If you’ve opened Claude’s Customize settings and weren’t sure whether a task called for a Skill, a connector, or a Plugin, you’re not alone.
The names sound similar, and people often use them interchangeably in casual conversation even though they do distinct jobs. This guide walks through what each one actually is, how they relate, and which one fits a given workflow, using a real example along the way instead of a hypothetical one.![Connect HubSpot to Claude For Personalized Insights [Free Connector]](https://no-cache.hubspot.com/cta/default/53/64493837-f4dd-43b0-abee-3ef154e06092.png)
TL;DR: Skills teach Claude a procedure, MCP connectors give it authorized access to an outside tool or data source, and Plugins package both together so a team can install everything at once.
Table of Contents
What are Claude skills, plugins, and MCP connectors?
The easiest way to keep these three straight is to separate access from procedure from distribution. Each layer answers a different question.
MCP connectors are access.
A connector links Claude to an outside service — your files, your calendar, an issue tracker, a CRM — through a live connection called an MCP server, which the service itself runs so Claude can reach it. Once you turn a connector on, Claude can read from that service and take action in it, inside your existing permissions.
Skills are procedure.
A skill is a set of written instructions, sometimes with scripts and reference files attached, that Claude loads when your request matches what the skill covers. A skill teaches the how: your team’s report format, your review checklist, the brand voice you always want applied. A skill can tell Claude how to use a connector’s data, but it can’t reach a live service on its own. That’s what the connector is for.
Plugins are distribution.
A plugin bundles skills, MCP connectors, commands, and sub-agents into a single package that installs once and can be shared with a whole team, instead of asking every person to add each piece separately.
If you want a broader primer on how plugins work as a category of software, What are plugins? covers the concept outside of Claude specifically. Inside Claude’s ecosystem, a plugin can hold any mix of the pieces above, and installing it adds everything in one step.
A short glossary, since these terms get thrown around loosely:
- Skill: written instructions Claude follows when a task matches the skill’s description.
- Plugin: an installable package of skills, connectors, commands, and sub-agents.
- MCP (Model Context Protocol): the open protocol that lets Claude talk to outside tools and services in the first place.
- MCP connector: one live connection, built on MCP, between Claude and a specific service.
- Sub-agent: a specialist Claude can delegate part of a task to; this only runs in Cowork and Claude Code, not in regular chat.
How do these three layers fit together?
Picture a request moving through Claude in three stages: a skill can shape how Claude approaches the task, a connector lets Claude reach outside the conversation to get or change real data, and a plugin is the optional wrapper that ships a skill and a connector together as one package.
None of the three sits “above” the others in a power ranking.
A connector without a skill still works fine, as Claude will simply use its own judgment about how to phrase requests to your CRM.
A skill without a connector also works fine — it just can’t touch live data, only whatever you paste into the conversation or whatever else Claude already has access to.
A plugin is only useful when you actually want several of these installed and shared as a unit; a single skill or a single connector doesn’t need one.
Which should I use: a skill, plugin, or MCP connector?
Match the problem to the layer, not the other way around:
- Use an MCP connector when you need live access to an outside tool or data source. If the task involves anything Claude can’t see just by reading the conversation — your calendar, your ticketing system, your CRM — a connector is the only layer that reaches it.
- Use a skill when you need a repeatable procedure, a template, or a brand voice applied consistently. Anything you find yourself correcting Claude on the same way every time — a report format, a QA checklist, a tone guide — is a good candidate.
- Use a plugin when you need a team-ready bundle that other people can install in one step. Once you’ve got a skill and a connector that work well together, packaging them into a plugin means your team doesn’t have to hunt down and configure each piece separately.
Combinations are common, and often the right call. A connector supplies the data, a skill decides what to do with it, and a plugin is how both reach everyone else on the team at once.
What does a combined workflow look like?
Take a weekly pipeline review, for example. On its own, none of the three layers gets you the full workflow:
A connector supplies the pipeline context. With the HubSpot Claude connector turned on, Claude can pull deal stages, recent activity, and engagement history straight from HubSpot, scoped to whatever a given user is permitted to see in HubSpot already.
A skill defines the review format. A skill can tell Claude exactly how your team likes a pipeline review structured — which fields matter, how to flag stalled deals, what counts as “needs attention” — so the output looks the same every week regardless of who asks for it.
A plugin packages it for the team. Bundling that skill with the connector into a plugin means a sales manager installs one thing, rather than separately connecting HubSpot and then hunting down the right skill, and rolling it out to a whole team is a single install instead of five.
This pattern is worth building toward because most teams end up connecting to more than one AI tool at once. HubSpot’s own changelog announcing the connector noted that over half of its customers already use multiple LLMs to get work done, which is part of why having your CRM context available consistently, regardless of which layer is doing the work, matters more than picking a single “winning” tool.
What are examples for marketing, sales, and service Teams?
Marketing
- Access: a connector supplying campaign performance and content-analytics context.
- Procedure: a Skill that applies your team’s brand voice to any draft, consistently.
- Distribution: a Plugin that packages both so every writer on the team gets the same setup.
