A skill is one folder with a SKILL.md file that tells Claude how to do a task. A Claude Code plugin is a package that installs several pieces as one unit: skills, subagents, hooks that run commands at set moments, MCP servers and more. Of 171 plugins we read in public skill repositories on November 15, 2026, 52 contain only skills and 19 run code of their own.
What is the difference, side by side?
| Skill | Plugin | |
|---|---|---|
| What it is | A folder with SKILL.md, plus optional scripts and reference files | A folder of components, usually with a manifest at .claude-plugin/plugin.json |
| What it can contain | Instructions, examples, scripts, reference files | Skills, commands, subagents, hooks, MCP and LSP servers, executables in bin/, monitors, output styles, themes, workflows, default settings |
| Where it comes from | Copied into a skills folder, such as ~/.claude/skills/ |
Installed from a marketplace with /plugin, or loaded from a folder with --plugin-dir |
| How a skill is named | /review |
/my-plugin:review |
| What runs on its own | Nothing: scripts run when Claude runs them through its tools | Hooks fire at their events and MCP servers start with each session where the plugin is enabled |
| Updates | You copy the new version | From its marketplace, in the background once auto-update is on |
| Switching it off | Move or delete the folder | /plugin or claude plugin disable |
Claude Code behavior from its plugins overview, components and security pages, checked November 15, 2026.
In short: a skill is one ingredient, and a plugin is a box that can hold many ingredients, some of which run by themselves.
What do real plugins contain?
Use This Skill lists every skill folder it finds in public GitHub repositories. On November 15, 2026, we cloned the 47 repositories in our latest crawl (November 9, 2026: 1,966 skills) and read every plugin their marketplace files list.
27 of the 47 repositories publish a plugin marketplace with 174 plugin entries. 171 of those entries point to a folder in the same repository, and we read each one; the other three fetch from npm or from another repository.
| What the plugin contains | Plugins | Files | Repositories |
|---|---|---|---|
| Skills | 123 | 750 skill folders | 23 |
Subagents (agents/) |
97 | 235 | 4 |
Commands (commands/) |
63 | 117 | 3 |
| MCP servers | 11 | 11 servers | 10 |
| Hooks | 6 | 13 hooks | 3 |
Executables (bin/) |
2 | 2 | 2 |
| LSP servers, monitors, output styles, default settings | 0 | 0 | 0 |
| Only skills, nothing else | 52 | ||
| No skills at all | 48 | 5 |
Read by script from each repository's default branch on November 15, 2026. Components were found in their default folders and in any paths the plugin's manifest or marketplace entry declares. A plugin can appear in several rows.
Three things stand out:
- Most plugins are skill packs. 52 contain nothing but skills, and 688 of the 1,966 skills in our directory sit inside a listed plugin folder.
- Subagents and commands come from a few large collections. wshobson/agents alone accounts for 82 of the 97 plugins with subagents and 53 of the 63 with commands. Claude Code's documentation calls commands the older format and recommends writing new ones as skills.
- Code that runs by itself is uncommon. 19 plugins in 14 repositories include hooks, MCP servers or executables. Those are the ones to read before you install.
What runs when a plugin is enabled?
Claude Code's documentation says an enabled plugin is part of every session, not only the sessions where you use it: its MCP servers run alongside each session, and its hooks fire at their events. Here is what the 19 plugins that run code declare.
Hooks. Six plugins declare 13 hooks: 11 run shell commands, and the other two, in one plugin, send a prompt to a model when Claude or a subagent stops. The command hooks run when a session starts (three plugins, for example obra/superpowers, which loads its using-superpowers skill text at startup), before or after a tool call (three plugins, such as two that check each tool call against a policy file), or when a session ends (one).
