Claude MCP is how Claude connects to tools outside the chat. MCP stands for Model Context Protocol, an open standard that lets an AI assistant call external servers to read a website, pull a file, generate an image or create a share link. For marketers, it means Claude can go past "here is some ad copy" and hand back finished, on-brand ad images, sized for each placement and ready for a client to review.

This guide covers what MCP is in plain words, how to add a server to Claude Desktop or Claude Code, which kinds of MCP servers are useful for marketing work, and a complete example workflow you can copy.

What is Claude MCP, in plain words?

On its own, Claude only knows what is in the conversation and what it learned in training. It cannot open your analytics, read your brand guidelines folder or render an image file. MCP fixes that by giving Claude a standard way to talk to other software.

You connect an MCP server, and that server offers Claude a set of tools. A tool is a named action with clear inputs and outputs. For example, a tool called research_url might take a web address and return notes on the business, its audience and its offers. When you ask Claude for something, it decides which tools to call, shows you the calls in the chat, and uses the results to keep working.

Three terms are worth knowing:

Client - the app you chat in, such as Claude Desktop, Claude Code, Cursor, VS Code or ChatGPT.
Server - the outside service that exposes tools, usually at a URL with an API key.
Tool - one action the server can perform, like scanning a site or rendering an image.

The practical difference is big. Without MCP you copy a prompt from Claude into another app, download the result and paste it somewhere else. With MCP, Claude runs those steps itself and you stay in one conversation.

How to add an MCP server to Claude

Every server has its own setup details, so always follow the server's own instructions rather than a generic snippet. The overall pattern is the same everywhere: get an API key from the service, add a small block of configuration to your client, restart, and test.

Claude Desktop

In Claude Desktop, go to Settings, then Developer, then Edit Config. That opens a JSON config file where you add the server's block, including its URL and your key. Some servers can instead be added as a connector under Settings, then Connectors, if they support signing in. Quit the app fully and reopen it, because the config is only read on a cold start. Our step-by-step guide for connecting an ad server to Claude Desktop has the exact block to paste.

Claude Code

Claude Code, the command-line version of Claude, adds MCP servers from the terminal or a config file. You give the server a name, its URL and an authorization header with your key, and the tools appear in your next session. The Claude Code setup guide walks through it without guesswork.

Other clients

MCP is not limited to Claude. Cursor, ChatGPT, VS Code and Windsurf can all connect to the same servers, so a team can use different apps with the same brand assets. The MCP overview page lists setup for VS Code, Windsurf and any other MCP-capable client.

Once connected, test with a harmless read-only request first, such as "List my brands" or "What is my credit balance?" If Claude answers using the server's tools, you are set up.

What marketing work MCP servers make possible

Marketing is full of hand-offs: research to strategy, strategy to copy, copy to design, design to client review. MCP lets one conversation cover several of those steps. Here is how the jobs map to real tools, using the tool names from Claivra, an ad-creative MCP server:

Job	What the tool does	Example tools
Brand setup	Reads a website and saves logo, colours and voice as a reusable brand kit	scan_brand_from_url, create_brand
Research	Pulls the business model, audience, offers and selling points from a URL	research_url
Concepts	Writes ad angles, then full concepts with headline, image prompt and CTA	generate_idea_seeds, generate_concepts_from_url
Images	Renders finished ad images, variations, layout compositions and targeted edits	generate_image, generate_variations, compose_ad, edit_image
Video and audio	Makes short video ads, voiceovers, lip-sync and dubbed versions	generate_video, generate_voice, lipsync, dub
Client review	Bundles ads into a link the client can open, pick from and comment on	create_share, revise_share
Reporting	Shows credit use per client, so agencies can bill accurately	get_client_usage

Ad creative is only one category. Many teams also connect servers for documents and file storage, project management, analytics and team chat. Claude can combine them in one request: read the brief from a shared doc, make the ads, then post the review link to your team channel. The more of your stack is reachable, the fewer copy-paste steps are left.

Example workflow: from a URL to a client review link

Here is a realistic request for a fictional home-cleaning brand, Acme. You can type something close to this in one message:

"Research acme.com, write 3 concepts for Meta, render 4:5 and 1:1 with the brand logo, then make a review link."

