Agent Plugins does not replace MCP. It packages it, and your stack should do nothing
OpenAI announced Agent Plugins on 6 August 2026 with AWS, Cursor, GitHub, VS Code and Vercel. We read the v1.0.0 spec instead of the press release. It is a folder format for two standards Anthropic already wrote, and for a 5 to 50 person ops team there is nothing to action today.
By Ishan Vats · Founder of IV Consulting, builds AI agents and automations for 150+ teams
Procedural knowledgeskills/SKILL.md
Loads itChatGPT, Codex
Loads itCursor
No, a small ops team does not need to do anything about Agent Plugins yet. Agent Plugins is a packaging layer, not a new capability: the v1.0.0 spec defines a folder with a plugin.json manifest, Agent Skills in skills subdirectories, and MCP servers described in a single mcp.json. It does not replace MCP, it bundles it. That makes it a distribution convenience for people who ship agent extensions, not a decision for people who use them. The one signal that changes this: when the tools your team already pays for start shipping installable plugins, it stops being a developer tooling story and becomes an ops one.
The verdict
Do Agent Plugins change anything for your ops stack?
No. Agent Plugins does not change anything for a 5 to 50 person ops stack, and we would push back on anyone telling you otherwise this month. Agent Plugins is a real standard with serious backing, and it is also a packaging format. Packaging formats matter enormously to the people doing the packaging and almost not at all to the people installing the result.
So rather than restate the announcement, we read the specification and the v1.0.0 spec repository. Three things came out of that:
- It wraps, it does not replace. A plugin is a directory with a required plugin.json manifest, optional Agent Skills in immediate subdirectories of skills, and MCP servers described in one mcp.json at the plugin root. Both things inside the wrapper already existed. Nothing you have wired stops working.
- Version 1.0.0 is still a Working Draft. The specification carries that status rather than being a frozen final standard, which is the single best reason not to rebuild anything around it yet. A format that can still move is not a format to couple your operations to.
- The problem it solves is a developer problem. Before this, an extension had to be packaged separately for each client. Now one folder loads in ChatGPT and Codex, Cursor, GitHub Copilot, Kiro and VS Code. That is a distribution win for people publishing extensions, not a capability win for people running a business.
- It packages two standards Anthropic wrote. MCP came from Anthropic and was donated to the Agentic AI Foundation, under the Linux Foundation, on 9 December 2025. Anthropic then published Agent Skills as an open standard on 18 December 2025. Anthropic is not among the announced co-developers or the initial Technical Steering Committee, which lists core maintainers from Amazon, Cursor, Microsoft, OpenAI and Vercel.
The honest counter case, because it is real: standards that get adopted early tend to get adopted quietly, and the teams that ignored MCP in its first six months did end up paying a catch up cost. The difference is that MCP unlocked something an agent could not previously do. Agent Plugins makes something that already worked easier to hand to someone else. Those are not the same kind of announcement, and only the first kind is worth interrupting a roadmap for.
The standard
What are Agent Plugins?
Agent Plugins is an open standard for packaging the things that extend an AI agent into one distributable folder. It was announced on 6 August 2026 by OpenAI, co-developed with AWS, Cursor, GitHub, VS Code and Vercel, and published at version 1.0.0. Note the status on that page: 1.0.0 is a Working Draft, not a frozen final standard. Beyond the six co-developers, AWS and Google have both published on it from their own developer blogs.
The spec is deliberately small, which is the most useful thing about it. A plugin is a directory. At the root sits plugin.json, and only two fields in it are required: the schema reference and a name. Everything else is optional.
Inside that directory, two things can live. Agent Skills go in immediate subdirectories of skills/, each containing a SKILL.md. There is no recursive search, so the layout is fixed rather than discovered. MCP servers are described in a single mcp.json at the plugin root, covering stdio, Streamable HTTP and the legacy HTTP plus SSE transports.
That is the entire idea. The smallest useful plugin is a folder, a manifest, and one skill file. If that sounds less dramatic than the coverage suggested, that is because it is, and the restraint is the point: a small spec is one that clients can actually implement, which is why six of them shipped support at launch.
| The three layers | Agent Plugins | MCP | Agent Skills |
|---|---|---|---|
| What layer it is | Packaging and distribution | Connectivity and transport | Procedural knowledge |
| What it actually does | Bundles skills and MCP configs into one installable folder | Connects an agent to systems outside the conversation | Tells an agent how to do a specific job in your context |
| Who created it | OpenAI with AWS, Cursor, GitHub, VS Code, Vercel | Anthropic | Anthropic |
| Key date | Announced 6 Aug 2026, v1.0.0 | Donated to the Linux Foundation 9 Dec 2025 | Published as an open standard 18 Dec 2025 |
| Where it lives | plugin.json at the folder root | mcp.json at the plugin root | skills/<name>/SKILL.md |
| Governance | Steering committee: Amazon, Cursor, Microsoft, OpenAI, Vercel | Agentic AI Foundation, under the Linux Foundation | Published openly at agentskills.io |
| Who it is actually for | People distributing agent extensions | Anyone connecting an agent to their own tools | Anyone encoding a process they repeat |
| Our call | Ignore until your vendors ship plugins | This is the layer that matters | Worth writing once a process repeats |
The layer model
Agent Plugins vs MCP: what is the actual difference?
