MCP, Skills, Apps and Plugins: A Terminology Map

Vendors use the words MCP, connector, skill, app, extension and plugin differently, and some overlap. In their own docs, MCP is a protocol, a connector is a connection to an MCP server, a skill is a folder of instructions, and a plugin is a package. This map lists the definitions, with sources and dates.

How to read this map

We fetched and read every page listed below on 2026-10-03. Each row states what that vendor's documentation says, and only that. We did not test any product, and Anavem has not audited any software named here. The map does not rank vendors or options.

One gap is deliberate. We did not read OpenAI's documentation for this guide, so this article makes no claim about ChatGPT apps, connectors or plugins. A name that is missing here is not evidence that it does not exist.

What the sources say: the terms, by owner

Term	Who documents it	What that documentation says	Source (read 2026-10-03)
MCP	MCP project	An open-source standard for connecting AI applications to external systems.	MCP introduction
MCP host, client, server	MCP project	The host is the AI application. The client keeps a connection to one server. The server provides context.	Architecture overview
MCP extensions	MCP project	Optional additions to the core protocol that need support from both client and server. The specification names Tasks, Skills over MCP and MCP Apps.	Specification overview
MCP Apps	MCP project	An extension for interactive interface elements, such as charts, forms and video players, shown inline in conversations.	Specification overview
Connector	Anthropic (Claude)	A link between Claude and an outside app or service. Each is a connection to an MCP server that the service runs.	Connectors overview, extend overview
Desktop extension	Anthropic (Claude)	A local MCP server packaged as an MCP Bundle (MCPB) that runs on your device in the Claude desktop app.	Add an unlisted connector
Skill	Anthropic (Claude)	Written instructions, optionally with scripts and reference files, that Claude loads when a task calls for them.	Extend overview
Plugin	Anthropic (Claude)	An installable package that can hold skills, MCP connectors, commands and agents.	Plugins overview
MCP servers in Claude Code	Anthropic (Claude Code)	Claude Code connects to external tools and data sources through MCP. Servers can be local stdio processes or remote HTTP servers.	Claude Code MCP docs
Agent Skills	Agent Skills project	A folder with a SKILL.md file of metadata and instructions, which can also bundle scripts and other resources. The format was originally developed by Anthropic and released as an open standard.	agentskills.io
Agent skills in VS Code	Microsoft (VS Code)	Folders of instructions, scripts and resources that agents load when relevant. The page says the open standard works across multiple AI agents, including GitHub Copilot in VS Code, Copilot CLI, the Copilot app and the Copilot cloud agent.	VS Code docs
MCP in GitHub Copilot	GitHub	GitHub defines MCP as an open standard for how applications share context with large language models. The page lists IDEs, Copilot CLI, the Copilot app and the Copilot cloud agent as supported places.	GitHub Docs
MCP and plugins in Cursor	Cursor	MCP lets Cursor connect to external tools and data sources. Servers are installed from the Customize page or configured in mcp.json. The page also mentions Cursor plugins and a Marketplace.	Cursor docs
Extension (Gemini CLI)	Google (Gemini CLI)	A package of prompts, MCP servers, custom commands, themes, hooks, sub-agents and agent skills.	Gemini CLI docs

Where names overlap or changed

• "Extension" has at least four meanings. The MCP specification uses it for optional protocol add-ons. Claude uses "desktop extension" for a local MCP server in the desktop app. Gemini CLI uses it for a whole package.
• "App" is used for more than one thing. In the MCP specification, MCP Apps is a protocol extension for interface elements. Anthropic's connector page says claude.ai supports remote connectors "including MCP Apps that display interactive content in the conversation." Other vendors use "app" for other things; we did not read them.
• Integrations became connectors in the documentation. Anthropic's announcement of 2025-05-01 introduced "Integrations," described as letting Claude work with remote MCP servers. The 2026 documentation we read uses "connectors" for connections to MCP servers. None of the pages we read states that one name replaced the other, so we only report that both names exist.
• Commands merged into skills in Claude Code. The Claude Code skills page says custom commands have been merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way (read 2026-10-03). Plugin pages still list "commands" as a plugin component, so the word remains in use.
• One skill folder, several products. The VS Code page lists .claude/skills among the project skill locations it reads, next to .github/skills and .agents/skills. A skill written for one product may therefore be read by another. The agentskills.io page says many AI tools support the format and points to a Client Showcase, but we did not read that showcase or test any of those products.
• "Skills" appears in two layers. The MCP specification lists "Skills over MCP" as an extension for delivering instructions through MCP. Anthropic's skills are files loaded from folders or uploads. These are different mechanisms that share a word.

What the sources say: how Anthropic describes the relationship

Anthropic's page on connectors, skills and plugins states the split plainly: MCP connectors give Claude access to a tool or data source, skills teach it how to do a task the way you or your team does it, and plugins package skills, connectors, commands and agents into one unit. You add all three from the Customize page in claude.ai or the Claude desktop app.

Its skills overview adds projects to the comparison. Projects hold background knowledge that is always loaded in that project's chats. Skills are task-specific procedures that load when a request matches.

What to consider: matching a job to a term

This is our reading of the sources above, not a vendor statement and not a ranking.

If you want to	The term the docs use	Where to read more
Let an AI app read from or act in another service	MCP server, shown as a connector in Claude	What is an MCP server
Have the assistant follow your procedure for a task	Skill	Claude Skills explained
Install several of those together and share them	Plugin	Claude Code profile

An assistant that plans and acts across several steps is a different topic. See AI agents vs AI automation.

Each option has its own risk profile. Connectors reach outside services with the account you connect. Skills can carry scripts. Plugins can carry both and, in Claude Code, hooks that run commands. Read are MCP servers safe and are Claude Skills safe before you install anything.

Limitations, and when this map is not enough

• It covers only vendors whose documentation we read. Microsoft's and Google's consumer assistants, and everything from OpenAI, are not covered.
• Documentation changes quickly. Menu names already differ between pages. Anthropic's platform documentation says custom skills are added under Settings > Features, while the claude.com skills page says Customize > Skills. Check the live interface.
• A shared word does not mean shared behavior. Do not assume a "plugin" or "extension" from one vendor works in another.
• If you need to choose between specific products, this page will not do it. Read the vendor's own docs.