An open standard Anthropic introduced in November 2024 lets AI agents use outside services. The Model Context Protocol (MCP): clients, servers, actions.
MCP stands for Model Context Protocol. It is an open protocol, introduced by Anthropic in November 2024, that lets AI agents connect to external services and data in a standard way. Instead of building a custom link for every pair of agent and service, developers follow one shared set of rules.
Do I need to be a developer to use MCP?
Not to understand it, and often not to use it. Connecting a server usually means pasting an address and a key into the settings of your AI agent, or running one command. Building a server is a developer task, but using one is closer to installing an extension.
What is an MCP server, and how is it different from an API?
An MCP server is a program that tells an AI agent what a service can do. An API is a way for software to call another service. An MCP server describes what a service can do in a form an AI agent can discover and use on its own: a list of actions, what each one needs and what it returns. Many MCP servers sit on top of an existing API.
Is it safe to let an AI agent use my social media tools through MCP?
It depends on how the server is built and how you set it up. Look for scoped access, keys you can revoke, confirmation for risky actions and a way to keep work as drafts. In Clautia, new posts are saved as drafts, and publishing now or cancelling asks for a confirmation token.
Which AI agents can connect to Clautia's MCP server?
The documented setups today are Claude Code, Cursor and Codex. Any other MCP-compatible agent that can send an Authorization header with an API key may work too, but check that your agent supports it. Today Clautia connects with an API key in the Authorization header. OAuth (one-click connection) is not available yet.
Does MCP replace APIs?
No. MCP builds on top of the idea of an API and makes it easier for AI agents to use one. Clautia offers both an MCP server and a REST API, and both use the same kind of API key, so developers and AI agents can work side by side.
MCP, short for Model Context Protocol, is an open standard that lets an AI agent (Claude Code, Codex, Cursor…) connect to outside services and data through one shared set of rules. This guide from Clautia, the agentic social media manager, explains the idea in plain English for marketers and small teams. Anthropic introduced the protocol in November 2024, and at the end of the guide you will see how it works in a real tool: Clautia's own MCP server.
What is MCP, in one minute
An AI agent on its own can write and reason, but it can't see your calendar, open your files or schedule a post. To do those things, it needs a way to reach other software. Before MCP, every connection like that was built one at a time: one integration for this agent and this service, another for that agent and the same service.
MCP replaces that patchwork with a common language. A service describes itself once, in a standard way. Any AI agent that speaks the same language can then discover what the service offers and use it.
A useful comparison is the USB port. You don't need a different cable for every device you plug in, because the connector is standard. MCP aims to do the same for the connection between AI agents and the services they work with. Like any comparison, it simplifies, but it gets the main idea across.
The three pieces: client (your agent), server and actions
Three words cover most of what you will read about MCP.
The client is the AI agent you talk to. It could be a coding agent, a chat or any program that can act on your instructions. The client decides when it needs an outside service and sends the request.
The server is the program that offers the outside actions or data. It sits in front of a service, such as a calendar, a database or a social media scheduler, and answers the client's requests.
The actions are the specific operations the server makes available to the agent. Each one has a name, a description, the information it needs and the result it returns. "List my brands" and "schedule this draft" are examples. The specification calls them tools; here we call them actions.
MCP also allows servers to share other things, such as reference data the agent can read. For most marketing work, actions are the part you notice.
Here is the flow in practice:
You type a request into your AI agent.
The agent looks at the list of actions its connected servers offer.
It picks the ones that fit and calls them, passing the details they need.
The server does the work in the real service and sends back the result.
The agent uses that result to answer you or to take the next step.
You stay in the conversation. The protocol handles the plumbing.
MCP, an API and a plugin: how they differ
People mix these up, and for good reason. They overlap. This table separates them by what each one is for.
Concept
Who uses it directly
What it does
Typical example
API
Developers and their code
Lets one piece of software call another service with fixed requests
A program that creates a post through a URL
MCP server
AI agents
Describes a service's actions so an AI agent can discover and use them
An agent that lists your brands and drafts
Plugin or app
People, inside one product
Adds features to a specific application, usually built for that one app
An extension for a single design or chat product
An MCP server often sits on top of an API. The API is how the service works under the hood, and the MCP server is the layer that makes it readable for an AI agent. The two don't compete.
