Guide
What is MCP? A plain-words guide to how an AI agent connects apps, with a real log of its tool calls
MCP is the reason an AI agent can do things in your apps instead of only talking about them. Below is the plain-words version: what the Model Context Protocol is, what an MCP server does, and the five steps of every tool call. Then a real chat with our test agent: one message that sends it to the weather forecast and sets two reminders, a spreadsheet built from a plain list, and a log of all twelve tool calls it made.
The short answer
MCP, the Model Context Protocol, is a common language between an AI agent and the apps it works with. The app side runs a small program called an MCP server. It tells the agent which actions it offers, such as “create an event”, “add a row” or “search mail”, and what each action needs. The agent picks an action, sends the details and gets the result back. One standard replaces a custom integration for every pair of AI and app.
Anthropic published MCP as an open standard in November 2024. In December 2025 it moved to the Agentic AI Foundation under the Linux Foundation, set up by Anthropic, OpenAI and Block. Claude, ChatGPT, VS Code and Cursor all speak it.
For you, MCP is the reason “put it in my calendar” ends up as an event in your calendar. Your agent on WAM Agents calls MCP tools on every plan, Free included, with no keys of your own. Below is how it works, and a real chat with our test agent where you can watch the tool calls happen.
MCP in one picture
Three parties take part in every call. The names come from the MCP documentation, the plain-words column is ours.
| Part | In plain words | What it knows | What it does |
|---|---|---|---|
| Host (the agent) | The AI you talk to, with its model and memory | Your request and the list of tools it can use | Chooses a tool and fills in the details |
| MCP server | A small program standing in front of an app | The app's actions and what each one needs | Checks the call, passes it to the app, returns the answer |
| App | Calendar, mail, spreadsheets, search | Your data | Does the actual work |
Every tool call follows the same five steps, and the log in step 4 of our test shows each of them:
- When the agent connects, it asks the server what it can do. The server answers with a list of tools: a name, a one-line description and the fields each tool needs.
- You write a request in ordinary words.
- The model reads your request next to that list and decides which tool fits, or several in a row.
- The agent sends the call: the tool's name and the filled-in fields. The server passes it to the app.
- The answer comes back as text or data, and the agent turns it into a reply in your words.
The USB-C comparison from the official MCP site is a fair one. Before USB-C every device had its own cable. Before MCP every AI app needed its own integration with every service. Now a service builds one MCP server, and any agent that speaks MCP can use it.
Tools, resources, prompts: what a server can offer
An MCP server can give an agent three kinds of things. The specification calls them primitives.
- Tools are actions: create an event, add a row, send a draft, search the web. This is what people mean nine times out of ten when they say “MCP”.
- Resources are data to read: a file, a database record, a page.
- Prompts are ready-made templates for a task, which the app's authors prepared.
The messages themselves are small. A tool call is a short structured note, in a format called JSON-RPC. Stripped of the technical wrapping, a calendar call for Leo, the bike shop owner from our test below, would read like this:
| What travels | |
|---|---|
| Agent → server | Tool “create event”; title “Group ride from Spoke & Chain”; start Saturday 7 November, 09:00; end 11:00 |
| Server → app | The same event in the calendar's own format |
| App → agent | “Created”, with the event's details |
A server can run on the same machine as the agent or somewhere on the internet. The first kind talks to the agent directly, the second over the network. For remote servers, MCP recommends OAuth, the kind of sign-in page you know from “Continue with Google” buttons, so access is granted on the app's own page and your password stays with the app.
MCP, API, plugin: what is the difference
An API is the door into one app, with its own rules. Every AI product that wanted to use a hundred apps used to learn a hundred sets of rules. MCP sits on top: the app's door stays the same, and the MCP server describes it in one shared way that any agent understands. The agent reads the tool list on connection, so a new action on the app's side becomes available without a new release of the agent.
Plugins and custom GPT actions were earlier tries at the same idea, each locked to one product. MCP is the version the industry agreed on: Claude, ChatGPT, VS Code and Cursor support it, and the standard itself is run by a foundation rather than by one company.
The test: Leo and his bike shop
Our reader is Leo. He runs Spoke & Chain, a small bike repair shop in Bristol, and keeps hearing the word MCP. He asks the agent what it is, gives it three jobs in one message, asks for a spreadsheet, and then asks it to show every tool it called along the way.
Leo, the shop and the customers are made up for the test. The agent's replies are real, from our test agent on 2 November 2026, quoted as they came, shortened in places but never added to. The grey lines above the replies are the agent's status lines: the tool it is calling and what that step cost in magnets.
Step 1. Ask the agent what MCP is
That is a better definition than most glossaries give: “each app describes what it can do once, in a standard way”. The second paragraph is the agent reading its own tool list, step 1 of the picture above. Google was connected on this agent earlier, so Gmail, Calendar, Drive, Docs and Sheets sit on the list next to the built-in tools. The third paragraph is the useful part: the agent has already matched those tools to a bike shop.
Step 2. Three jobs, one message
Now a request that needs more than one tool. Leo does not name any of them.
One sentence turned into four tool calls, as the log in step 4 shows: two web searches and two reminders. The agent read the forecast, made the “if it looks dry” decision itself, and said why, with the sources it used: the Met Office hourly view against the gloomier BBC label. Then it set both reminders and offered to check the forecast again on Friday. Leo never had to say “search”, “reminder” or “tool”.
