Claude skills vs. connectors vs. plugins: what is the difference?

Claude skills, connectors, and plugins all extend what Claude can do, but they solve different problems. A skill teaches Claude a repeatable way to work. A connector gives Claude access to an app, service, or changing data. A plugin packages skills and other components so a capability can be installed, shared, and updated as one unit.

The confusion is understandable. Claude's unified directory places all three under Customize, and a plugin may contain both skills and MCP server configuration. Shared discovery does not make the mechanisms interchangeable. Choose based on what the workflow is missing: procedure, access, or packaging.

The difference in one table

This leads to a useful rule: a skill tells Claude how; a connector gives Claude where; a plugin makes the combination installable.

Skills package the method

A skill is an on-demand instruction package. It can define a checklist, decision process, output format, examples, scripts, references, or templates. Claude loads it when the task matches its description, or a user can invoke it directly when the client supports that behavior.

Imagine a source-review skill. It might tell Claude to extract factual claims, match each one to evidence, mark unsupported statements, and return a review table. The source documents can arrive in the conversation; the skill supplies the method. No external service is required.

Use a skill when:

• the same instructions or corrections recur;
• order and completeness matter;
• the output follows a stable structure;
• the necessary input is already in the conversation or workspace; and
• people should be able to inspect and improve the procedure.

A skill does not automatically grant access to Slack, Google Drive, an issue tracker, or a database. It can explain how to use tools that are already available, but the connection and authority come from somewhere else. For the file format and loading model, read what AI agent skills are.

Connectors provide live capabilities

Connectors let Claude access external apps and services, retrieve current data, and—where supported— take actions. Anthropic's connector documentation gives examples such as searching Google Drive, creating Linear issues, and sending Slack messages. Claude inherits the connected person's source- service permissions: if that person cannot access a channel, file, or record, the connector should not make it reachable through Claude.

Most Claude connectors use the Model Context Protocol underneath. MCP gives the host a standard way to discover tools, resources, and prompts exposed by a server. The user-facing term connector focuses on the app connection and authentication experience; MCP server names the protocol-side component. They often describe two views of the same integration, but availability and setup vary by Claude surface.

Use a connector when Claude needs to:

• read information that may have changed since the conversation began;
• search a system of record rather than a pasted export;
• create or update an object in another service; or
• act using the identity and permissions of a connected account.

Connecting a service does not tell Claude your business process. A CRM connector can expose an update_opportunity tool; it does not know which evidence your team requires before a stage changes. That procedure belongs in a skill or an application policy.

Plugins package a complete extension

A plugin is a distribution unit. In Claude Code, a plugin can bundle skills, agents, hooks, MCP servers, language-server configuration, and other components. It gives the bundle a name and update path, namespaces its skills, and lets a team or community install the pieces together.

That makes plugins useful when a workflow has outgrown a single repository folder. A release plugin might contain:

release-workflow/
├── .claude-plugin/
│   └── plugin.json
├── skills/
│   ├── prepare-release/
│   │   └── SKILL.md
│   └── write-release-notes/
│       └── SKILL.md
├── hooks/
│   └── hooks.json
└── .mcp.json

The skills define the release procedure and public-writing rules. A hook can run a deterministic check. The MCP configuration can connect the installed plugin to a release service. The plugin makes the whole set installable and versionable; it is not a fourth way to perform every underlying job.

Claude's web and Desktop directory also exposes plugins. Anthropic says installing one makes its skills available in chat and Cowork. Product surfaces do not necessarily expose every Claude Code component in the same way, so check the listing's stated availability rather than assuming a plugin behaves identically everywhere.

Three choices in practice

A personal writing workflow

You repeatedly turn interview notes into customer stories using the same structure and redaction rules. The notes are uploaded directly. Start with a skill. It carries the method, needs no live integration, and remains easy to inspect.

If the notes must be fetched from Drive, add a connector for Drive. The skill still defines how to select evidence and structure the story; the connector supplies the current files.

A team sales workflow

The team wants Claude to prepare an account brief from current CRM records and recent call notes, then draft—but not send—a follow-up. This needs connectors for the live systems and a skill for the research order, evidence rules, and approval boundary.

Package the combination as a plugin only when the organization needs a consistent, versioned installation across teams or projects. Central packaging is useful after the workflow is proven, not before.

A developer workflow

A repository needs a code-review checklist that understands its architecture. Put a project skill in the repository. If the review must read live issue acceptance criteria, add the issue tracker's connector. If multiple repositories need the same review skill, hooks, and MCP server, distribute them as a plugin.

The unified directory changes discovery, not definitions

Claude's unified directory places skills, connectors, and plugins in one browsing experience. That reduces navigation work, but each tab still has a different action: install a skill, connect a service, or install a plugin.

Review the detail page before choosing. For a connector, check read and write capabilities, authentication, and where it is available. For a skill, inspect the instructions and bundled code. For a plugin, inspect every included component and the update mechanism. A friendly listing page is discovery metadata, not a security assessment.

Security and administrator questions

The risk follows the capability, not the label. Ask:

1. What information can this read under the connected identity?
2. What can it create, change, send, or delete?
3. Which skill instructions or plugin scripts enter the agent's workflow?
4. Which external hosts receive data?
5. Can an administrator restrict, disable, audit, or centrally update it?
6. Which actions require a fresh human confirmation?

Least privilege matters most for connectors because they cross into systems of record. Supply-chain review matters most for plugins because one installation can add several active components. Skills need instruction review as well as code review because natural-language steps can redirect an agent even when no script looks suspicious.

Frequently asked questions

Is a Claude connector the same as an MCP server?

Not exactly. MCP is the protocol; connector is the Claude product experience for connecting an app or service. A custom connector can point to an MCP server, and directory connectors use MCP infrastructure, but the words emphasize different layers.

Can a skill call a connector?

Yes. A skill can tell Claude when and how to use connector-provided tools. The connector supplies the capability; the skill supplies the procedure.

Does every plugin include a connector?

No. A plugin may contain only skills, or it may bundle agents, hooks, MCP servers, and other Claude Code components. Inspect the manifest and directories rather than inferring contents from the word plugin.

Which should I build first?

Start with the smallest missing layer. Write a skill for repeated procedure. Add a connector when live data or external action is essential. Package a plugin when distribution and versioning become the problem.

Next step: Browse LLM Mart's skills and MCP server directory, then choose the smallest capability that closes the actual workflow gap.

Sources

• Claude Help Center: unified directory
• Claude Help Center: connectors
• Claude Code: skills
• Claude Code: plugins