Agent Plugins: what the new standard means for social media teams
The open plugin format from Vercel, OpenAI, Microsoft and friends, translated for social media teams — and how to run the MCP half of it today.
William Lindberg · Customer Success
On August 6, 2026, Vercel and a group of companies that rarely agree on anything — AWS, GitHub, Microsoft, OpenAI and Anysphere — announced Agent Plugins, a shared standard for adding extras to AI assistants. At launch it works across ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code. The announcement is written for developers, but it matters for anyone who wants an AI assistant that can actually do work with them — including your social media. Here's what it is in plain language, and how you'd actually use it.
What was announced
An agent plugin is a small package you install into your AI assistant to give it two things: knowledge about how you work, and the ability to use your tools. Think of it like a browser extension, but for ChatGPT or Cursor instead of Chrome.
The problem it solves is simple. Until now, every AI app had its own way of adding extensions. Something built for ChatGPT didn't work in Copilot. Something built for Cursor didn't work anywhere else. If a tool you relied on supported your assistant, you got lucky. From now on, a company can build one plugin and it works in every app that supports the format — and six large companies just agreed to support it.
One name is missing from that list: Anthropic. That's less strange than it sounds. The two building blocks inside a plugin — skills and MCP — both started in the Claude ecosystem, and Claude Code has had its own plugin format for a while. So this is really the rest of the industry adopting the same pieces Claude users already work with. Wherever you sit, the building blocks are now the same everywhere.
The two things inside a plugin
A plugin holds two kinds of things, and the difference between them is worth two minutes of your time:
| A skill | An MCP server | |
|---|---|---|
| What it is | A document that teaches the assistant how you work | A connection that lets it use a real tool |
| What it gives the assistant | Knowledge: your voice, your rules, your process | Ability: draft, schedule, publish, report |
| Social media example | Your brand voice and posting rules | "Schedule this carousel for Tuesday at 9" |
An easy way to remember it: when a freelancer starts working on your social media, you hand them two things — your brand guidelines and a login to your tools. Skills are the guidelines. MCP is the login.
The two need each other. A skill without an MCP server writes good drafts it can't publish. An MCP server without a skill publishes things that don't sound like you. The clever part of the plugin format is that both travel in one box.
What this looks like for your social media
Picture a plugin your team installs once, into whichever assistant each person prefers. Inside it:
- A voice skill: your tone, phrases you never use, your hashtag policy, and how your LinkedIn voice differs from your TikTok voice
- A process skill: how far ahead you plan, who approves what, and what a launch week looks like
- Your scheduler's connection: the MCP server that turns an approved draft into a scheduled post
The first two are documents your team writes once and updates when things change. The third comes from your scheduler. With those three in place, your week starts to look different:
- Monday planning: "Look at what's scheduled this week and suggest posts for the empty days." The assistant checks your real calendar, sees Tuesday and Thursday are empty, and drafts something in your voice.
- Getting more out of what you make: "Turn this blog post into a LinkedIn post, an Instagram caption and a TikTok script." One idea becomes a week of content, each version shaped for its platform.
- Quick checks from anywhere: "What's going out today?" or "Move Friday's post to next Monday." No opening the app, no clicking through a calendar.
- Simple reporting: "Which of last month's posts did best, and what do they have in common?" The assistant pulls the numbers and gives you an answer instead of a dashboard.
None of this replaces your calendar or your approval process. The assistant is a faster way to operate them.
You don't have to wait
Here's the practical part: the connection half of this already works today. Plugins make it easier to package and share, but MCP itself is live now. Lemonpost ships an MCP server — connect Claude, ChatGPT, Codex or Cursor, and your assistant can draft, schedule and check your calendar using the same operations the app itself uses.
We're not guessing that this works. Lemonpost's own social accounts are run this way, at roughly 500,000 views and counting, with most posts drafted and scheduled through Claude talking to Lemonpost's MCP server. It's how we found out, for example, that the planning conversations matter more than the drafting ones — "what should this week look like" is where the assistant saves the most time.
If you want to try the same setup, the order that works:
- Pick the assistant you already use. There's no wrong answer here — that's the whole point of a shared standard.
- Connect your scheduler. With Lemonpost, that means adding its MCP server to your assistant — a few minutes of setup, once.
- Write your voice down. One page: tone, three example posts you loved, three things you'd never say. Give it to the assistant at the start of a conversation. When plugins reach your assistant, this page becomes your skill file — nothing is wasted.
- Start with drafts only. Let the assistant write while you schedule. Trust builds fast, but let it build.
- Keep approvals on when you let it schedule. More on that below.
Letting an AI near your accounts sounds like losing control. It isn't, if the tool underneath has guardrails: with approval workflows turned on, nothing the assistant drafts goes live until a person signs it off — the same rule you'd apply to any new teammate.
What's still missing
This is a version 1.0, and it's honest to say what isn't there yet:
- There's no app store. The format exists; a place to browse and install plugins with one click mostly doesn't, yet. That's the moment this goes mainstream.
- Tools have to show up. The format only helps if schedulers, analytics tools and asset libraries actually package themselves as plugins.
- Some pieces stay app-specific. The standard deliberately covers only skills and MCP servers for now. More advanced extras still work differently in each app.
Our bet: within a year, "does it work with my AI assistant" becomes a normal question when choosing software, the way "does it have a mobile app" once was. Teams whose tools can talk to their assistant will delegate the busywork. Teams whose tools can't will keep copy-pasting between tabs.
FAQ
Do I need to be a developer to use agent plugins? No. Installing a plugin is designed to be a one-step action inside the assistant. Writing skill files is writing, not coding — if you can keep a brand guideline doc up to date, you can keep a skill up to date.
Does Lemonpost have an agent plugin? Lemonpost ships the MCP server half today, which is the part that lets an assistant actually schedule and publish. As the apps roll out plugin support, that's the piece that slots straight in — and you can already connect Claude, ChatGPT, Codex or Cursor directly.
Which AI assistant should I use for social media? Whichever you already use for everything else. The point of a shared standard is that this choice stops mattering — the same plugin will follow you if you switch.
Will AI assistants replace social media schedulers? No. The assistant is a way of operating your tools, not a replacement for them. Something still has to hold the calendar, run the approvals, retry a failed publish and collect the numbers. That's the scheduler, whoever — or whatever — is pressing its buttons.
Is it safe to let an assistant post for you? As safe as you set it up to be. Keep approval workflows on so every draft waits for a human yes, and expand what the assistant may do gradually — the way you would with any new hire.
