Coding agents are powerful, but out of the box they’re stuck on your computer. They can read your code and run commands. They can’t check your deployments, query your production database, or read your Notion pages.

There are three ways to extend what an agent can do:

CLI tools
skills
MCP servers

They overlap. Here’s how I think about it. If a good CLI exists, that’s the best option: the agent already knows how to run commands. A skill can teach the agent how to use that CLI well. And MCP is the right choice when you need a direct integration, especially when a complex login is involved or many people share the setup.

Skills teach a process

A skill is a way to “teach” an agent something. Remember The Matrix? Something like that.

Concretely, a skill is a folder with a Markdown file, and sometimes scripts and reference documents. It describes how to do a specific kind of work: a pull request review checklist, how to write a Remotion video, how your team publishes a release.

Skills started in Claude Code and are now a standard most tools support. Each tool reads them from its own folder, per project or globally for your user. A few examples:

.claude/skills/<name>/
.cursor/skills/<name>/
.codex/skills/<name>/
.agents/skills/<name>/

Installing a skill means copying files into that folder. Sometimes by hand, sometimes with a command like npx skills i vercel-labs/agent-skills. Then you’re done. When you type a prompt, the agent decides whether a skill is relevant and activates it. If you know a skill will help, mention it in the prompt.

The AI Agent Skills course shows you how to build one from scratch.

MCP connects capabilities

MCP, the Model Context Protocol, is an open standard for connecting AI applications to external tools and data. Anthropic created it, and it works across tools and model providers.

Think of it as an adapter. Instead of the AI being stuck in a bubble, it can reach your development environment and third-party services.

The pieces:

the host, your AI application
an MCP client the host manages
an MCP server, which exposes capabilities

A server can expose tools that do things, resources that provide data, and prompts that are reusable templates. The host connects, discovers what the server offers, and the model picks a capability when it needs one. The host applies your permissions and may ask you to approve. The server does the work and returns the result.

You can run several servers at once. One for your database, one for your deployment platform, one for the browser, one for documentation. Then you can say investigate the spike in 500 errors from the last hour and the agent pulls the logs, cross-references your code, checks the library docs, and proposes a fix. In one conversation.

Know-how and access are different

A skill explains how to review a pull request. A GitHub MCP server supplies the pull request, the files, and the action to post a comment.

Either works alone. You can follow a review skill with plain Git commands. You can use a GitHub MCP tool with no review process at all. Together you get both know-how and access.

Keep it focused

Every MCP server you connect adds tool descriptions to the context. That’s weight on every request. Add the servers you actually use, not every one you find interesting. Disable the ones you’re not using right now.

Standard doesn’t mean trusted

MCP standardizes the communication. It doesn’t make a server safe or correct. A server can expose broad tools, mishandle credentials, return malicious content, or do something different from what its name suggests.

Before connecting one to anything important: check the source, review the exposed tools, use API keys with minimal permissions, run sensitive servers locally, and know what data it can reach.

Skills deserve the same care. A skill can include scripts or tell an agent to run commands. Installing one is like adding code to your development environment. Trust, but verify.

The MCP course goes into clients, servers, configuration, and trust boundaries in detail.

Skills and MCP