Claude Code mods, plugins, skills, hooks and MCP
Choose by the job, not by the label. A plugin packages extensions; a mod observes, intercepts or draws inside Claude Code; a skill guides the model; a settings hook automates an event; MCP supplies tools. They can work together in one workflow.
The short comparison
| Extension | Use it for | How it acts | Check first |
|---|---|---|---|
| Plugin | Install and share a bundle | The distribution unit for skills, hooks, mods, MCP configuration and other extensions | The source and every feature included |
| Mod | A live view or in-session behavior | A plugin hooks module in JavaScript or TypeScript that can observe events, intercept supported calls and draw UI | Declared reach, event behavior and drawing surface |
| Skill | A reusable model workflow | Instructions Claude loads when relevant or you invoke them; can reference scripts and resources | Instructions, allowed tools, scripts and data sent to the model |
| Settings hook | Automation at a lifecycle event | A configured handler, such as a shell command, HTTP endpoint or prompt/agent check | Trigger, execution rights and any external destination |
| MCP | Tools from another service or process | A server exposes capabilities Claude Code can call through the Model Context Protocol | Server trust, scopes, credentials and read/write actions |
A plugin is not a competing behavior type. You might install one plugin that contains a skill, a mod and an MCP server configuration. Installing the bundle does not make each included action equally safe or equally costly.
Start with the result you need
- “Show me what is happening while Claude works.”
- A mod fits a persistent view. Ant Farm follows subagent activity; Redline presents context and usage estimates. Use the Ant Farm guide or Redline guide to read the view, not to direct the model's task.
- “Review changes using our checklist.”
- A skill can hold the checklist and review workflow so the model can reuse it. A view mod such as Play-by-Play can help you inspect changes alongside that workflow, but does not replace the instructions or prove the review is correct.
- “Run a fixed check whenever a file changes.”
- A command settings hook can run a known script at the relevant event. Review the script and failure handling. If the check needs judgment rather than a fixed command, a prompt or agent hook may be more suitable and can use the model.
- “Let Claude read tickets or update an external system.”
- MCP fits external tools. Start with narrow permissions, such as read-only access, and inspect write actions before approving them. A skill can explain your team's triage workflow while MCP supplies the actual ticket tools.
- “Change the working view's look.”
- A drawing mod such as Wardrobe fits. The Wardrobe guide explains which parts it can style and how to restore the default.
Combine layers without confusing their jobs
For a repository review, a skill might define the questions to ask. Claude uses its file tools, or an MCP server for external context. A settings hook can run a local check after a supported event. Ant Farm can display delegated activity and Redline can display estimated usage. A plugin can distribute the reusable pieces together.
Mods themselves use hooks, but they are not the same thing as settings hooks. Mod code runs through the plugin's function-hook interface and can share state between its event handlers and UI drawing. Settings hooks declare event handlers in configuration. Both can react to events; neither name alone tells you whether a handler only observes, blocks an action or calls a model.
Do not treat a skill as a deterministic program: it gives instructions to a model, and results can vary. Do not treat a mod as automatically token-free: some mods call the model. Read each mod's reported reach and source rather than extrapolating from a display-only example.
Review authority, not only the install command
- Plugins and mods: installing third-party code is a trust decision. Check source, updates and the listed files, shell, network, model and blocking-hook capabilities. A validator report is useful evidence, not a sandbox or a promise about future versions.
- Skills: instructions enter model context. Referenced scripts can execute with available tools and permissions. Review both the workflow text and any code; a harmless-sounding name is not a security boundary.
- Settings hooks: a shell command runs automatically at its trigger under the local environment's authority. HTTP handlers can send data out; prompt and agent handlers involve model reasoning. Do not assume they all wait for an ordinary interactive tool approval.
- MCP: remote servers and local processes have different hosting, but both require trust. Restrict credentials and scopes, review server instructions and tool descriptions, and understand where data goes. Tool results are external content, not automatically trusted instructions.
Model usage and service charges are separate. A display-only mod can avoid its own model calls while the observed task still costs tokens. An MCP provider may charge separately from the model. Redline's category estimates do not measure those external bills.
Check the extension and the surface
Use Claude Code 2.1.287 or later in a terminal, or a local Code session in Claude Desktop. Desktop includes its own Claude Code; the official minimum there is 2.1.286, while this collection conservatively requires 2.1.287.
The current official mods documentation also allows mod UI in the VS Code chat panel with extension 2.1.299 or later, in a session that fetches feature flags. That platform capability is not a compatibility test of every mod. Desktop WSL sessions have no plugins. Remote Control shows mod UI in the terminal on the host machine.
Platform support for mod UI is not proof that a particular drawing works there. These practical tutorials use terminal or local Desktop Code. Wardrobe's reviewed version explicitly does not style VS Code or mobile; Redline does not report tests there. Skills, settings hooks and MCP tools have their own configuration and execution constraints, so consult the relevant documentation rather than using a mod's drawing table for everything.
Source and next steps
Reviewed against current official Claude Code documentation on 15 Oct 2026. Examples describe how the features combine, not results of a new live product test.