AI tool of choice​

Because MCP is a standardised protocol, you don't need to adopt a proprietary interface or change how you work. You can plug the RTGS.global MCP server into whatever AI development tool your team already prefers—including Claude Code, Claude Desktop, Cursor, and VS Code.

Getting connected takes just a couple of minutes using our server endpoint: https://mcp.rtgs.global/mcp.

For example, in Claude Code, simply run:

claude mcp add --transport http rtgs-docs https://mcp.rtgs.global/mcp

Or add it to your project's .mcp.json or Cursor config:

{
  "mcpServers": {
    "rtgs-docs": {
      "type": "http",
      "url": "https://mcp.rtgs.global/mcp"
    }
  }
}

Faster answers, same engineering discipline​

The MCP server can save you time. It cannot save you from testing.

AI-generated code is still a suggestion, not a production deployment. Test your flows against the local emulator, keep your signing keys safe, validate the schemas, and use normal peer review before anything gets near production.

The difference is that you can now get from "How does this work?" to "Show me the relevant example" without first reading the entire internet.

To get started, visit the MCP Server documentation, connect it to your AI development tool of choice, and ask it something useful.