Behind that single message, Claude works through a sequence like this:

Build the brand kit. It calls scan_brand_from_url on acme.com to pick up the logo, colours and tone of voice, then create_brand to save them for every later render.
Research the business. research_url returns who Acme sells to, what it offers and which claims the site actually makes.
Write the concepts. The concept tool returns three ideas, each with a headline, an image prompt and a call to action.
Render every size. generate_image renders each concept at 4:5 for the Facebook and Instagram feed and 1:1 for square placements, with the saved logo applied from the brand kit.
Fix details. If a headline wraps badly, edit_image changes that one element without starting over.
Share for review. create_share bundles the six images into one link. The client opens it without an account, picks favourites and leaves comments. After changes, revise_share updates the same link.

The concepts might come back looking like this:

Concept 1: Headline "A spotless kitchen before the kettle boils" - bright countertop scene, product in hand - CTA "Shop the starter kit".
Concept 2: Headline "Strong on grease. Gentle on you." - split image of a greasy pan and a clean one - CTA "See the range".
Concept 3: Headline "Refill, don't rebuy" - refill pouch next to a reused bottle - CTA "Start refilling".

You stay in control between steps. A good habit is to ask Claude to stop after the concepts so you can approve or rewrite headlines before any images are rendered.

Tips for better MCP prompts

Claude is good at choosing tools, but a clear request saves time and credits. These habits help:

Name the brand. If you manage several clients, say "use the Acme brand kit" so nothing renders under the wrong logo.
State the placement and sizes. "Meta feed, 4:5 and 1:1" beats "some social ads".
Give the offer and audience in one line. "Refill pouches for busy parents who hate harsh smells" gives the concepts a real angle.
Quote exact text. Put headlines, prices and CTAs in quotation marks so they render word for word.
Concepts first, images second. Approving ideas before rendering keeps wasted renders low.
Edit, don't regenerate. "Keep the composition, change only the headline to..." preserves what already works.
Let the brand kit supply the logo. Avoid describing the logo in words; descriptive text can end up drawn into the image instead of the real file being used.
Ask for a recap. End with "list what you rendered and the credits used" so you have a record.

Limits and cautions

MCP makes Claude more capable, which also means it can do more on your behalf. Keep these in mind:

Review every output. Check spelling, prices, dates, legal lines and any claim before an ad goes live. AI can misread a site or phrase a claim more strongly than the business can support.
Credits cost money. Image, video and voice tools spend credits on each call. Check your balance before a large batch, render one test image before twenty, and remember that video usually costs more than a still.
Keep client brand kits separate. Name kits clearly, confirm which one is active before a batch, and never reuse one client's assets for another.
Read tool approvals. Most clients ask before running a tool. Take a second to check what is about to happen, especially for anything that shares or deletes.
Protect API keys. Treat them like passwords. Do not paste them into chat, screenshots or shared config files.
Platform rules still apply. Ad platforms have their own policies on claims, text and targeting. An MCP server does not check those for you.

Try it with your own brand

If you want to see the workflow above run end to end, Claivra works inside Claude Desktop, Claude Code and the other clients mentioned here. The free plan includes 10 credits and 5 concept runs with no credit card, which is enough to scan a site, write concepts and render a few test ads. Paid plans and top-up packs are on the pricing page.

FAQ

What does MCP stand for in Claude?

Model Context Protocol. It is an open standard that lets Claude call tools on outside servers, such as research, image generation or file storage.

Is Claude MCP free to use?

Adding an MCP server to Claude is free. The server itself may charge for usage, for example credits for each image or video it generates.

Do I need to code to use MCP servers?

No. Setup usually means pasting a short config block and an API key, then restarting the app. After that you work in plain language.

Can Claude make ad images through MCP?

Yes, if you connect a server that offers image tools. Claude writes the concept, calls the render tool and returns the finished image in the chat.

Which clients support MCP besides Claude?

Cursor, ChatGPT, VS Code and Windsurf all support MCP servers, along with any other MCP-capable client.