Magic Comes to Prompting (aka MCP)

What happens when your LLM grows arms and legs?

For all the amazing feats AI can achieve, it is still prone to embarrassing stumbles. Take this classic maths example. Ask Claude which is bigger: 9.11 or 9.9?

Oops.

ChatGPT is a bit better:

But replying ‘yes’ and then contradicting the original question isn’t ideal either.

Enter tools

The problem is the models aren’t inherently good at mathematics; these are language models after all. But give them access to tools - a calculator or code, for example - and they do better. Here’s Claude again, but told to use code to answer the question.

OK. So it couldn’t resist sharing it’s initial (wrong) gut feel. But it corrected itself. Note, though, it took three tries with a coding tool before the answer from the tool was strong enough to override Claude’s initial gut feel. (I can empathize - many times I’ve had to repeat a calculation because I didn’t want to believe the first answer!)

So tools are good. And some models use them extensively. If you’ve used o3 you’ll be well aware how extensively it uses tools - it’s often hard to get it do anything without pulling in a tool. Here is o3 calculating the square root of 3.

But what if you want to use your own tools? That’s where MCP - Model Context Protocol - comes in. It’s a mechanism that gives AI models access to arbitrary tools. Want integration with PowerPoint? Yup there’s a MCP server for that. Or Word. Or Slack. Or Git. Or the filesystem.

MCP is a relatively recent innovation created by Anthropic last November. Although ChatGPT seems to think it was invented by OpenAI last May…

Claude, the main beneficiary of MCP, doesn’t even know about its new super-power since MCP postdates its knowledge cutoff.

There’s a long list of publicly available MCP servers here. And if the one you want doesn’t exist, well, you can build your own. I got Claude to build me one to connect my OpenAI vector database to Claude. Kwool.

Magic Comes to Prompting aka MCP

MCP adds another layer of magic to AI. Now the tools can interact with the files on your machine. Start to perform agentic flows. For example…

Earlier this week I was asked to support an AI app I created 3 months ago. In doing so I discovered an unexpected consequence of creating things quickly. If you only spent two weeks creating a thing - and you’re me - you’re unlikely to remember much about it. Oh dear.

But, not to worry. My Claude is hooked up with a raft of MCP tools. It soon found the source code and gave me an overview:

And a little later:

So now I’ve got a doc to remind me how the tool works. Even better it came complete with Mermaid diagrams documenting the key flows; I’m a sucker for a good diagram so I was delighted.

A little later I wanted to find out when a feature called “joe.pst” was added. And so the git MCP comes to the rescue..

MCP simplifies things; no more copying and pasting into other apps. Everything can be done from the Claude desktop app.

MCP is powerful for coding, but it can do so much more as well.

It’s not just coding

I was interested in learning more about AI music generation. So I asked Claude to research the subject. But rather than give me an artefact (in markdown format) I wanted the output as a Word doc. Step forward Word MCP server.

And several minutes later I had a Word document. Admittedly the formatting isn’t quite how I’d like it. And there’s a lack of diagrams. But it’s more than good enough to give me a leg-up. No longer do I have to copy and paste between Claude and Word. Converting the doc into the bones of a PowerPoint presentation is simplicity itself thanks to the PowerPoint MCP server. Adding the output to a git repo is trivial. I can describe what I want to happen; I don’t have to use the raw tools themselves. I don’t have to remember weird syntax.

Last December I wrote about how AI spelled the end-game for Microsoft Office - MCP brings us one step closer to that reality.

The downsides

MCP sounds wonderful. When it works, it is wonderful. But, currently, it’s a wild-west. It’s hard to use. First you’ve got to use the Claude desktop app. Then you’ve got to install the MCP servers. There are multiple install options. NPX. Smithery. PIP. Python. UVX. Assuming you win the install battle then how do you start the server. Manually? Or will Claude magically start it? What if it doesn’t start? Fortunately you can feed the logs to Claude and get it to help diagnose the problem.

Configuration is handled via your claude_desktop_config.json file. First you’ve got to find that file. Then you’ve got to manually set multiple parameters. Config is only read on first startup; but there are multiple processes behind the Claude desktop app - and closing the window only gets rid of some of them. So you may need to use Process Explorer to kill the rest of them otherwise the config won’t get re-read.

Then, sometimes, the server crashes or fails without warning. You’ll look at the Claude output and discover you’re back at the previous prompt, questioning what happened. No diags. No explanation. Just some gentle gas-lighting.

For me, I can cope. Sure, it’s a little frustrating. But I know how to make it work. I’m willing to live with the rough corners. But for the normal user? Not a chance.

So what?

Despite the current bumpy experience, the direction of travel is clear. MCP is far too powerful to be ignored. The integrations will improve. Claude will learn how to debug the configuration problems itself. Maybe even fix them. The labs will start to provide many of the more common integrations out of the box. You’ll be able to configure the integrations via Claude - not via a cryptic config file.

The end for old-fashioned clunky interfaces (I'm looking at you Microsoft Office) is drawing closer - and some may say about time too. Once we can create documents, presentations, and code through natural conversation who would want to use a confusing clutter of menus, ribbons and dialog boxes?

And then we'll all have access to a new set of magical tools that enable our AIs to do things we never imagined before.