What is an MCP server?

Key points

• MCP is an open standard for connecting AI assistants to outside systems, and both ChatGPT and Claude support it.
• A server gives the AI assistant tools. Some tools only read data, and others can change it.
• ChatGPT and claude.ai connect to remote servers only. A server running on your own computer does not work there.
• Before you approve a server, check who runs it, which tools it has and which permissions it asks for.
• Mitos is a hosted MCP server, read-only by default, that reaches one organization per connection.

Local and remote MCP servers

What the Model Context Protocol is

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to outside systems, such as files, databases and business tools. Anthropic introduced it on 25 November 2024 as an open standard and an open-source project.

The official MCP site compares it to a USB-C port for AI applications: one shared way to connect instead of a separate integration for each app. Among the apps that support it, the site lists Claude, ChatGPT, Visual Studio Code and Cursor.

What an MCP server offers

The AI assistant asks the server for a list of what it offers. When you ask a question, the app picks a tool, calls it and writes its answer from the result. The MCP specification defines three kinds of offering:

• Tools: functions the AI assistant calls, such as a traffic report or an order search
• Resources: data that gives context, such as the contents of a file
• Prompts: ready-made templates that help the AI assistant use the tools

Why ChatGPT and Claude on the web need a remote server

A local server runs on your computer and talks to the AI assistant on the same machine. A remote server runs on its provider's servers, and the AI assistant reaches it over the internet.

OpenAI says ChatGPT does not connect to a local server directly, only to remote ones. Anthropic says Claude reaches a custom connector from Anthropic's cloud, even in Claude Desktop, and that local servers set up in Claude Desktop are not available on claude.ai. So for ChatGPT or Claude to read your data, you add a remote server's address as a custom connector. In ChatGPT that needs developer mode turned on, and OpenAI calls the result an app.

What read-only tools mean

Whoever builds a server decides what each tool does. A tool can read data, or it can create, change or delete it. A read-only server gives the AI assistant only tools that read.

A tool can declare that it only reads. ChatGPT asks for confirmation before write actions by default, and treats any tool without that declaration as a write action. The MCP specification adds that apps must treat such declarations as untrusted unless the server is trusted. Before you approve a server, check:

• Who builds it and who runs it
• Its tool list, and whether any tool changes data
• The permissions it asks for when you sign in
• How you take access back when you no longer need it

How sign-in to a remote server works

The MCP specification bases sign-in for servers reached over HTTP on OAuth 2.1. The AI assistant sends you to the service's sign-in page, you approve access, and the app receives a token instead of your password.

Anthropic says this way Claude never sees your password, and that you revoke access by disconnecting the connector in Claude or in the service's security settings. ChatGPT accepts servers that use OAuth, no sign-in, or a mix of both.

One example: Mitos

Mitos is a hosted MCP server at https://mcp.sar-soft.com/mcp. In the Mitos app you connect Google Analytics, Search Console, WordPress, WooCommerce or Magento, choosing one site or store for each WordPress or Magento connection and the Google properties you want. Then you add the address to ChatGPT or Claude, sign in with OAuth and authorize one organization.

Other options exist, such as servers that software vendors run themselves or open-source servers you run on your own. Put them through the same checks. For Mitos the answers are:

• By default the AI assistant gets only tools that read. It changes your site, your store or Google Analytics only if an owner or admin allows it, and anything that would go live waits for approval, in the chat where your AI assistant supports it, otherwise in Mitos.
• One connector reaches one organization, and each organization is isolated from every other.
• You revoke an AI assistant in the Mitos app, and it takes effect within about a minute.
• Mitos has no AI model of its own. ChatGPT or Claude writes the answer, which can be wrong or incomplete.

Questions to try

• How much traffic did we get in the last 30 days, compared with the 30 before?Sources: Google Analytics
• Which search queries brought us the most clicks in the last 28 days?Sources: Search Console
• What were our best-selling products this month?Sources: WooCommerce
• Which pages get plenty of search impressions but little actual traffic?Sources: Search Console · Google Analytics

Questions people ask

References

• Model Context ProtocolWhat is the Model Context Protocol (MCP)?
• Model Context ProtocolArchitecture overview
• Model Context ProtocolAuthorization
• Model Context ProtocolTools
• AnthropicIntroducing the Model Context Protocol
• AnthropicGet started with custom connectors using remote MCP
• OpenAIChatGPT developer mode
• OpenAIDeveloper mode and MCP apps in ChatGPT