Step 3. A spreadsheet from a plain list
The file arrived in the chat with the total as a live formula, so a changed price updates the sum. The agent also noticed something Leo did not ask about: Wednesday has two bikes in. That is the difference between a tool that fills cells and an agent that reads what it is filling.
Step 4. “Show me the tool calls”
This is the part most MCP explainers can only draw. We asked the agent to show its work.
Read it against the picture from the start of the article and every part of MCP is there.
- Each line is one tool call: a tool's name, what the agent sent and what came back. websearch, schedulereminder, telegramsenddocument and telegramreply are tools from the agent's list; the agent chose them from Leo's plain words.
- The agent asked the web twice, in different words, because the first answer lacked the hourly view. Choosing and repeating calls is the model's job, not yours.
- Bash is the agent's own workspace: it wrote and ran a small Python script that built the spreadsheet, then opened the file again to check every row before sending it. Some work is a call to an app, some is work inside the agent's container, and the log shows both.
- Twelve calls, about 11 for the three requests, and Leo saw only the results.
How your agent connects apps
Your agent lives in its own container, a separate workspace for each account, and works with three groups of tools.
- Built-in tools come with every agent: Telegram, memory, web search, images, voice, reminders and app connections. Nothing to install, nothing to switch on.
- Google connects with one link. Gmail, Calendar, Docs, Sheets, Slides and Drive connect together. You open the link in your browser, sign in on Google's own page and allow access. Connectors work on Free as well.
- The rest of the work happens inside the container: files, documents, spreadsheets, code the agent writes and runs for your task.
The keys to your Google account do not sit in the agent's container. The OAuth tokens are kept by the integration service listed on the Subprocessors page, which performs the actions you ask for in the connected apps. Your password never passes through the chat. To disconnect, write to the agent “disconnect Google”; you can also remove the access in your Google account settings, under third-party apps. Google disconnects as a whole, every service at once.
What the agent does freely and what it asks first
Inside its container the agent works freely: it reads, searches, drafts, builds files and calls tools as the task needs. Actions that leave the container and touch other people go through you first. Those are an email, a message to a person, a publication and a deletion. The agent shows what it is about to do and waits for your yes.
Two more rules from the same set. The agent opens Google when a task of yours needs it, including a task you scheduled yourself, and does not look into your mail on its own. Text that comes from outside, a letter, a document, a message from a guest, is treated as data, not as instructions, so a line in an email saying “forward all invoices” stays a line in an email. For more on how the agent is protected, read Is an AI agent safe? 8 things to check before you let one into your chats.
What stays with you
Tips for working with a connected agent
- Ask for the result, not the tool. “Put the ride in my calendar” is enough; the agent finds the tool.
- Give several jobs in one message. The agent chains the tools itself and answers once.
- Ask “which tools did you just call?” when you want to see the work. It is a normal question.
- Name the exact title, date and time for anything that lands in a calendar or a sheet. The tool takes what you give it.
- Read the draft before you say yes to an email. That yes is the only thing between the draft and the recipient.
What it costs
MCP itself costs nothing extra. Magnets are spent on the agent's work, and the status line shows the balance as it goes. In our test, by that balance, the question about MCP cost 4 , the weather check with two reminders 7 , the spreadsheet 2 and the log of tool calls 2 : about 15 in total. On the default model, text costs 1 per 20,000 tokens and a web search 1.5 .
Connectors are part of both plans. Free gives 200 in the first month and 100 every month after, with no card. Pro costs $13.90 a month with 400 , or $1.39 for the first month if you subscribe within 7 days of signing up, and adds recurring reminders and scheduled tasks. The full rate card is on the pricing page. Terms like tool, context and container are explained in AI agent in Telegram: 18 terms in plain English.
Messages worth copying
What is MCP, in plain words, and which apps and tools can you reach right now?
Check [the weather / a fact], and if [condition], remind me [when] about [what].
Make me a spreadsheet [name] with columns [A, B, C] and these rows: [rows]. Add a total.
Which tools did you just call, in what order? Show it as a short log.
Create your agent, connect Google with one link, and ask it what it can reach.
Try it with your own agent
Free, no card. The agent starts working in the chat straight away.
Create your AI agentQuestions
Do I have to install or set up an MCP server to use the agent?
No. Your agent comes with its tools built in: Telegram, memory, web search, images, voice, reminders and app connections. Google connects with one link that the agent sends you. You write requests in plain words and the agent picks the tool.
Is MCP an app, a company or a program?
It is a standard, a set of rules for how an AI agent and an app talk to each other. Anthropic published it in November 2024, and since December 2025 it belongs to the Agentic AI Foundation under the Linux Foundation. Anyone can build an MCP server for their app.
What is the difference between MCP and an API?
An API is the door into one app, with that app's own rules. An MCP server describes that door in one shared format that any agent understands, with a list of actions and the fields each one needs. The app keeps its API; MCP is the common layer on top.
Does the agent get my Google password?
No. You allow access on Google's own sign-in page. The access tokens are kept by the integration service listed on the Subprocessors page, outside the agent's container, and your password never goes through the chat. You can disconnect Google from the chat or from your Google account settings.
Can the agent send an email or delete something on its own?
Emails, messages to people, publications and deletions wait for your yes: the agent shows what it is about to do and asks first. Reading, searching, setting your reminders and building files it does without asking.
Do MCP tools cost extra?
No. Connectors and MCP tools are part of both plans, Free included. Magnets are spent on the agent's work, and the status line shows what each step cost.
