Short answer. In Claude, an MCP server you connect is called a connector. On claude.ai and in the desktop and mobile apps you add one from the Connectors Directory or as a custom connector by URL. Claude Desktop can also run local servers as desktop extensions. In Claude Code it is one command, claude mcp add. In the Claude API, the MCP connector does the same job. We use MCP servers in client work every day: store jobs on Horeca Papatya's live Shopify store are asked for in a plain sentence, and a clinic's sales manager changes Kommo lead routing by writing to an agent. Below: every way to connect, which plan allows what, and the safety rules we keep. Facts are as of November 2026.
What does MCP mean in Claude?
MCP (Model Context Protocol) is the open standard Claude uses to call tools and read data in other systems. If the idea is new to you, start with what an MCP server is; here we only cover the Claude side.
Anthropic's products call a connected MCP server a connector. Anthropic's help center says connectors work across Claude, Claude Desktop, Claude Code and the API (through the MCP connector). Claude inherits your permissions from the connected service: if you can't open a file or a record there, Claude can't reach it either. The exception is a custom connector that uses a shared credential such as an API key; it reaches whatever that key can access.
What we have seen running Claude with MCP servers
Two of our client projects run on this setup, and both taught us the same thing: the person doing the work only types a sentence, and the safety comes from the server, not from the chat.
- A live B2B Shopify store. For Horeca Papatya, a hotel and restaurant supplier in Romania, we built a custom Shopify MCP server. Requests come in plain language (“show items with stock under 5”). Bulk jobs run as a dry run first, new products start as drafts and deletes need explicit approval. Because these rules sit in the server's code, Claude Code, Codex and Cursor all work by the same rules.
- Lead routing run by chat. For a health tourism clinic in Türkiye, the sales manager writes “A is off tomorrow” to an agent. The agent picks the right exception type, deploys it, checks it live and reports back. All 13 routing changes in about three months were made this way.
- A practical limit. Some harnesses need the server to be reachable from the internet. That matters for Claude too: a custom connector on claude.ai can't reach a server that only runs on your laptop.
Claude MCP setup: four ways to connect a server
Which door you use depends on where you work with Claude and where the server runs. As of November 2026 there are four:
- Where: claude.ai, Claude Desktop, mobile, CoworkHow you add it: Directory connector or custom connector by URLWhere the server runs: Remote, reached from Anthropic's cloudWho can add it: Free (one custom connector), Pro, Max; Owners on Team and Enterprise
- Where: Claude DesktopHow you add it: Desktop extension (.mcpb) or the claude_desktop_config.json fileWhere the server runs: On your computerWho can add it: Any Desktop user; admins on Team and Enterprise control which ones
- Where: Claude CodeHow you add it: claude mcp add, or a .mcp.json file in the projectWhere the server runs: Remote (http) or on your machine (stdio)Who can add it: The developer; claude.ai connectors also show up
- Where: Claude APIHow you add it: MCP connector: mcp_servers in the Messages API request (beta)Where the server runs: Remote, public over HTTP; no local stdioWho can add it: Developers with an API key
| Where | How you add it | Where the server runs | Who can add it |
|---|---|---|---|
| claude.ai, Claude Desktop, mobile, Cowork | Directory connector or custom connector by URL | Remote, reached from Anthropic's cloud | Free (one custom connector), Pro, Max; Owners on Team and Enterprise |
| Claude Desktop | Desktop extension (.mcpb) or the claude_desktop_config.json file | On your computer | Any Desktop user; admins on Team and Enterprise control which ones |
| Claude Code | claude mcp add, or a .mcp.json file in the project | Remote (http) or on your machine (stdio) | The developer; claude.ai connectors also show up |
| Claude API | MCP connector: mcp_servers in the Messages API request (beta) | Remote, public over HTTP; no local stdio | Developers with an API key |
The Connectors Directory: ready-made connectors
The Connectors Directory is Anthropic's catalog of MCP servers. One catalog serves claude.ai, Claude Desktop, mobile, Claude Code and Cowork. You open it from Customize > Connectors, or from the + button in a chat. Pick a connector, click Connect, sign in to the service and review what it asks for. Then you switch connectors on or off per conversation from the same + menu.
Every listing has a label. Verified means Anthropic tested the tools for quality and compatibility. Community means Anthropic screened it but did not review it in depth. Custom is one you added yourself, with no Anthropic review. Anthropic is clear that verification is not a security audit, and the developer can change the tools after review. Connectors for services like Shopify or Meta Ads are good examples of what you can plug in. Our Google Ads MCP guide shows a server you host yourself and add as a custom connector.
