The short answer
An Agent Skill is a folder containing a SKILL.md file — a name, a one-line description and a set of instructions — plus any scripts, templates or reference files those instructions use. The assistant reads only the short description up front, and loads the full instructions when a request matches it. That keeps dozens of skills available without filling the context window.
Skills vs. MCP servers
They solve different problems and work well together. An MCP server gives the assistant new capabilities: tools it can call to reach a database, a browser or an API. A skill gives it know-how: how your team writes a changelog, how to fill in a particular spreadsheet, which steps a code review should follow. A skill can tell the assistant to use an MCP server's tools in a specific way.
How to install a skill in Claude Code
Skills are distributed through plugin marketplaces, which are ordinary GitHub repositories:
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skillsRun these inside a Claude Code session. You can also drop a skill folder into .claude/skills/ in a project, or ~/.claude/skills/ for every project.
In Claude, Cursor and Codex
- Claude (app and desktop): upload the skill folder under Settings → Capabilities → Skills.
- Cursor: there is no native skill format; paste the
SKILL.mdinstructions into a rule under.cursor/rules/. - Codex and other agents: reference the instructions from
AGENTS.mdat the repository root.
Every skill page in the catalog has these steps written out for that skill.
Are skills safe?
A skill is instructions plus, often, scripts the assistant may run. Read the SKILL.mdand any scripts before installing one from a source you don't know — the same care you would take with any code you execute.