What is the difference between Agent Plugins, MCP, and Agent Skills?

They are three layers. MCP (Model Context Protocol) is the connection layer that lets an agent reach live tools and data. Agent Skills are the expertise layer — reusable instructions and workflows for a task. Agent Plugins is the packaging layer that bundles Skills and MCP configs into one portable package so an extension runs across many clients.

Does Agent Plugins replace MCP?

No. Agent Plugins wraps MCP; it does not replace it. MCP remains the standard for connecting an agent to external tools and data. Agent Plugins standardizes where the MCP config and Agent Skills live so the whole extension is portable across clients like ChatGPT, Cursor, Copilot, and Codex.

Which layer do I need to build an agent capability?

Usually all three. Use MCP to connect to the tool or data source, write an Agent Skill to encode the task expertise, and ship them together as an Agent Plugin so the capability is portable. You can use MCP or Skills alone, but the plugin package is what makes it 'build once, run anywhere.'

Quick Answer

Agent Plugins vs MCP vs Agent Skills Explained (2026)

Published: August 8, 2026

The Short Answer

They’re three layers of the same stack, not competitors. MCP connects an agent to live tools and data. Agent Skills package task expertise. Agent Plugins (launched August 6, 2026) bundle both into one portable package so a capability runs across many AI clients. MCP is the plumbing, Skills are the know-how, Plugins are the box you ship them in.

The Three Layers

Layer	What it is	Analogy
MCP (Model Context Protocol)	Open standard (from Anthropic) that lets an agent connect to external data and live tools	The wiring / API sockets
Agent Skills	Reusable packages of expertise — instructions, workflows, context for a task	The know-how / playbook
Agent Plugins	Portable package (with `plugin.json` manifest) bundling Skills + MCP configs	The shipping box that holds both

How They Fit Together

MCP gives the agent hands — it can reach your database, a live API, or a file system.
Agent Skills give the agent a brain for a task — the steps, context, and workflow to actually do the job well.
Agent Plugins put both in one standardized folder so the same extension works in ChatGPT, Cursor, GitHub Copilot, OpenAI Codex, VS Code, and Kiro — “build once, run anywhere.”

Why the Distinction Matters

Before Agent Plugins, every client wanted a different folder layout, so a capability built for one tool had to be re-packaged for the next. The plugin standard fixes where the MCP config and Skills live and how a client maps them into its runtime — so distribution stops being a per-tool chore.

What Agent Plugins Does Not Do

It’s a low-level packaging standard. It does not define installation, security, permissions, distribution, or UX — those stay with each client. It’s also distinct from OpenAI’s Plugin Directory, which is about distributing workflow packages, not the package format itself.

Which One Do You Actually Build?

Usually all three together: connect with MCP, encode the task in an Agent Skill, and ship them as an Agent Plugin. You can use MCP or Skills standalone, but the plugin package is what makes the capability portable across the whole ecosystem.

Verdict

Stop thinking “vs.” Think stack: MCP = connection, Skills = expertise, Plugins = portable packaging. Together they’re the 2026 blueprint for reusable agent capabilities.

Sources

The Next Web — OpenAI’s Agent Plugins open standard (Skills + MCP): thenextweb.com
9to5Mac — OpenAI shares new Agent Plugins standard (Aug 6, 2026): 9to5mac.com
AAIF — From skills and tools to portable agent plugins: aaif.io