Claude Connectors vs MCP Servers: What's the Difference? A Plain Breakdown

You open Claude's settings, spot a menu called Connectors, and ten minutes later a tutorial tells you to "install an MCP server." Same thing? Different thing? The two terms get swapped constantly, and that is where the confusion starts.

Here is the short version. A connector is what Claude calls the feature: a switch in your settings that lets Claude reach an outside app such as your calendar, your files or your project tracker. An MCP server is the program sitting behind that switch, speaking the Model Context Protocol. One is the product label, the other is the machinery. This breakdown pulls them apart, shows where each one lives, and helps you pick the right one for your own setup.

The Short Answer

Connector describes how you use it. MCP server describes how it is built. Most connectors you click on inside Claude are MCP servers underneath, so asking "connector or MCP server?" is a bit like asking "app or code?" You are usually looking at the same thing from two different angles.

💡 Quick proof: connectors you switch on in claude.ai show up inside Claude Code as MCP servers, listed with a "claude.ai" prefix. Same plumbing, different label.

Here is the one-glance version:

Think of the connector as the wall socket you actually touch. The MCP server is the wiring behind the wall. You can live your whole life using sockets without ever opening a wall, but the day you need an outlet where none exists, you will be glad you know how the wiring works.

The Two Pieces Explained

Before any comparison, each piece deserves a plain definition.

Directory Connectors

Directory connectors are the ready-made ones listed in Claude's Connectors menu: Google Drive, Gmail, Google Calendar, Notion, Asana, Linear, Figma, Atlassian and many others. You click Connect, sign in to the other service, and Claude gets a defined list of actions it can perform there. The company behind the service (or Anthropic) maintains the server. You never see it, and you never need to.

This is the lowest-friction route by a wide margin. If your calendar is already on the list, the whole setup takes under a minute. It is the equivalent of a USB-C hub: one small box, several devices, no soldering.

Custom Connectors by URL

A custom connector is the same idea with the training wheels off. You point Claude at a remote MCP server that is not in the directory, usually one your team built. In claude.ai the steps look like this:

1. Open Settings and choose Connectors.
2. Click Add, then Add custom connector.
3. Paste the remote MCP server URL.
4. If the server needs a registered OAuth app, fill in the optional client ID and client secret fields.
5. Follow the sign-in screen the server shows you.

The word "custom" is the giveaway: you just added an MCP server, you just did not call it that. On Team and Enterprise plans, an owner typically adds the connector for the organization first, and members then sign in with their own accounts. Plan availability shifts over time, so check Anthropic's help page before you promise it to a team.

One detail catches people out. On the web and mobile apps, Claude reaches your server from Anthropic's cloud, so a server running on localhost is invisible to it. The URL has to be public and served over HTTPS.

The Protocol Underneath

MCP stands for Model Context Protocol, an open standard Anthropic introduced in November 2024 and later handed to the Agentic AI Foundation, a Linux Foundation project. Under the hood it is a set of JSON-RPC 2.0 messages. A client (Claude) asks a server what it can do, then calls those things when a conversation needs them.

The protocol is the shared language. That is the whole point: write a server once, and any client that speaks MCP can use it. Claude is one such client, but code editors, other assistants and home-grown agents can be clients too. Picture a conference interpreter: the speakers never need to study each other's language, because the booth in the middle handles it.

Tools, Resources, and Prompts

Every MCP server offers some mix of three building blocks:

Most servers you meet lean heavily on tools. Resources and prompts appear less often, but they matter when you want Claude to read a specific document or follow a repeatable routine.

Local or Remote

MCP servers come in two flavors, and the flavor decides where they can be used.

• Local servers run on your own machine as a child process and talk over standard input and output (stdio). They can reach your files, your local database and your scripts, and nothing is exposed to the internet.
• Remote servers run on a host somewhere, speak Streamable HTTP, sit behind an HTTPS URL and usually sign users in with OAuth. These are the ones that become connectors.

Side by Side

With both pieces defined, here is the full comparison:

Where they overlap. Directory connectors and custom connectors are both remote MCP servers. Every time you click Connect or paste a URL, MCP is doing the talking behind the scenes.

Where they split. A local stdio server is not a connector in the claude.ai sense: it lives in Claude Desktop's config file, in a desktop extension, or in Claude Code. And an MCP server can serve clients that have nothing to do with Claude. In that direction, the word "connector" stops applying altogether.

Where Each One Lives

Claude Apps

The web, desktop and mobile apps all show the Connectors menu. Claude Desktop does one extra trick: it can also run local servers. You either edit claude_desktop_config.json by hand or install a desktop extension, a packaged .mcpb file that bundles a local server so you skip the config editing.

Claude Code

Claude Code is the command-line tool, and it treats everything as an MCP server. Add a remote one with a single command:

claude mcp add --transport http shipping-rates https://example.com/mcp

Copy Prompt

Add a local one by passing the command that starts it:

claude mcp add shipping-rates -- python /path/to/server.py

Copy Prompt

Servers can be saved at a local, project or user scope. Project scope writes a .mcp.json file that teammates share through git, which makes it the easiest way to give a whole team the same tools.

