Flori / Blog

What is MCP? The Model Context Protocol explained for small teams

By the Flori team · Published 2026-10-27 · Updated 2026-10-29

MCP, short for Model Context Protocol, is an open standard that lets an AI assistant such as ChatGPT, Claude or Cursor connect to another app, read information from it and, if you allow it, make changes there. Instead of copying a client brief into a chat, you connect the app once and the assistant looks up what it needs.

As of October 2026. Protocol details come from the official MCP documentation and the vendor pages linked below, and can change.

What is MCP in plain English?

MCP, the Model Context Protocol, is an open standard that connects AI applications to external systems.

The official MCP site describes MCP as "an open-source standard for connecting AI applications to external systems" and compares it to a USB-C port: one standard plug instead of a different cable for every device. Before MCP, every AI tool needed its own custom integration with every app. With MCP, an app builds one connector and any assistant that speaks MCP can use it.

What problem does MCP solve?

If you run a small agency or a one-person business, you probably already do this every day:

1. Open your project tool and copy the brief.
2. Paste it into ChatGPT or Claude.
3. Ask for a draft or a plan.
4. Copy the answer back into the project tool by hand.

Every copy and paste loses something: the latest decision, the call note from Tuesday, the task someone already finished. MCP removes steps 1 and 4. The assistant reads the current record directly and, with permission, writes the result back.

What is the difference between an MCP client and an MCP server?

An MCP client is the assistant-side connection, while an MCP server exposes an app's tools and records.

MCP has two sides. The architecture overview uses three terms, but for everyday use you only need two:

One assistant can connect to several servers at once.

Remote vs local MCP servers

MCP servers run in one of two places, and the difference matters for non-engineers.

Remote servers run on the vendor's computers and you reach them with a web address. The official docs give the Sentry server as an example of a remote server. You add a URL to your assistant, sign in through your browser, and you are done. Nothing is installed on your machine. For most small teams, remote servers are the easy option.

Local servers run on your own computer, usually started by the assistant app with a command. They are common for developer tools, such as one that reads files in a folder. They are powerful, but they run with the same access your account has, and the official security guidance warns that a malicious local server can run any command your computer allows. Only install local servers from sources you trust.

In settings screens, local servers are often labeled stdio and remote ones Streamable HTTP.

How does sign-in work? OAuth in plain English

When you connect a remote server, you usually see a browser window that says something like "Claude wants to access your workspace." That is OAuth, the same sign-in pattern you use for "Sign in with Google."

The MCP authorization specification builds on OAuth 2.1 for remote servers. In practice, three things are true:

• You sign in on the vendor's own page, so the assistant never sees your password.
• The vendor gives the assistant a token that only works for that server.
• The server can ask for limited permissions, called scopes, such as read-only or read and write.

Local servers work differently. The spec says they should take credentials from the computer's environment, which usually means an API key you paste into a settings file. That is another reason non-engineers usually prefer remote servers.

What can MCP let ChatGPT, Claude or Cursor do?

The MCP intro page lists examples such as assistants that use your Google Calendar and Notion. For a small business, useful jobs look like this:

• "What is overdue across my client projects this week?"
• "Read the brief for the Acme website and draft the kickoff email."
• "Turn the notes from today's call into tasks on the right project."
• "Which deals have had no activity in two weeks?"
• "Write this week's status update from what was actually finished."

Is MCP safe? Scoped access, read vs update

MCP is a plug, not a security guarantee. Safety depends on the server you connect and the permissions you give. A sensible checklist:

1. Start read-only. Many servers let you choose read access or update access. Give an assistant read access first and widen it only when you trust the results.
2. Limit the scope. If a server lets you share one project or one workspace instead of everything, do that.
3. Keep tool approvals on. Claude asks for approval before using a tool unless you choose "Allow always," and lets you switch individual tools off. Cursor also asks before running MCP tools by default.
4. Prefer remote servers from the vendor itself. A connector published by the app maker is easier to trust than a random script.
5. Know how to disconnect. Check where to revoke access, both in the assistant and in the app.

How do I try an MCP server today?

You can try one in about ten minutes, without writing code.

1. Pick an assistant you already pay for. Claude supports custom connectors on Free (one connector), Pro, Max, Team and Enterprise per its help article. ChatGPT lets you add a custom MCP server too, described in OpenAI's MCP docs. Cursor reads servers from a settings file or its marketplace (Cursor docs).
2. Pick an app with a remote server. Many work tools now publish one. Notion's MCP server is a well-known example: it connects through OAuth and lets clients search, read and update pages.
3. Add the server URL. In Claude on Pro or Max, the path is Customize, then Connectors, then Add custom connector. In Cursor, see our Cursor MCP guide.
4. Sign in and choose access. Read the permission screen. Choose read-only if it is offered.
5. Ask a real question. Use a question from your actual week, like the examples above, and check the answer against the app.

Where Flori fits

Flori is a CRM with project management for small teams, and it was built to be used through MCP. Its remote server lives at https://useflori.com/mcp and connects with browser sign-in (OAuth), so there is nothing to install. You can connect Claude, ChatGPT, Grok, Cursor, Codex or Gemini CLI.

When you sign in, you choose a whole workspace or a single project, and read-only or update access. From there, your assistant can read a project brief, plan tasks, update status and save results. A few things are designed for assistants doing real work:

• Tasks marked Ready for AI are reserved by one assistant at a time, so two assistants never do the same job.
• Submitted work has a review path; authorized direct updates can save immediately.
• Every save records which assistant made it.

Flori sells no AI credits; the work runs on the assistant plan you already have. The MCP project management guide walks through setup for each assistant, and the connect page has the per-assistant steps. If you want to see why this matters for client work, read about CRM with project management.

FAQ

What does MCP stand for? MCP stands for Model Context Protocol. It is an open standard for connecting AI applications, such as Claude, ChatGPT and Cursor, to outside systems like project tools, files and databases. The assistant is the client and the connected app provides a server, so one connector works with any assistant that supports MCP.

Do I need to be a developer to use MCP? No. Remote MCP servers only need a web address and a browser sign-in. You add the URL in your assistant's connector settings, approve access, and start asking questions. Local servers, which run on your own computer, are more technical and usually need a settings file or an API key.

What is the difference between an MCP client and an MCP server? The client is the AI app you talk to, such as Claude or Cursor. The server is the connector another app provides so assistants can use it. The server offers tools, like "create task," and information, like a project brief. The client decides when to use them, usually asking you first.

Can an MCP server change my data? Only if the server offers actions that change data and you grant that access. Good servers let you choose read-only or update access, and most assistants ask before running a tool. Start read-only, limit access to one project or workspace, and check the change history.

Which assistants support MCP? The official MCP site lists Claude, ChatGPT, Visual Studio Code and Cursor among supported clients. Support and plan requirements vary by assistant and change often, so check each vendor's own help page before you rely on a feature.