MCP & Skills
An agent's capability layer has two halves. MCP gives the agent dynamic tools it can call at runtime. Skills give it injectable domain capability — instructions, references, and (when needed) bundled scripts, loaded on demand. This lesson covers both, and how they fit.
The capability layer
Every production agent extends the base LLM along two orthogonal axes:
| Axis | Concern | Pattern |
|---|---|---|
| What the agent can do | Run actions, read data, call APIs | MCP tools, resources, prompts |
| What the agent knows | Best practices, conventions, runnable recipes | Skills (on-demand instructions + optional scripts) |
Both can be added incrementally. Agents typically start with a few MCP tools, add domain skills as patterns crystallize, and grow toward a rich capability surface over time.
MCP: a universal tool protocol
The Model Context Protocol (MCP) is an open standard introduced by Anthropic in November 2024. It defines a single, transport-agnostic protocol for AI applications to connect to external systems — replacing the per-LLM, per-app custom integration sprawl that came before.
Think USB-C for AI tooling: a shared protocol instead of a bespoke connector for every integration. MCP adoption is broad but still evolving: