MCP is the open standard that lets Claude reach your files, calendars, databases and design tools. This article separates servers, connectors and apps, shows how to add each one in Claude, and explains the permissions worth checking before you click Allow.
Open the connectors menu in Claude and you will meet three words that people use almost interchangeably: servers, connectors and apps. All three trace back to one idea, the Model Context Protocol (MCP), and the differences between them decide what Claude can read, what it can change, and what you see on screen when it does. This article sorts out the vocabulary, shows how a request travels from your chat box to a tool and back, and walks through adding MCP access in claude.ai, Claude Desktop and Claude Code. It also flags the permission settings worth a second look, and ends with a practical example: connecting an image and video generator to Claude.
💡 Quick answer: MCP is an open standard that lets an AI app talk to outside tools in one consistent way. In Claude, a server provides the tools, a connector is how Claude reaches that server, and an app is a connector that also draws an interactive screen inside the chat.
MCP is an open protocol that Anthropic introduced in November 2024 so AI applications could connect to outside systems without a custom integration for every pair. Before it, wiring five assistants to ten tools meant up to fifty separate pieces of glue code, each with its own login flow, data format and bugs. With MCP, a tool is wrapped once as a server, and any compatible client can use it. People often compare it to USB-C for AI: one plug shape, many devices.
Claude is one MCP client among many. Editors such as VS Code and assistants such as ChatGPT also speak the protocol, which means a server you build once is not locked to a single vendor.
A language model only knows what sat in its training data and what is inside the current conversation. Anything live, such as your calendar, a database row or a design file, has to be handed over at the moment you ask. Early integrations did that with one-off plugins. MCP standardizes three things instead:
A company ships one MCP server for its product, and Claude, code editors and other hosts can all connect to it.
The three terms sit on top of each other like layers. Anthropic's own documentation describes a connector as a link between Claude and an outside app or service, such as your calendar, your documents or your issue tracker. A developer breakdown of the same idea puts it bluntly: every connector is an MCP server, and an app is a connector that also ships a user interface.
An MCP server is a program that exposes capabilities. It might wrap a database, a file system, a project tracker or an image generator. It never talks to users directly. It answers structured requests such as "list your tools", "run this tool with these arguments" or "read this resource". A server can run on your own machine or at a public URL.
A connector is Claude's name for a configured link to an MCP server. In claude.ai and the desktop app you open Customize, then Connectors, and either pick a service from the directory or add a remote server by its URL. Directory listings carry a Verified or Community label, so you can see how much review a listing has had. After connecting, you switch the connector on per conversation from the + menu in the message box.
Connectors come in two shapes:
A Claude app is a connector that also returns an interactive screen. When Claude calls one of its tools, the result carries two things: data for the model, and a linked interface resource for you. Claude draws that resource inside the conversation, usually as a sandboxed iframe, so you can click, filter, select rows, fill in forms or approve an action without leaving the chat. The extension behind this is called MCP Apps, and services such as Asana, Box, Canva, Figma, Hex, monday.com and Slack have shipped interactive versions of their connectors.
| Term | What it is | Who sees it | Example |
|---|---|---|---|
| Server | A program that exposes tools, resources and prompts | Developers and admins | A server wrapping a project tracker |
| Connector | Claude's configured link to a server | You, in the Connectors page | A calendar or document connector |
| App | A connector that also renders an interactive screen in chat | You, inside the conversation | A task board or design canvas you can click |
Not every connector is an app, but every app is a connector.
💡 If a tool only needs to return text or numbers, a plain connector is enough. Add an app screen only when a person needs to pick, review or approve something.
Three roles do the work. The host is the application you use, such as claude.ai or Claude Desktop. Inside it, an MCP client keeps a one-to-one connection with each server. The server exposes the capabilities. Say you ask, "What is on my calendar tomorrow afternoon?" The flow looks like this:
Servers offer three building blocks, and each one has a different party in charge of using it.
| Building block | Purpose | Who triggers it |
|---|---|---|
| Tools | Actions or lookups with defined inputs | The model decides |
| Resources | Readable data such as files or records | The app or the user attaches them |
| Prompts | Reusable instruction templates | The user picks them |
Most connectors people use every day are tool-heavy. Claude picks well among tools when each one has a clear name and a short, honest description, so a well-written description matters more than clever code behind it.
The protocol defines two standard transports. With stdio, the client launches the server as a subprocess on your machine and they exchange JSON-RPC over standard input and output. With Streamable HTTP, the server lives at a URL and messages travel over HTTP, which is what makes remote connectors possible. Older servers sometimes still advertise the earlier HTTP plus SSE style, but new work should target Streamable HTTP.
