Skills vs MCP vs plugins: what's the difference?

Comparison table (what / when / example / who maintains)

A skill teaches HOW, MCP gives ACCESS, and a plugin PACKAGES extensions for installation and sharing.

That shorthand matches Anthropic's extensibility overview. A plugin can contain both skills and connections, but it does not have to contain every type of extension. My judgement: choose the missing piece in your workflow before adding a bundle.

What	When I would choose it	Example	Who maintains it
Skill: reusable procedure	The assistant has the material but needs your method	Weekly update format and evidence checklist	Process owner or skill author
MCP connection: access to exposed capabilities	The task needs an authorised system	Retrieve approved account records	Server operator and connection owner
Plugin: installable package	Several extensions belong together	Team procedure plus its connection	Package maintainer, with component owners

Skills

Build a skill when you keep correcting the same approach: the wrong structure, missing evidence, inconsistent terminology or a skipped review step. You are teaching a method that should survive beyond one conversation.

For a weekly update, I would specify the inputs, sections, source references and handling of missing information. “Write professionally” is too vague. “List decisions needed separately from completed work, and attach evidence to each claim of completion” gives the assistant something observable to do.

A skill is a directory with SKILL.md instructions and optional supporting material. Anthropic's skills overview describes progressive loading: the name and description are available first, the instructions load when relevant, and additional files are opened as needed. Skills therefore have a context cost, even before the full procedure loads.

Keep the description useful for selection and the body useful for execution. Put a long reference document in a supporting file rather than repeating it throughout the instructions. Start with the smallest procedure that fixes an observed problem. Browse skills or follow how to build your first skill.

MCP/connectors

Add a connection when the missing piece is access. A skill cannot retrieve a private record merely by mentioning the system's name. The application needs a working connection, suitable exposed operations and appropriate authorisation.

MCP supplies a common protocol for that access. “Connector” is the product-facing label you may see during setup; some vendors use it for integrations beyond MCP. Our MCP explainer covers servers, clients and discovery, while the connectors directory helps frame the practical options.

I would test access separately from method. First confirm that the connection returns the right record. Then check whether the skill interprets and formats it correctly. Otherwise, an empty result can be mistaken for a writing problem, or an accurate retrieval can hide a poor procedure.

Anthropic's skills and MCP article treats them as complementary. For my client-update example, the connection could supply authorised source material and the skill could define how to turn it into a reviewable draft. Neither component decides whether the client should receive it.

Plugins

Use a plugin when distribution becomes the problem. A team may want the same procedures and connections installed together, with a clear package to maintain. That can be useful after the individual workflow already works.

The Claude Code plugin overview documents bundles of skills, agents, hooks, MCP servers and other components. Those pieces can also work independently. A plugin is therefore a packaging choice rather than a requirement for every reusable procedure.

Inspect the components before installation. Ask who updates them and what changes when the package updates. A familiar package name does not tell you whether it contains a simple writing checklist or tools that can change shared records.

Picking one: decision flow

Start by describing the failure in the existing workflow. Then follow this sequence:

1. Missing method? Write a skill with explicit inputs, steps and acceptance criteria.
2. Missing information or action? Evaluate a suitable connection and its permissions.
3. Need both installed together? Consider a plugin containing the relevant pieces.
4. Still unclear what good work looks like? Clarify the task before adding extensions.

In Craft Assist, AI helps me think; I decide. In Augment, AI carries several steps; I review. In Adapt, AI works on files within rules I set; I stay on call. These modes describe responsibility, not extension types.

The seven-layer map is model, harness, connections, skills, product, agency, my role. MCP fits connections; skills carry procedures; plugins package product extensions. My role still includes deciding the review point and what the assistant is permitted to change.

Cross-tool equivalents

Skills extend beyond Claude. Cursor's Agent Skills documentation explicitly supports the open standard, and GitHub's agent skills guide documents support in named Copilot surfaces. Calling them only rough equivalents would understate that support.

Tool	Verified position	What I would recheck
Claude	Skill directories and account-enabled skills documented	Installation surface and dependencies
Cursor	Agent Skills standard supported	Discovery location and available tools
GitHub Copilot	Agent skills supported across listed agent clients	Specific client and organisation policy

A portable format does not guarantee identical behaviour. A procedure might assume a tool, path or permission absent from another application. Read the instructions and test a representative task after moving it. This comparison makes no claim that rules files or custom instruction files are automatically interchangeable with skills.

FAQ

Should I build a skill or add MCP?

Build a skill when the assistant needs your procedure; add an MCP connection when it needs access to a system. You may need both. Identify the failure first, then test the missing component separately. An elaborate procedure cannot repair unavailable records, and a successful connection cannot define your preferred review standard.

Does every plugin contain skills and MCP?

No. A plugin packages supported components and can contain a subset of them. Check its contents rather than inferring capability from the label. I would look for the procedure it adds, the systems it reaches and the maintainer responsible for changes before deciding whether the package belongs in my workflow.

Do skills work outside Claude?

Yes. Cursor and GitHub Copilot document Agent Skills support in their stated clients. That verifies the format, not identical execution of every workflow. Check file locations, available tools and permissions, then run a representative task. A skill tied to one product's capabilities may need adjustment before it works elsewhere.

Do unused skills consume context?

Yes, their available metadata uses context, while the full instructions load when selected and supporting files load as needed. Progressive loading reduces the amount carried upfront. Keep names and descriptions concise but specific, and disable irrelevant extensions when maintaining a clear working setup matters more than having a large catalogue available.

Build your first reusable workflow in Craft Adapt, with explicit rules and a human checkpoint.

Verification notes

• Cross-tool support is verified for the documented Agent Skills surfaces. Identical behaviour, plugin portability and conversion of rules files are not assumed.
• Examples, ownership assignments and Craft mode guidance are editorial recommendations.
• No unresolved factual claims are presented as verified. No prices or model versions are needed for this distinction.