How to add a custom connector in Claude
A custom connector is any remote MCP server you add by its URL, listed in the directory or not. Custom connectors are available on Free, Pro, Max, Team and Enterprise plans; Free users are limited to one. On an individual Pro or Max plan the steps are:
- Step 1Open connectors
Go to Customize > Connectors, click + Add, then Add custom connector.
- Step 2Name and URL
Enter a name and the remote MCP server URL, then click Continue.
- Step 3Check authentication
Review the settings Claude detected. Choose Sign in now, Sign in when needed or No sign in.
- Step 4OAuth or a header
Pick the OAuth client option, or add a fixed API key as a request header if the server uses one.
- Step 5Add and connect
Click Add, sign in, then turn the connector on per chat from the + menu.
One detail catches many teams: Claude connects to a custom connector from Anthropic's cloud, not from your device. That is true even in Claude Desktop and Cowork. A server on a private network, behind a VPN or a firewall won't connect, even if you can open it from your own machine. You either make it reachable over the public internet or allowlist Anthropic's IP ranges. To edit a custom connector, you remove it and add it again.
Team and Enterprise: what admins control
- Who adds connectors. On Team, only Owners and Primary Owners add custom connectors, in Organization settings > Connectors. On Enterprise, so can people with a custom role that includes Manage access to Libraries.
- Everyone signs in for themselves. Adding a connector for the organization grants no access by itself. Each member connects with their own account, so Claude sees only what that person can see.
- Tool permissions. For each connector, tools can be set to Always allow, Needs approval or Blocked. Owners can apply this across the organization, for example read email but never send it.
- Requests. Members can request a connector from the directory. Requests are on by default on Team and off by default on Enterprise.
- Desktop extensions. Owners can switch public desktop extensions on or off and upload their own .mcpb extensions for one-click install.
Local MCP servers in Claude Desktop
Some servers should run on your own computer: one that reads local files, for example. Claude Desktop handles these as desktop extensions, single-click packages with the .mcpb extension. Go to Settings > Extensions, click Browse extensions and install an Anthropic-reviewed one, or install your own .mcpb file under Advanced settings. Sensitive fields such as API keys are stored in the operating system's secure storage (Keychain on macOS, Credential Manager on Windows).
Servers set up by hand in claude_desktop_config.json still work and use your local network. Keep in mind that they are a separate mechanism: they are not available in Cowork or on claude.ai.
Claude Code: claude mcp add
In Claude Code you add a server from the terminal. The documentation recommends the HTTP transport for remote servers; the older SSE transport is deprecated. For a local server the transport is stdio, and everything after the -- is the command that starts the server. These are the examples from the Claude Code docs:
# Remote server over HTTP
claude mcp add --transport http notion https://mcp.notion.com/mcp
# Remote server that needs a fixed token
claude mcp add --transport http secure-api https://api.example.com/mcp \
--header "Authorization: Bearer your-token"
# Local server over stdio: everything after -- starts the server
claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable \
-- npx -y airtable-mcp-server
# List, inspect and remove servers
claude mcp list
claude mcp get notion
claude mcp remove notionEach server has a scope. Local is the default: the server loads only in the current project and stays private to you. User (--scope user) makes it available in all your projects. Project (--scope project) writes it to a .mcp.json file in the project root, so the team gets it through version control; Claude Code asks for approval before it uses servers from a .mcp.json. Inside a session, /mcp shows server status and runs the OAuth sign-in in your browser. If you are logged in with a claude.ai account, the connectors you added on claude.ai show up in Claude Code automatically; they don't load when an API key is the active login. Servers already set up in Claude Desktop can be imported with claude mcp add-from-claude-desktop.
MCP in the Claude API: the MCP connector
If you build your own product on the Claude API, the MCP connector lets the Messages API talk to a remote MCP server without writing an MCP client. You pass the server in mcp_servers and enable its tools with an mcp_toolset in tools. It is in beta, with the mcp-client-2025-11-20 or the newer mcp-client-2026-09-15 header; the newer one can pin a server's tool list so it doesn't change in the middle of a conversation. Limits: only tool calls are supported, the server must be public over HTTP, local stdio servers can't be connected directly, and you handle the OAuth flow yourself and pass the token.
Claude MCP security: what to check first
A connector does more than read. Depending on its tools, Claude can create, change or delete data in the connected service on your behalf. Anthropic's own guidance and the MCP project's security pages come down to a few rules:
- Only connect servers you trust. Custom connectors have not been verified by Anthropic, and even a Verified connector can change its tools after review.
- Read the OAuth screen. With OAuth, Claude never sees your password. Check the scopes the server asks for and refuse ones it doesn't need. A shared API key gives everyone who uses the connector that key's access.