- Best for: teams that produce a high volume of on-brand content and want consistency without a manual style review every time.
Sales
- Access: a connector supplying account and pipeline context.
- Procedure: a Skill that defines a standard follow-up sequence or note-taking format after a call.
- Distribution: a Plugin bundled as a “rep toolkit” so new hires get the same setup on day one.
- Best for: teams onboarding reps regularly, where you’d otherwise have to explain the same CRM habits to each new person.
Service
- Access: a connector supplying ticket and knowledge-base context.
- Procedure: a Skill that defines how tickets get triaged and routed.
- Distribution: a Plugin that packages the triage skill with the ticket connector as an agent toolkit.
- Best for: teams that field a high ticket volume and need consistent triage regardless of who’s on shift.
Pro tip: Start with the smallest layer that solves the actual problem. If a single skill already gets the format right, don’t reach for a plugin just because it feels more “official.” Package things into a plugin once you’ve confirmed the skill and connector actually work well together and more than one person needs them.
How do I set up and govern these layers?
Setup and governance are worth covering together, since the controls at each stage carry over into the next.
Start with one repeatable task rather than trying to stand up your whole stack at once. Pick a process you already do by hand every week, build or find the skill for it, connect the data source it needs, and see whether the combination actually saves time before you package anything for a wider team.
Test the skill before anyone relies on it.
Anthropic’s own skill-creator skill runs a skill against test prompts you define and reports back so you can revise the description or instructions before rolling it out.
Authorize connectors deliberately.
When you add an MCP connector, you review and approve the access it’s requesting, and you can set tool-level permissions afterward: Always Allow, Needs Approval, or Blocked, for individual actions rather than the connector as a whole.
Verify plugins before installing from outside the directory.
Anthropic runs automated validation and a security scan on every plugin version submitted to its public directory, plus a human review before a new listing goes live. That review doesn’t extend to a plugin someone uploads from an unofficial marketplace URL or a local file, so only add those from sources you actually trust.
Pilot before you roll out organization-wide.
On Team and Enterprise plans, you can share a plugin with a handful of specific people first, confirm it works the way you expect, and only then set it to install by default or make it required for everyone.
Set permissions deliberately, not just once.
The HubSpot Claude connector, for instance, respects whatever permissions already exist in HubSpot. A rep only sees the pipeline data they’re already allowed to see in HubSpot itself, so existing HubSpot permissions constrain the context available through the connector rather than the connector introducing a separate permission system.
Keep versioning and rollback in mind.
Organization admins on Team and Enterprise plans can see a plugin’s full version history and revert to an earlier version if a new one causes problems, without losing any of the version history in the process.
Know where audit logs live if you need them.
On Enterprise plans, Organization Owners can export audit log data from Organization settings and Data and Privacy, and the same data is available programmatically through Anthropic’s Compliance API for teams that need to feed it into their own compliance tooling.
None of this amounts to a security or legal guarantee. It’s a description of the controls that exist today, and you should evaluate them against your own team’s requirements rather than treating any of it as a compliance sign-off.
Frequently Asked Questions About Claude Skills, Plugins, and MCP Connectors
Do I need a developer to use Skills, Plugins, or MCP connectors?
No, not to use ones that already exist. Adding a connector from Anthropic’s directory or installing a plugin is a point-and-click process from Claude’s Customize settings. Writing your own skill from scratch is also mostly writing plain instructions in a text file — no code required unless you want to add an optional script. Building your own MCP connector or submitting a plugin to the public directory is the part that takes engineering work.
What is the difference between Claude Skills and MCP in one sentence?
A Skill teaches Claude how to do a task; MCP is the protocol that lets Claude reach outside tools and data to get the information it needs to do that task in the first place.
When should I standardize a process as a Plugin rather than a single Skill?
Once more than one person needs the same skill, or the same skill needs to travel together with a specific connector, package it as a plugin. A single skill is fine as long as you’re the only one using it and it doesn’t depend on a connector someone else still has to set up separately.
How do Claude Skills compare with tools or slash commands?
A “tool,” in Claude’s terms, is usually something a connector exposes — a specific action Claude can take in an outside service. A slash command is a named shortcut, often bundled inside a plugin, that runs a specific skill or action when you type it directly rather than describing the task in plain language. A skill is the instructions themselves; commands and tools are two different ways of triggering something.
Is there an example of a CRM-grade MCP connector I can try?
Yes. The HubSpot connector for Claude is a working example you can install today. It requires an active HubSpot account on any tier and a paid Anthropic plan (Pro, Max, Team, or Enterprise), and it can create and update records like contacts, companies, deals, and tickets, though it doesn’t delete anything on any object.
Solve with the right tool.
Skills, plugins, and MCP connectors aren’t competing tiers of the same feature. They’re three different jobs, and the workflows that actually stick tend to use more than one at a time. If you want to see the access layer in action with a real CRM rather than a hypothetical one, the HubSpot connector for Claude is a reasonable place to start, and it pairs naturally with whatever skill or plugin your team builds around it.