Pick stdio for data that stays on your machine, and Streamable HTTP for a hosted service that many people share.
For a service that is not in the directory, add a custom connector with its remote MCP server URL. On Team and Enterprise plans an owner adds the connector for the organization first, and members then connect with their own accounts.
⚠️ Custom connectors can point at services nobody has reviewed. Add a URL only when you trust whoever runs it.
Connectors you add in claude.ai are available in Claude Code when you sign in with your Claude account. You can also register a server straight from the terminal:
claude mcp add --transport http my-server https://example.com/mcp
claude mcp list
The second command confirms the server is registered and shows its connection status.
Claude Desktop can also launch local servers from a JSON file named claude_desktop_config.json. This example starts the reference file system server for a single folder:
{
"mcpServers": {
"files": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents"]
}
}
}
Restart the app after editing the file. Pointing the server at one folder instead of your whole disk is the most useful safety choice you can make here. For most people, a desktop extension is easier still, because it bundles a local server that installs in one click and skips the JSON entirely.
A connector gives Claude real reach. That is the whole point, and it is also the risk. Three failure modes show up most often:
On each connector's page, Tool permissions lets you set every tool, or every group of tools, to Always allow, Needs approval or Blocked. A sensible baseline:
A small shop makes a good test. Picture a bakery owner who opens a laptop at six in the morning, asks Claude for today's orders, the week's delivery slots and a draft reply to a wholesale customer. Three connectors, one question, no tab switching.
| Job | Connector type | What Claude does |
|---|---|---|
| Morning planning | Calendar and email | Reads tomorrow's events and drafts replies |
| Sales follow-up | CRM | Pulls account notes and summarizes the last call |
| Bug triage | Issue tracker | Lists open tickets and proposes a fix |
| Dashboard review | Analytics app | Renders a chart you can filter in the chat |
| Design handoff | Design app | Opens a canvas inside the conversation |
| Local research | File system server | Reads a folder of PDFs and answers questions |
Start with one read-only connector for a task you repeat every week. Add a second one only after the first has saved you time twice.
MCP is not limited to office data. Media generation suits it well, because a prompt goes in and a file URL comes back. PicassoIA provides an MCP connector that works inside Claude, so a chat can plan a visual, generate it and revise it without opening another tab.
Through the connector, Claude can call four PicassoIA models:
| Model | Output | Best for |
|---|---|---|
| PicassoIA Image | Image | Hero images, blog art, social visuals |
| PicassoIA Image Editor Pro | Edited image | Retouching or changing an existing picture |
| PicassoIA Video | Video | Short clips from text or a photo |
| Seedance 2.5 Lite | Video with audio | Clips of up to 10 seconds with sound |
Its tools handle image generation, image editing, video generation, status checks, past generations, model listing, account details and job cancellation. The connector notes say its own GPU models are free on the Infinite and Wonder plans, so check your plan on the pricing page before a large batch.
Generation is asynchronous. A request returns a prediction ID once a GPU accepts the job, and Claude checks the status tool after the suggested wait until the job reports succeeded or failed. A failed job is final, so Claude submits a new one. An account allows up to five predictions at once, shared across every connection.
Planning the work is the other half. If you want a Claude model on PicassoIA itself to draft prompts, write an MCP server or debug a config file, Claude Sonnet 5 sits in the Large Language Models category. Here is how to run it:
💡 A strong first prompt: paste a draft tool description and ask for three tighter rewrites, each under 20 words. Clear descriptions are what help Claude choose the right tool.
For quick lookups, Claude 4.5 Haiku answers faster. For heavier reasoning, Claude Opus 4.7 is the one to reach for.
Reading about MCP only goes so far. The quickest way to make it click is to connect one thing, ask Claude for one job, and watch the approval prompt appear. Then give the connector something creative to do.
Open PicassoIA Image and write a scene in plain words: a place, a light direction, a lens, a mood. Pick the 16:9 ratio, set the number of outputs to 2 and compare both results. Lock a seed when one frame is close, so the same prompt gives the same picture again. When an image is nearly right, hand it to PicassoIA Image Editor Pro instead of starting over, and send your best frame to PicassoIA Video when you want motion.
Then connect the same models to Claude and let it run the whole loop from one chat. Create your own images with Picasso IA, change one prompt detail at a time, and see how much a single sentence moves the result. Every model mentioned here is listed at picassoia.com/en/all-models.