MCP servers. Eleven plugins declare one server each:
| How the server starts | Servers |
|---|---|
| Remote server reached over HTTP (Sentry, two variants) | 2 |
Package fetched at startup, pinned to a version (for example chrome-devtools-mcp@1.13.0) |
3 |
Package fetched at startup with @latest, or from a Git repository without a commit |
4 |
| Code in the plugin folder, or a tool you install yourself | 2 |
A server that starts from @latest can be a different version tomorrow from the one you read today. Claude Code's security page makes the same point about updates: with auto-update on, the files you reviewed can change on disk.
Executables. Two plugins ship a file in bin/, which Claude Code adds to the PATH of the shell Claude runs commands in.
Is a plugin riskier than a skill?
It can do more, so it deserves a closer read. Claude Code's security page says a plugin "can execute arbitrary code on your machine with your user privileges", that command hooks run with your full user permissions, and that Claude Code runs hooks and MCP servers outside the sandbox. Skills, commands and subagents enter Claude's context as instructions, so they act through the tools Claude already has, and your permission rules apply to those tool calls.
What our scans found in the 19 plugins that run code:
- The skills inside them. 289 of the skills in our directory sit in those plugins. In the November 9, 2026 weekly scan, 285 were clear, 3 marked for review and 1 flagged. The flagged one, modern-python, is flagged for installer lines (
curl ... | sh) in its reference files, not for its hook. - The files around them. Our weekly scan and the live skill check read skill folders, so they do not cover
hooks/,.mcp.jsonorbin/. On November 15, 2026 we ran the same nine patterns over those files in the 19 plugins, 66 files in all. None matched. A clean pattern scan is not a review: a hook that runs a script is fine or not depending on what the script does.
For what a hostile skill or hook could attempt, see what a malicious skill can do. For our weekly results, see are agent skills safe?.
How do I check a plugin before installing it?
Claude Code's documentation gives four steps, and they take a few minutes:
- Run
claude plugin marketplace listto see where each marketplace comes from. - Run
/plugin, select the plugin, and read the Will install list of commands, agents, skills, hooks, and MCP and LSP servers. - Open the plugin's source and read
hooks/hooks.json(the command each hook runs),.mcp.json(each server's command or URL) and every file inbin/. The Will install list shows that a hook exists, not what it runs. - Run
claude --plugin-dir <folder> plugin details <name>on a clone to list everything Claude Code finds, before it is installed.
For the skills inside, paste the plugin's GitHub folder into the live skill check: it reads each skill folder for nine risky patterns and shows the file and line for anything it finds. If you no longer trust a plugin, claude plugin uninstall <name> with the scope you installed it at removes it; the documentation notes the cached files stay for 14 days unless you delete them.
Should I make a skill or a plugin?
Make a skill when you have one method to teach: a review checklist, a release procedure, a writing style. Keep it in ~/.claude/skills/ for yourself, or commit it to the repository so your team gets it.
Make a plugin when you want several pieces to travel together, such as skills plus a subagent, a hook or an MCP server, or when you want to give the same setup to teammates across many repositories with versioned updates. Our guide to sharing skills with your team covers marketplaces, and skills vs subagents vs slash commands and MCP server vs skill explain the pieces a plugin can bundle.
How we counted
The repository list comes from data/skills.json (crawl of November 9, 2026: 1,966 skills, 47 repositories). On November 15, 2026, each repository was cloned at depth 1 from its default branch. For every relative-path entry in a root .claude-plugin/marketplace.json, we read the plugin folder: SKILL.md folders under skills/, Markdown files under commands/ and agents/, hooks/hooks.json, .mcp.json, .lsp.json, bin/, monitors/, output-styles/ and settings.json, plus any of these the manifest or the marketplace entry declares. One entry that fetches a subfolder of its own repository was read from the clone. Skill scan results come from data/scan.json (November 9, 2026). The nine patterns are those in our scanner's rule file, run over every file in hooks/ and bin/, plus .mcp.json and plugin.json. We did not install or run any plugin. Repositories change, so a few counts could differ by the time you read this.