The Claude API

Developers building their own apps get an MCP connector feature in the Messages API: you list remote server URLs in the request, and Claude calls their tools during the turn. It only reaches remote servers over HTTP, so a local stdio server is out.

💡 Naming trap: inside the API docs, "MCP connector" is the name of that feature, not the menu in claude.ai. Same word, third meaning. When someone says "connector," ask which surface they mean.

Which One to Pick

Pick a Directory Connector

Choose this route when:

• The service is already listed.
• You need it working in minutes.
• Nobody on your team wants to maintain code.
• Standard permissions are good enough.

Build Your Own Server

Build when:

• The data is internal and no vendor offers a server for it.
• You need custom actions, like "quote a shipment" or "check stock in our warehouse."
• You want the same capability in more than one MCP-aware tool.

The official Python SDK makes a first version short enough to read in a minute:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("shipping-rates")

@mcp.tool()
def rate(weight_kg: float, zone: int) -> float:
    """Return the shipping price in USD."""
    return round(4.5 + weight_kg * 1.2 * zone, 2)

if __name__ == "__main__":
    mcp.run()  # stdio by default

Copy Prompt

Swap the last line for mcp.run(transport="streamable-http"), host it behind HTTPS, and the same code becomes a remote server you can paste into Add custom connector.

Stay Local

Run a local server when the data should never leave the machine: a folder of contracts, a SQLite file, a script that touches your own network. In Claude Desktop the setup is a few lines of JSON:

{
  "mcpServers": {
    "shipping-rates": {
      "command": "python",
      "args": ["/path/to/server.py"]
    }
  }
}

Copy Prompt

💡 Rule of thumb: start with a directory connector. Build your own server only when you catch yourself wishing for a tool that does exactly one specific thing for you.

Safety and Common Mistakes

Connectors and servers both give a language model the power to do things, not just say things. A little discipline goes a long way.

Approval Prompts

Claude asks before it runs a tool, and most clients let you allow once, allow always or block it. Read the prompt. A tool named "search" and a tool named "delete" or "send" deserve very different amounts of attention.

Least Privilege

Grant read-only access when read-only is enough. Use a dedicated account or a narrowly scoped token for each server. Disconnect connectors you have not touched in a month. Fewer connected tools also means fewer chances for Claude to pick the wrong one.

Another risk is text written by someone else. A web page, an email or a ticket can contain instructions aimed at the model, a trick known as prompt injection. If the same assistant can read untrusted content and run powerful write actions, keep approvals switched on.

Three mistakes show up again and again:

1. Treating the words as rivals. You do not pick between a connector and an MCP server. You pick between a ready-made server, one you build and one you run locally.
2. Installing servers from random repositories. A local server runs with your user permissions. Read the code or trust the author, exactly as you would with any program.
3. Pointing a web connector at localhost. It works on your laptop in a terminal and fails in the browser. Host it publicly or run it locally instead.

Use Claude Sonnet 5 on PicassoIA

Drafting server code, tool descriptions or test prompts is where a good model earns its place. Claude Sonnet 5 on PicassoIA is built for coding and tool-use tasks, and it runs from a plain prompt box. One honest note: it is a hosted model page, not a client that attaches to your connectors. Use it to write and debug the server, then wire that server into Claude yourself.

1. Open the Claude Sonnet 5 page on PicassoIA.
2. Type your request in Prompt, the only required field. Try: "Write a Python MCP server with one tool that converts kilograms to pounds, using the official SDK."
3. Set Effort. The default is low for fast answers; raise it when you are chasing a bug that touches several files.
4. Add a System Prompt if you want a standing role, such as "You write small, well-commented MCP servers in Python."
5. Attach an Image if it helps, like a screenshot of an error message.
6. Generate, then copy the code into your project.

Here is what each setting does:

💡 Tip: heavier jobs, like untangling a server that spans many files, may suit Claude Fable 5 or Claude Opus 4.7, both available on PicassoIA.

Make Your Own Images on PicassoIA

Your next blog post, slide deck or documentation page deserves better than a stock photo of a padlock. PicassoIA turns a sentence into a photorealistic image, and the same platform handles text to video when a still is not enough. Four text to image models worth a first try:

• Flux 2 Pro for sharp, balanced everyday scenes.
• GPT Image 2 for prompts packed with specific details.
• Seedream 5 Pro for polished, high-resolution results.
• Nano Banana Pro for quick experiments you can iterate on.

Write the prompt the way a photographer would brief an assistant: the subject and what it is doing, the setting, the direction of the light, the lens and the textures you want to see. Something like this works well on the first run:

A developer plugging a braided USB-C cable into an aluminum hub on a walnut desk, soft window light from the left, 50mm lens, shallow depth of field, visible wood grain, fine film grain.

Change one detail at a time, such as the lens or the light, and compare the results side by side. That habit teaches you more about prompting than any checklist will. When you are ready to browse everything on offer, open the full model list, pick a model, and make your first image today.