- Keep write tools on approval. Read Claude's tool approval requests and click Allow always only for tools you would let run unsupervised. In Research, Claude can call connector tools without asking, so switch off write tools there.
- Watch for prompt injection. A malicious server can hide instructions in its output that try to make Claude do something you didn't ask for. Claude has built-in protections, but you should still watch tool inputs and outputs.
- Know about tool poisoning. The MCP project's security guide notes that tool names and descriptions are read by the model as instructions, and a server can change them after you approved it. Watch for unexpected changes in tool behavior.
- Treat local servers as software you install. A local server runs with your user's privileges. Read the exact command before you add it, the same as any program you download.
A prompt that says “be careful” is lost the moment the harness changes. Put dry runs, drafts and delete approval in the server's code, as we did for Horeca Papatya.
Which way should you use?
- Who you are: Business userBest route: A directory connector on claude.ai or Claude DesktopWhy: No code; sign in with OAuth and switch it on per chat
- Who you are: Business user with an in-house toolBest route: A custom connector to a remote MCP serverWhy: Works in chat, desktop and mobile; the server must be public
- Who you are: DeveloperBest route: Claude Code with claude mcp add, http or stdioWhy: Local and remote servers; .mcp.json shares the setup with the team
- Who you are: CompanyBest route: Team or Enterprise with admin-added connectors and tool permissionsWhy: Central control; each member signs in with their own access
| Who you are | Best route | Why |
|---|---|---|
| Business user | A directory connector on claude.ai or Claude Desktop | No code; sign in with OAuth and switch it on per chat |
| Business user with an in-house tool | A custom connector to a remote MCP server | Works in chat, desktop and mobile; the server must be public |
| Developer | Claude Code with claude mcp add, http or stdio | Local and remote servers; .mcp.json shares the setup with the team |
| Company | Team or Enterprise with admin-added connectors and tool permissions | Central control; each member signs in with their own access |
If your product needs Claude to call your tools from your own backend, use the MCP connector in the API instead. And if the system you want to connect has no MCP server yet, someone has to build one; how to build an AI agent covers the steps around it. If you would rather wire tools together without code, read what n8n is and when it is enough.
What is Claude MCP?
Claude MCP is how Claude uses the Model Context Protocol to connect to outside tools and data. In Claude's apps a connected MCP server is called a connector. You can add connectors on claude.ai, Claude Desktop and mobile, add servers in Claude Code with claude mcp add, or connect from the Claude API with the MCP connector.
Can I use MCP on the free Claude plan?
Yes, with a limit. As of November 2026, custom connectors using remote MCP are available on Free, Pro, Max, Team and Enterprise plans, and Free users are limited to one custom connector. Desktop extensions are available to all users of Claude Desktop.
How do I add an MCP server to Claude Code?
Run claude mcp add in your terminal. For a remote server: claude mcp add --transport http <name> <url>. For a local server: claude mcp add --transport stdio <name> -- <command>. Add --scope user to use it in all projects or --scope project to share it with the team in .mcp.json. Use /mcp inside Claude Code to check status and sign in.
Why won't my custom connector connect to Claude?
The most common reason is the network. Claude reaches custom connectors from Anthropic's cloud, not from your computer, even in Claude Desktop. A server on a private network, behind a VPN or blocked by a firewall won't connect. Make it reachable over the public internet or allowlist Anthropic's IP ranges. For a server that must stay local, use a desktop extension or Claude Code.
Is it safe to connect Claude to MCP servers?
It is as safe as the server and the permissions you give it. Connect only servers from organizations you trust, review the OAuth scopes, keep write and delete tools on approval, and watch for prompt injection and unexpected changes in tool behavior. For business systems, put hard rules such as dry runs, drafts and delete approval in the server itself.
What is the difference between a connector and a desktop extension?
A connector is a remote MCP server that Claude reaches over the internet; it works on claude.ai, Claude Desktop, mobile and in Claude Code. A desktop extension is a local MCP server packaged as a .mcpb file that runs on your computer, and it is available only in Claude Desktop.
Sources
- Claude Help Center: Get started with custom connectors using remote MCP
- Claude Help Center: Use connectors to extend Claude's capabilities
- Claude Help Center: Getting started with local MCP servers on Claude Desktop
- Claude docs: Connectors directory
- Claude docs: Connector verification
- Claude Code docs: Connect Claude Code to tools via MCP
- Claude Platform docs: MCP connector
- Model Context Protocol: Security best practices
- Model Context Protocol: Local server security (draft)
- Claude
- MCP
- AI agent

