Cowork plugins vs skills vs connectors, explained
A skill tells Claude how to do a task. A connector gives it access to a service. A plugin packages both. How to tell them apart and which you need.
A skill tells Claude how to do a task. A connector gives Claude access to a service. A plugin packages skills, connectors and other components into one install. The three overlap in everyday conversation, but they solve different problems, and picking the right one saves you from installing more than you need.
This comparison is based on Anthropic's current documentation, linked at the end. Where the docs are version-sensitive, we say so.
What a skill is
A skill is a set of instructions for a workflow: how to draft a campaign brief, how to check brand voice, how to structure a financial review. It tells Claude *how* to do something, step by step, in your team's way.
What a skill does not do is give Claude access to anything. A skill that describes how to write a sales email still needs you to paste in the account notes, or a connector that can reach your CRM. Instructions and access are separate things.
Skills travel well. According to Anthropic's plugin documentation, the skills and commands bundled in a plugin work in chat, in Cowork, and in Claude Code. If the missing piece in your work is the *procedure*, a skill is the smallest fix. Our skills section lists the ones bundled with plugins in this directory.
What a connector is
A connector links Claude to an external service through MCP (the Model Context Protocol). Google Drive, Gmail, Slack and similar services are reached through connectors. A connector is about *access and available actions*, not about knowing what to do with that access.
Two things worth knowing before you connect anything:
- In Cowork, connectors reach external services through Anthropic's cloud. A custom connector must point at a server that is reachable over the public internet from Anthropic's IP ranges. A service behind your office firewall or on a private network will not work through that path.
- Plugins can also include *local* MCP server components, and those behave differently: they run on your computer through the desktop app. Anthropic's docs list plugins with local MCP servers as desktop-only.
So "this plugin has a connector" is not one fact but two: where the connector runs, and what it can reach. A connected account can still lack the permissions or the data your task needs.
What a plugin is
A plugin bundles skills, connectors, sub-agents, hooks and commands into a single package, so a whole workflow installs at once instead of you assembling the pieces. A plugin does not have to contain every component type. Some are mostly skills with no connector at all.
The current Anthropic docs describe host support per component, not per plugin:
- Chat: plugin skills and commands work. Hooks and sub-agents do not run in chat; they appear grayed out.
- Cowork: skills, commands, hooks and sub-agents run. Connectors go through Anthropic's cloud, and local MCP components work through the desktop app.
- Claude Code: the plugin runs in full on your computer, including skills, sub-agents, hooks and any local MCP servers. Code reads installed plugins from your Claude account (one-way sync) when you sign in with the same account; the Help Center names v2.1.273 or later and notes that sync does not happen with API-key or cloud-provider sign-in such as Bedrock.
Earlier versions of Anthropic's Cowork guide said plugins were not used in chat. The current Help Center and Cowork docs both say plugins you add are saved to your account and work across chat, Cowork and Code, with the component limits above. If you see older articles claiming plugins are Cowork-only, they are describing an older state.
Which one do you need?
Choose the smallest thing that closes the gap:
- The procedure is missing. You have the source material but want the task done a specific, repeatable way. A skill is enough. Paste in your notes and check the output against your own facts.
- The access is missing. The task needs live data from a service, like real campaign numbers or open tickets. You need a connector, and a minute spent reviewing what it can read and do before enabling it.
- You want a checked, role-specific setup. A curated plugin bundles the procedure and the access for a function like sales, legal or marketing. Inspect what is inside before installing.
For a first test, supplied text is often enough. If you want to analyze a fictional campaign brief, paste the brief; you do not need to connect your ad account to find out whether the output is useful. Connect only what the task needs.
What to check before trusting any of the three
Whether you are browsing our plugin directory or Anthropic's own marketplace, look for:
- Publisher and source. Official marketplace inclusion and who wrote the package are two different facts.
- Component inventory. Does it contain skills, commands, sub-agents, hooks, connectors, local MCP servers? A label like "works everywhere" needs evidence per component.
- Where connectors run. Cloud-reachable service, or a local component that needs the desktop app.
- Install path and version. How it installs, and whether compatibility is documented or actually tested.
A manifest or listing is metadata. It does not prove safety, output accuracy or fitness for your organization. Our plugin security guide goes deeper on the checks, and the official plugins page shows which packages come from Anthropic itself.