MCP, Skills & Plugins: the big three
An inflection-point day. Three building blocks that level up your Claude Code — built on one idea: tools. Here's the map, the breakthrough underneath it, and MCP, the first of the three.
Some days move the needle. This is one of them. Welcome to week two, day three — the day your expertise crosses a line. We're covering the three building blocks that everyone in this space talks about, and by the end you'll have complete clarity on when to use each.
Today is a giant of a day, and some of it may feel overwhelming the first time. That's fine — we come back to this material again and again all week and into the next. Enjoy the first pass; you'll hear it a lot.
01The big three at a glance
This ties straight back to Andrej Karpathy's tweet that opened the course — the confusing array of ways to get more out of a coding agent. Here they are, and the relationship between them is the whole game.
🔌 MCP
A standard way to connect Claude Code to tools someone else wrote — the "USB-C port for AI."
📄 Skills
Expertise as simple Markdown (plus optional scripts). Easy to build, easy to share, light on context.
📦 Plugins
Bundles of MCP servers + skills + commands, made properly discoverable through marketplaces.
All three can be installed, uninstalled and activated independently, which is exactly why they confuse people. So we'll build them up in order and finish with a clear decision rule.
02The one idea underneath everything: tools
Before any of the three makes sense, you have to see the innovation they all dance around. Cast your mind back to the four tricks that turn a plain LLM — predictive text on steroids — into something agentic. One of those four stands head and shoulders above the rest: tools.
As it generates text, the model can emit tokens that describe an action to take. We interpret those as "you're calling a tool," run it, and hand back the result. That's the whole trick — and it's why you see Claude Code checking off a to-do list, editing files, and searching. Anthropic baked a set of tools right into Claude Code, and that was the game-changer.
The exciting thing is tools — the ability to take action. MCP, skills and plugins are all just different ways to give the model more tools (or tool-like abilities). Keep that anchor and the rest falls into place.
03MCP — the USB-C port for AI
Back in late 2024, Anthropic realised that if you could bolt anyone's tools into Claude Code in a general way, you'd unlock enormous capability. Frameworks like LangChain already had tool libraries, but they needed a shared standard. So Anthropic announced MCP — the Model Context Protocol, describing it as "the USB-C port for AI applications": an open standard for connecting AI apps to external systems.
MCP originally supported tools, prompt templates and "resources" (sets of information). It's the tools part that took off. But the key point is subtle and worth repeating: MCP isn't the tools, and it isn't the technology that implements them. It's a protocol — an agreed way to package tools so they can be shared into any MCP-ready app.
Not because it's ingenious — other approaches exist (LangChain and friends). It's a big deal because it was so widely adopted. That ecosystem — thousands of servers, ratings, marketplaces — is what makes a single one-line command enough to teach Claude Code an entirely new skill.
Two honest caveats. First, MCP is the interface, not the magic — be excited about great tools, not the tissue that connects them. Second, it's not context-efficient: stitch in too many juicy servers and you fill the window fast. Recent 2026 improvements help (Claude Code no longer eagerly loads every tool), but overuse still degrades performance and pushes you into compaction too soon. That limitation is exactly what skills set out to fix — coming up shortly.
✓ Key takeaways
- The big three — MCP, Skills, Plugins — all add tool-like abilities; each installs and activates independently.
- The core innovation is tools: tokens that describe an action, turning content-generation into action.
- MCP is a protocol — the "USB-C port for AI" — for connecting other people's tools into your app.
- Its power comes from wide adoption, not cleverness: one command equips Claude Code with a whole toolset.
- Watch the downside: MCP can be a context hog and is complex to build. Skills address that next.