Model Context Protocol (MCPs) are one of the more ambitious developments to hit the AI tooling ecosystem, and they have support from all the major players in the space so are likely to stick around as a standard. Frequently people refer to it as “a usb-c port for AI” but I think this doesn’t quite explain the full picture. I think it’s easiest to explain what it is by explaining how you might use it as a tool developer, and as a service provider.
Tool Developer (e.g.: ChatGPT) – I can now more easily hook into external services that provide MCP integration.
Service Provider (e.g.: Salesforce) – I can now use MCP to standardize the way that tool developers utilize my services.
Notice that this doesn’t necessarily give us anything new, per se, as AI models have been using tools for years already! But this concept reflects where things are headed in terms of model-to-tool interactions, and addresses a constant need to re-invent the wheel with LLM/Service interactions.
But It’s Still Early
That said, MCP is still new. The developer experience isn’t quite streamlined yet. Testing flows aren’t fully nailed down. Documentation is thin. And since MCPs are too new to have been included in most model training data, vibe-coding around MCPs is out of the question.
Most of what’s out there today is high-level or promotional. The clearest information tends to come from talks and demos by the teams working on the spec directly, but this will become out of date fast.
This is where we are in the lifecycle. New standards always start out with a bit of telephone until the dust settles.
A Modern Shift in How We Learn
There’s also a broader shift happening: new tooling like MCP doesn’t always come with robust written documentation initially. Instead, learning happens through video walkthroughs, social posts, discord channels, and community threads. In the very recent past, things happened much more slowly and these things would have been written in indexed and searchable forum posts or blog posts.
That makes it harder to evaluate things critically or learn them on your own terms, especially in a world where we’ve gotten used to asking LLMs for instant context. It’s a not-so-subtle shift in how we engage with emerging technology.
More Surfaces, Not Less Simplicity
The push behind MCPs is about expanding what’s possible by giving models more hooks into external systems, and giving developers more consistent ways to integrate with external systems. It increases surface area and, most interestingly, accelerates tool development for integrations with external systems.
MCP will absolutely be part of shellA’s future, especially for enterprise users who want to make their internal tools callable by name through a shared interface. As the MCP spec matures, we’re keeping a close eye on how it will fit into our broader ecosystem.