What is an MCP server? A plain-English guide for people who run a business

What an MCP server does

Software has always had APIs. An API is a contract written for a programmer: here are the addresses, the fields, the error codes, go and write some code. It works, and it costs a developer a week.

MCP, the Model Context Protocol, is an open standard, originally published by Anthropic and now supported by a range of AI clients. It defines a way for a program to hand an AI assistant three things at once:

1. 1.A list of the tools it offers: what can be done here.
2. 2.A description of each tool: what it takes, what it returns, and what it means.
3. 3.An authentication handshake: who is asking, and what they are allowed to do.

An MCP server is the piece of software that publishes those three things for a specific product. Your CRM might publish one. Your accounting tool might publish one. Your project tracker might publish one. When you connect your assistant to it, the assistant reads the menu and can then act inside your account, with your permissions.

The critical word is discover. Nobody writes an integration for each thing you want to do. The assistant reads what is available at the moment you ask, picks the tools it needs, and sequences them itself.

What the model gets that it does not get from a plain API

That fourth row deserves a sentence of its own. An API key is a long-lived secret that ends up in a chat window, a notes app or a screenshot. A properly implemented MCP server uses a browser-based sign-in, so the credential never passes through the assistant at all. If a vendor asks you to paste a key into a chat box, that is worth a raised eyebrow.

What changes in practice

The honest answer is: the boring 80% of multi-step work.

What does not change: judgement, taste, and anything that needs someone to be accountable. An assistant with tool access is a very fast operator, not a decision-maker.

How to connect one

For most people it is three steps and under five minutes.

1. 1.Get the server URL from the vendor. It is a normal HTTPS address, e.g. https://app.example.com/mcp. Remote servers are the ones worth having. A "local" server that needs something installed on your laptop is a developer tool.
2. 2.Add it in your AI client. In Claude Code this is one line: claude mcp add example --transport http https://app.example.com/mcp. In desktop and browser clients it is a settings dialog where you paste the URL. Clients that support remote MCP servers today include Claude, ChatGPT/Codex, Cursor, Windsurf, VS Code and Cline.
3. 3.Sign in and read the consent screen. A browser window opens on the vendor's own domain. Read what is being granted before approving. This is the one screen in the whole process that matters.

Then test with questions before you test with actions. "How many contacts were created last week?" tells you the connection works and costs nothing if the answer is wrong.

Five things that go wrong

Tool sprawl. A product with a thousand operations cannot hand a model a thousand tool definitions. It will not fit in the context window, and accuracy collapses long before the limit. Good servers solve this with a small number of meta-tools that let the model search for the operation it needs. If a vendor exposes hundreds of raw tools, expect the assistant to get slower and less accurate the bigger your request.

Plausible-but-wrong arguments. A schema says a field is a string; it does not say the string must be one of four magic values. The model fills something reasonable, the call succeeds, and the result is quietly wrong. The fix is on the vendor's side: worked examples and behavioral notes attached to each operation, not just types.

No undo. Very few systems have one. Ask what is reversible before you let an agent touch anything in bulk.

Silent partial success. Step four of six fails, and the assistant reports the parts that worked. Ask for a summary of what actually changed, and verify it in the product's own interface the first few times.

Quotas. Some vendors meter tool calls; a few thousand per month is common. Fine for occasional work, not fine if you were planning to run something daily. Find the number before you build on it.

Nine questions to ask a vendor before you rely on their MCP server

Copy this into an email. The answers will tell you more than any feature page.

1. 1.Is the server first-party and hosted by you, or a community project?
2. 2.Can the agent write, or only read? If it writes, does anything publish automatically?
3. 3.How many of your product's operations are exposed, and which areas are excluded?
4. 4.Which operations are blocked from the AI surface entirely, and why those?
5. 5.How is access granted: browser sign-in, or a key I have to paste somewhere?
6. 6.Does the agent inherit my permissions, or does it act as a superuser?
7. 7.Are there quotas or rate limits, and what happens when I hit them?
8. 8.Where do I see a log of what the agent changed, and can I revert it?
9. 9.Which plan is it on, and does the price change if I use it heavily?

Question 4 is the one that separates a serious implementation from a demo. A vendor that has thought about this has removed dangerous operations from the AI surface (billing writes, token minting, role changes) rather than instructing the model to avoid them. Instructions in a prompt are a suggestion; a missing tool is a guarantee.

Where AXL fits

AXL's remote MCP server covers roughly 1,100 admin operations across 60+ modules. Around 110 verified recipes add working request bodies and the behavioral notes a schema cannot express.

The agent signs in through the browser and inherits that person's permissions. Twenty-one sensitive operations, including billing writes, API-token creation, and role changes, are unavailable. Building and publishing are separate calls, so creating a page does not put it live.

If you want the setup details, they are in connecting an agent to AXL, and the account-side permissions model is documented in the API keys guide. For the wider picture of what an agent can run end to end, see running your business with an AI agent.

FAQ