The difference between Agent Plugins and MCP is that MCP decides what an agent can reach, and Agent Plugins decides how you hand that setup to someone else. They operate at different levels, so comparing them as rivals produces a question with no answer.
The clearest way to hold all three in your head is as a stack. MCP is the plumbing. It is the protocol that lets an agent talk to your CRM, your database, your file store, your ticketing system. Without it, the agent can only reason about what you paste into it. We wrote the plain English version of this in MCP for small businesses explained.
Agent Skills, what Anthropic shipped as Claude Skills and then opened as a standard, are the instructions. A skill is essentially a markdown file that tells the agent how a particular job gets done in your business: how you qualify a lead, how you close a month, what your escalation rules are. It is procedural knowledge, and it is the layer most small teams underuse. That one is covered in Claude Skills for business.
Agent Plugins is the box. It takes a set of skills and a set of MCP server configurations and makes them one thing you can install, version and share. Before the standard, a team publishing an extension had to produce a different layout for each client. Now they produce one.
Which brings us to the question people are actually typing.
Do Agent Plugins replace MCP?
No. Agent Plugins do not replace MCP, they contain it. The spec describes MCP servers in an mcp.json file inside the plugin, so a plugin without MCP is a plugin that has nothing to connect to. MCP is a dependency of the format, not a competitor to it.
This matters because the announcement generated a wave of coverage phrased as a succession, and succession is the wrong mental model. MCP had more than 10,000 active public servers when Anthropic donated it to the Agentic AI Foundation on 9 December 2025. Standards with that much deployed surface area do not get replaced by a packaging convention. They get wrapped by one.
The ops question
Does a small ops team need to do anything?
A 5 to 50 person ops team does not need to do anything about Agent Plugins today. Here is the test we would apply, and it takes about a minute.
Ask whether anyone in your business is publishing an agent extension for other people to install. If the answer is no, which it is for almost every team we work with, the standard has no surface that touches you. You are a consumer of agent capability, and consumers experience a packaging standard as the absence of friction they were never conscious of in the first place.
The second test: are your agents currently able to reach the systems they need to reach? If yes, you already have the layer that does the work. If no, the fix is an MCP server or a proper integration, and it was the fix last week too. A new folder format does not add a connection you have not built.
What we would genuinely avoid is the reflex this kind of news produces, which is re-plumbing a working setup because a standard was announced. We have watched teams burn a fortnight migrating an automation stack toward whatever was announced that month, and the cost is never the fortnight. It is that the automation stopped being trusted while it was in pieces.
The trigger
When does this become your problem?
A do nothing verdict on Agent Plugins is only honest if it comes with the conditions that would change it. Here are the three, in the order we expect them to arrive.
Your own vendors ship plugins
The real signal. When the tools you already pay for start publishing installable plugins, the standard stops being developer tooling and becomes a procurement question: what does this plugin get access to, and who approved it. That is the moment to have an answer ready.
You start installing third party plugins
A plugin bundles MCP server configurations, and an MCP server is a connection to your systems. One folder that installs in one click is convenient in exactly the way that deserves a review step. Treat plugin installs like you treat browser extensions, not like you treat a font.
You are distributing to your own clients
If you are an agency handing the same agent setup to client after client, this is the one case where the standard is straightforwardly good news for you today. One package that installs across clients replaces a setup document nobody follows correctly.
None of those three is a reason to act this week. All of them are reasons to know what the format is, which is the actual job this post is doing. Our position gets revisited when the first of them lands, and we would rather revise a verdict in public than pretend we did not have one.
FAQ
Questions teams are asking about Agent Plugins
What are Agent Plugins?
Do Agent Plugins replace MCP?
What is the difference between Agent Plugins, MCP and Agent Skills?
Does a small business need to do anything about Agent Plugins?
Which AI tools support Agent Plugins?
Is Agent Plugins the same as ChatGPT plugins?
What version is Agent Plugins and is the standard final?
Is installing an Agent Plugin a security risk?
Ishan Vats
Founder, IV Consulting · AI and automation consultant
I build Notion and ClickUp operating systems, automations, and production AI agents for teams from startup to enterprise. 150+ ops transformations over 10+ years. We had already published our explainers on both MCP and Agent Skills before this standard packaged them, which is why this verdict tells most teams to do nothing rather than to migrate.
Run my own agent numbers →Keep reading
Related guides and work
MCP for small businesses explained
The connectivity layer underneath all of this, in plain English. Read this one first.
Read the explainer →
Claude Skills for business
The procedural knowledge layer a plugin packages, and how to write your first one.
Read the guide →
Zapier Agents vs Make vs n8n
The decision that actually moves your bill this quarter, unlike this one.
See the comparison →Want to know which agent standard actually affects you?
Book a free 30-minute strategy call. We will look at what your agents can currently reach, tell you honestly whether anything in this announcement touches your stack, and point at the one change that would move your numbers. If we are not the right team for you, we will say so and point you somewhere better.
Book a Free Strategy Call →Free 30-minute call. Honest take, even if that means "you do not need us yet."