Browse the Library for ten minutes and you will run into three words that sound alike: skills, connectors and plugins. All three add something to an AI agent, but they add different things, and mixing them up leads to installing the wrong one.
Skills teach a method
A skill is a folder with a file called SKILL.md in it, plus sometimes a script or a few examples. The file is written instructions for one kind of job: how to lay out a consulting-style chart, how to turn an article into a set of illustrated cards, how to write a grant proposal. When your request matches, the agent reads the skill and follows it.
Skills are the cheapest upgrade there is. They connect to nothing new, so there is little to set up, and a good one makes a visible difference to the result.
Many agents load skills, including Claude Code and Codex CLI. Each skill in the Library has a "Copy install request" button. Paste it into your agent: it reads the files, tells you in plain words what they do, and waits for your OK before installing.
Connectors reach your other apps
A connector lets the agent use something outside your computer's files: a browser, your notes, your code on GitHub, a database. Most are built on the Model Context Protocol (MCP), an open standard that many agents and apps support, so one connector usually works with several agents.
You want a connector when the request mentions another app. "Find last week's meeting notes in Notion" needs one; "summarise this PDF" does not.
A connector is a small program that acts with your permissions in the app it connects to, so treat it with a bit more care than a skill. Install it from the official source, preferably one published by the app's own company, and give it the narrowest access the setup guide allows. Browse the connectors in the Library.
Plugins bundle everything for one agent
A plugin is a ready-made pack for one specific agent: commands, skills, helper agents, connectors and settings, installed in one go through that agent's own plugin system. Claude Code has plugin marketplaces; Gemini CLI calls them extensions.
A plugin made for Claude Code does nothing in Codex, so check the "For" line on its page. Browse the plugins in the Library.
Picking one
What should the agent get better at?
- One kind of job, like writing grant proposalsAdd a skill
- Using another app, like your notes or GitHubAdd a connector
- Everything for one job, set up in one goInstall a plugin
They also combine well. Give an agent a connector to your notes app and a skill for meeting summaries, and yesterday's call becomes a tidy summary in your notes.
Whatever you add, let your agent read it first. Everything in the Library is made by someone else, and we introduce projects without reviewing or testing them. If the agent cannot tell you clearly what something does, whether it runs scripts and whether it talks to the internet, leave it out.