Why MCP matters for social media work
Social media is a field made of small, repetitive steps that cross several tools: you gather ideas, write text, adapt it for each network, pick a time, schedule and check the result. AI agents already help with the writing. The slow part is moving everything between the agent and the place where posts actually live. We cover the practical side in social media automation.
MCP changes what an agent can do inside that process:
Fewer copy-and-paste steps. Instead of taking text out of a chat window and pasting it into a scheduler, you ask the agent to create the draft in the scheduler.
Work with real context. The agent can read which brands and accounts exist, and what is already planned, before it writes anything. Drafts then fit your calendar instead of ignoring it.
One request, several steps. "Plan next week's Instagram posts for the bakery" (a prompt in plain words) can become a series of actions in your tools, not just a block of text.
Your tools, your choice of agent. Because the standard is shared, you are not tied to a single AI agent for a single scheduler.
There is a trade-off, and it's worth saying plainly. An agent that can act is more useful and also needs more care than one that only talks. That is why the details of a server matter: what the agent is allowed to do, what needs your confirmation, and what stays as a draft.
What to check before you connect an MCP server
Whatever service you are about to connect, run through this short list.
Who can do what. Does the access key give full control, or is it scoped? Can you revoke it?
Drafts first. Can the agent prepare work without publishing it?
Confirmation for risky steps. Is there a pause before something irreversible, such as publishing now or deleting?
Limits. Are there daily caps per key, so a runaway loop can't flood your accounts?
Data handling. Does the server treat content stored in your account as data, rather than as instructions the agent should obey?
Documentation. Are the actions listed, with what each one does?
These questions apply to any MCP server, not only to a social media one.
How Clautia's MCP server works today
Clautia has an MCP server, so an AI agent can work with your brands and posts in the same system as the app. This is what exists today, based on Clautia's MCP documentation.
Connection. The server lives at one address and uses a standard web transport (Streamable HTTP). You authenticate with an API key, sent as a Bearer token in the Authorization header. API keys are available on every plan, with 2 active keys on Creator, 5 on Pro and 20 on Agency.
Agents. The documented setups today are Claude Code, Cursor and Codex (how to choose between them, and how to add a server to Cursor). Other MCP-compatible agents that let you set an Authorization header may work as well. Connectors with one-click sign-in, like the planned Perplexity one, are not available yet, and the command-line tool is listed as coming soon on the Roadmap.
Actions. The server makes these actions available to the agent:
Action
What it does
Kind
brands_list, accounts_list
Lists your brands and their connected accounts
Read
posts_list, posts_get
Reads your posts
Read
posts_create
Saves a new post as a draft
Write
posts_schedule
Schedules a draft for a date and time
Write, with a safeguard
ai_draft, ai_adapt
Drafts a post with AI or adapts it to a network
Write, uses AI credits
posts_publish, posts_cancel
Publishes now or cancels
Delicate, needs a token
Safeguards. Four rules shape what the agent can do:
A new post is always a draft. Nothing goes out until it is scheduled or published.
Scheduling less than 10 minutes ahead asks for confirmation.
Publishing now and cancelling first return a confirmation token, and the call must be repeated with it.
Action responses are labelled as untrusted data, so content stored in your workspace is read as data and not as instructions.
One honest caveat: the confirmation token goes back to the same agent that asked for it. It prevents accidental calls, but it isn't a human review. If you want to approve every action yourself, tell your agent to stop and ask you before it confirms anything. One-click approval for agent work will come with Agents and Routines (automations), coming soon.
Limits. Each key has a daily budget (per UTC day), so one key can't run away with your account. A Creator key can publish up to 50 posts a day, a Pro key 200 and an Agency key 1,000, and the same caps apply to the workspace as a whole. Each key can also spend up to 50, 200 or 1,000 AI credits a day from the agent, by plan.
A first request to try once it is connected
Once your agent is connected, start with something that is easy to check and easy to undo. For example:
"For the brand Driftwood Books, list the accounts we have connected. Then write two short drafts for Saturday's local authors table, one for Instagram and one for Facebook. Leave both as drafts and tell me which account each one is for."
The agent lists brands and accounts, drafts the text with the AI functions and creates two drafts. You open the calendar, edit what you want and decide when to schedule them. The agent prepared the work and you kept the decision. Once you trust the flow, you can ask for a whole week.