How to connect MCP servers to Claude, with a deck tool as the example

Claude MCP means connecting Claude to external tools through the Model Context Protocol. In claude.ai and Claude Desktop you add a remote server as a custom connector; in Claude Code you run claude mcp add. After a one-time sign-in, Claude can call that server's tools in any conversation, with no plugin or glue code.

I build heydecks, which runs an MCP server that Claude uses to build slide decks, so I set up Claude MCP connections more often than most people. This post covers how it works, the exact setup in each Claude app, and a real run from a prompt to a published deck and a .pptx, using heydecks as the example server.

#What is Claude MCP, and how does it work?

MCP is an open standard Anthropic released in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025, so it is no longer a single-vendor protocol. Claude is the host; the thing it connects to is an MCP server. A server advertises tools, each with a name, a description, and an input schema. Claude reads that list, decides when a tool fits your request, calls it with arguments, and reasons over the result. Credentials stay with the server, not in the chat.

Servers come in two shapes. Local servers run on your machine and talk over stdio, which is how file-system and many community servers work. Remote servers run over HTTP at a URL, which is how hosted services connect. The shape decides where you can use a server: remote servers work in claude.ai, Claude Desktop, and Claude Code, while local ones need Claude Desktop or Claude Code on the same machine. For the protocol details, see what MCP is.

#How do you add an MCP server to Claude?

It depends on the app. Here is each one, using the heydecks server at https://heydecks.com/mcp as the example.

claude.ai and Claude Desktop. Open Customize, then Connectors, select "+", choose "Add custom connector," paste the server URL, and select "Add." The first time Claude uses it, a browser window asks you to sign in to the service. You can switch connectors on or off per conversation from the "+" menu in the chat box.

Claude Code. One command, then a browser sign-in on first use (OAuth 2.1 with PKCE, so no token lands in your shell history):

claude mcp add --transport http heydecks https://heydecks.com/mcp

For a headless box or CI where a browser is awkward, mint an mcp_ token in the heydecks dashboard and pass it as a header:

claude mcp add --transport http heydecks https://heydecks.com/mcp \
  --header "Authorization: Bearer mcp_your_token"

Claude Desktop with a config file. If you prefer config, or need token auth, Claude Desktop's claude_desktop_config.json only accepts command-based servers, so a remote server is bridged through mcp-remote (it needs Node.js installed). Open Settings, Developer, Edit Config:

{
  "mcpServers": {
    "heydecks": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://heydecks.com/mcp"]
    }
  }
}

On macOS, if Claude Desktop reports that npx is not found, put the absolute path from which npx in command. If you use several hosts, npx heydecks install wires up Claude Desktop, Claude Code, and Cursor in one step.

#How do you check the MCP server is connected?

In Claude Code, list your servers:

claude mcp list

You should see heydecks in the list, and /mcp inside a session shows its status and lets you re-run the sign-in. In claude.ai and Claude Desktop, the connector appears under Customize, Connectors and in the chat's "+" menu. Then send a small request that only reads, such as "list my heydecks brands," before you ask for anything that writes or spends credits.

#What can you build with MCP servers in Claude?

Roughly anything with an API. People connect Claude to GitHub for pull requests, Supabase or Postgres for live queries, Sentry for production errors, Playwright to drive a browser, and Context7 for current library docs. I keep a shorter, opinionated list of the best MCP servers.

Most servers fetch or change data. The category I care about produces a finished artifact. A deck is a good test case, because Claude is good at structuring an argument and writing the words, and bad at fighting a slide editor's layout. Hand the words to Claude and the rendering and brand to a server, and you skip the part everyone hates.

#What tools does the heydecks MCP server give Claude?

heydecks is the AI slide creator for AI agents. Its MCP server is a structured deck store, not a second writer: Claude designs the deck and writes every word, and the tools persist exactly what it sends. The tools you will see Claude call most:

There are more for versions, reordering, images, and brand edits; the full list is in the MCP docs.

#What does an end-to-end deck build look like?

You ask in plain language and watch the tool calls:

"List my slide templates, then create a deck called 'Q3 review' with a title, three stats, a process flow, and a CTA. Publish it and give me the PowerPoint."

Claude calls list_slide_templates to learn the fields, writes the content, builds the deck with create_deck and add_slides, publishes it, and calls export_deck for a .pptx link it pastes into the chat. The PowerPoint has real text boxes, not images of slides, so your team can keep editing. Point Claude at a company URL first and it can call extract_brand_from_url and create_brand to put the deck on that brand before any slide is written.

Here is a quarterly review deck of the kind that prompt produces, on a sample brand:

#Is Claude MCP free, and what does it cost?

The protocol is free and open source. You pay for Claude and for whatever each server charges. Custom connectors work on every Claude plan, with Free limited to one; Claude Code needs Pro or higher. The heydecks meter is the same on every channel: a deck costs 100 credits, reading and editing are free, the Free plan includes 200 one-time credits with watermarked exports, and editable PPTX export starts on Solo at $19 a month.

#Skills, MCP, or the REST API: which should you use?

They do different jobs and work together. A Skill teaches Claude a procedure, like a house style or a checklist, and can call MCP tools as part of it. MCP gives Claude a live connection to an outside system. The REST API is for code that needs a result without a model in the loop.

For decks: use MCP when Claude should design the deck in the conversation, which covers most agent work, and see building decks with Claude for prompts. Use the REST API when a script needs a deck from one call, since POST /v1/generate writes the deck from your brief. If what you actually want is any PowerPoint out of Claude, Claude PowerPoint options compares all five routes, MCP included.

To try it with a real server, the heydecks MCP server page has the connector URL and every tool Claude gets.

#Frequently asked questions

#How do I add an MCP server to Claude?

In claude.ai or Claude Desktop, open Customize, Connectors, select "+", then "Add custom connector," and paste the server URL. In Claude Code, run claude mcp add --transport http <name> <url>. Sign in when the browser opens on first use.

#Is Claude MCP free to use?

The protocol is free. Custom connectors are available on every Claude plan, though Free accounts get one, and Claude Code needs a paid plan. Each server may charge for its own work; heydecks charges 100 credits per deck.

#What is the difference between Claude Code and MCP?

Claude Code is Anthropic's coding agent in your terminal; MCP is how it, and the other Claude apps, get new tools. You add a server once with claude mcp add, and Claude Code can call it from then on.

#Does Claude Desktop support remote MCP servers?

Yes. Add the URL as a custom connector, which handles OAuth sign-in. The config file only accepts command-based servers, so for token auth you bridge the remote URL through mcp-remote.

#Can Claude make an editable PowerPoint over MCP?

Yes. With heydecks connected, Claude builds and publishes the deck, then export_deck returns a download link for a native .pptx with real, editable slides, locked to your brand. PPTX export needs the Solo plan or higher; PDF works on every plan.