The viewer’s question about MCP
In the latter stages of the livestream, a viewer who is also an agency owner asked a very straightforward but relevant question:
It seems that MCPs are popping up everywhere. What criteria should we use to evaluate them?
As the conversation progressed, they also asked about an example of a major MCP, and if they even need multiple MCPs. Together, these three questions cover precisely what most people and agencies want to know before connecting anything to their tools.
Below we’ll show you Nathan’s answers, his opinions, and expand on it all so you can leave knowing what your next steps should be.
What is an MCP?
MCP stands for Model Context Protocol. If that doesn’t immediately explain what it does, you’re in good company.
Like with most other protocols, the name doesn’t really show what it does. In reality, MCP is an open standard introduced by Anthropic (the folks behind Claude) as a means to allow AI models to talk to outside tools and data sources. The vital point here is that it’s meant to eliminate the need for custom-built integrations for every single combination of AI and external tool.
Before MCP was introduced, connecting an AI model to a CRM, for example, meant writing bespoke code for that one pairing. Multiply that across every model and every possible tool that exists online and, yes, it’s a mess. Engineers call it the MxN problem: M models times N tools equals so many one-off integrations we probably don’t even have a number for yet.
To make things even easier, Nathan explained what MCP is very well during the livestream:
An MCP, if you're not familiar, is like a translator that sits between some app and your AI tool, and lets your AI tool talk directly through it to that app. It's like an API but built for AI.
And that’s really the whole thing. The MCP server exposes (explains) what a tool can do, like sending an email, in a format an AI model can understand and act on.
Why MCPs are suddenly everywhere
After Anthropic made MCP open source, it’s become the closest thing we have to a standard now.
Many major software categories have incorporated an MCP in one way or another, and the rate at which that is happening has only increased recently. A couple examples from the livestream that our viewer asked about are QuickBooks (Nathan’s choice for accounting software) and Rocket.net. QuickBooks has an MCP for bookkeeping, while the Rocket.net MCP allows AI agents to talk directly to the hosting infrastructure.
That’s why it feels MCPs burst into the scene and are now everywhere. Companies are actively, even aggressively, adopting the protocol. Nathan’s agency, Brilliant Webworks has multiple that they use daily:
I don't know exactly how many we have connected in the Brilliant Web Works world, many, many dozens. Each MCP connects Claude to one thing.
And while each MCP adds more functionality to your workflow, it’s still code, often written by someone outside your organization, which you grant access to your files, client data, hosting environment, etc.
The more your MCP count grows, the more vetting you have to do so you can trust it will work properly and securely and not cause havoc in your systems.
How to evaluate an MCP before allowing it access
Ultimately, this is the heart of our viewer’s question, and Nathan’s answer mirrors something many people already know how to do: vetting a WordPress plugin.
Nathan explained it precisely:
It's a lot like evaluating a plugin, honestly. As you look at a plugin in the WordPress plugin directory: who was the author, is it the company that wrote the MCP for the thing you want to attach to, how many reviews does it have. It's the same sort of criteria.
Based on his answer, here are the things to look out for when evaluating a new MCP:
Beyond what Nathan mentioned, our blog about choosing and running plugins safely elaborates further with a few more list items, where hosting fits in, and has an actionable checklist. While it’s about plugins, the general advice still applies.
It’s vital to be careful what MCPs you connect to your systems. A large-scale review of nearly 1,900 public MCP servers found real risks: about 7% had exploitable general vulnerabilities, while about 5% had MCP-specific issues like tool poisoning.
The U.S. National Security Agency has also published guidance on MCP, following its concern that it allows a server to act on a client’s behalf, not just respond to requests. That opens attack paths that don’t exist in traditional server setups.
Of course, none of that means MCPs are inherently dangerous. It means that the vetting Nathan described is simply the baseline.
Do you need multiple MCPs?
Asking if you need multiple MCPs is like asking if you need some WordPress plugin. Are you going to make use of it? Then yes, you probably want it.
If you want your AI agent of choice talking to your project management tool, accounting software, hosting environment, or something else, those are multiple separate MCPs you’ll have to find and vet. Of course, broader MCPs like the Google Workspace one (covering Gmail, Drive, and Calendar) can trim that number, unless you want to limit what the AI can access.
Once you’ve properly vetted your MCP stack, you have to check if your sites themselves are ready to work with it. Watch our “Is your site agent-ready?” livestream to find out how.
Finally, if you have any questions about MCPs, AI in general, WordPress, or web hosting, join us every Thursday at 2 p.m. ET for Office Hours. Nathan and the community talk about various topics each time, so you are always bound to learn new things!
FAQ
Is an MCP the same thing as a "connector"?
Functionally, yes. Tools like Claude and ChatGPT use the term "connector" for the same kind of direct link to an outside app, think Claude to Slack, or Claude to Canva. Under the hood, a connector and an MCP are solving the same problem: giving an AI tool a standardized way to talk to a specific outside system.
Can an unofficial or low-quality MCP cause harm?
It can, in the same way an untrustworthy WordPress plugin can. Because an MCP is just code that your AI tool executes with a level of trust, a poorly built or malicious one can expose more access than intended. This is an active area of security research in 2026, and it's the exact reason Nathan's evaluation criteria (official source, real reviews, code review) matter.
How many MCPs does a typical agency need?
There's no fixed number. It scales with how many tools you want your AI assistant to actively work inside. Some agencies connect a handful; others, like Nathan's, run "many, many dozens." The right number is whatever matches the tools you're using day to day, not a target to hit.