Your AI knows the whole internet and nothing about your business. MCP fixes that: it lets your AI into the tools your team already runs on. Here’s how, and what to do with it.
Without MCP
AIThe tools you pay for
The support desk
The tracker
Accounting
Every tool’s door was its own shape. Reaching one meant building something for that one.
With MCP
AIMCP
The support desk
The tracker
Accounting
One shape, agreed by everybody. Nothing changed inside the tools and nothing changed about your AI. Only the way in.
Three tools, three shaped holes, and a plug that fits none of them. With MCP the holes are one shape and the cable reaches all three.The same picture drawn twice. Without MCP: an AI with a single plug on the end of its cable, and three tools you pay for, each with a way in for software drawn as a hole of a different shape, one round, one triangular, one a slot. The cable stops short, because one plug cannot fit three different holes. With MCP: the same AI and the same three tools, but every hole is now the same shape, the cable carries a mark reading MCP, and it reaches all three.
How it works
The door was already there. Nobody agreed on its shape.
Your tools have always been able to talk to each other. Every one just did it differently, so connecting any two was a project. MCP gives them all one common language, so your AI can plug into any of them without one.
ModelThe AI itself. The thing you type into.
ContextEverything it would need to know to be useful to you, which is sitting in the tools you already pay for.
ProtocolAn agreed way for two pieces of software to talk, so neither one has to be built specially for the other.
One tool, two ways in
A personAn AI
The support deskThe screenMenus and buttons, and somebody who learned themThe door for softwareBuilt to be used by other programs, in the agreed shapeaskyesterday’s messagesbackthree records
The same records, either way
The export ran and the file is emptyAnswered
Two of our people cannot sign inAnswered
The summary totals do not add upOpen
One tool, two ways in, and the same three messages behind both.One tool, drawn as a window holding its records: yesterday’s messages, each with the customer who wrote it and whether it was answered. On its face are two ways in. The first is the screen, with menus and buttons and somebody who learned where everything is, and a person is at it. The second is narrower, a door built for software in the way everybody agreed on, and an AI is at it. Both lead to the same records inside.
Connecting two tools isn’t new. What’s new is that your AI can walk through the door on day one, with nobody building anything first. The question is what it should go in there to do. The answer is smaller than you’d think.
What it is for
Start with the copy-and-paste. There’s more of it than you think.
Yesterday’s delivery times typed from one screen into another. A certificate pulled from one portal into another file. Half an hour a morning, in every team, and nobody has ever counted it.
Request for a piece of work
Sponsor
The business case
Estimated cost
Which budget it comes out of
Expected benefit, in numbers
Sign-off
Never filled in
The job it would have been forMove yesterday's bugs into the tracker30 minutesevery morning30 min × 5 mornings × 46 weeks= 115 hours a year
Six fields, never filled in, for thirty minutes a morning and a hundred and fifteen hours a year.A request form for a piece of work, six fields long: sponsor, the business case, estimated cost, which budget it comes out of, expected benefit in numbers, and sign-off. It is stamped never filled in. The job it would have been for is drawn at a fraction of the form's height: move yesterday's bugs into the tracker, thirty minutes, every morning. Under it, the arithmetic: thirty minutes by five mornings by forty-six weeks is a hundred and fifteen hours a year.
That’s one of them, and yours won’t be bugs. It’ll be a renewal, a delivery note, a timesheet. Count the teams in your business and multiply. Nobody assembles a sponsor and a business case to win back a morning routine, so start at the other end: not with a connection, with the job.
One job, written down
What goes in
Yesterday's customer messages
Everything, not a filtered list
What has to come out
A drafted job for each one that is a bug
Enough detail for an engineer to start
The support lead checks itbefore anything is opened
And then, last, how the facts travel
A scheduled export
A direct connection somebody builds
A script somebody wrote
MCP
What goes in, what has to come out, who checks it, and four possible routes with MCP the only one lit.One card with two halves. The half headed what goes in holds yesterday's customer messages, all of them rather than a filtered list. The half headed what has to come out holds a drafted job for each message that is a bug, with enough detail for an engineer to start. The support lead is marked on the card as the person who checks it before anything is opened. Last on the card, four ways the facts could travel: a scheduled export, a direct connection somebody builds, a script somebody wrote, and MCP. MCP is in ink and the other three are greyed.
Write the job down and the route becomes the easy part. The route is the last line on the card, and the only one you can change your mind about later. So pick one job. Here’s half an hour of it, running.
The fix
One prompt, both doors. A person still reads the list.
One person types one instruction and it reaches both tools. The AI does the reading and the drafting. The person does the deciding. Nothing gets opened in the second tool until they’ve read the list and ticked it.
Yesterday
The export file came out empty
Two of our people cannot sign in
Five customer messagesThe support desk's door
The AI reads and drafts
Export runs to an empty file
Two customers cannot sign in
Asked for a refund
Totals do not add up on the summary
Wants a feature we do not have
The support lead ticks three
Only thenThe tracker's door
Export writes an empty file
Three jobs openedThe other two stay where they are
Five messages read, five drafts written, three ticked, three jobs opened.The work runs through three stages. First, yesterday's five customer messages, each with the customer who sent it, read through the support desk's door: an export that came out empty, two people who cannot sign in, a refund request, totals that do not add up, and a request for a dark mode. Only the first two are written out. The last three are drawn as grey bars, because the drafted list writes all five out again. Next, the AI chip holds a drafted list of five: an export running to an empty file, two customers unable to sign in, a refund request, totals that do not add up on the summary, and a request for a feature that does not exist. The support lead ticks three of the five, the three that are bugs. Last, the tracker's door, with three jobs opened behind it: an empty export, a sign-in failure and wrong totals. Only the first is written out and the other two are drawn as grey bars. The two messages that were not bugs stay where they are.
The half hour that came back was the typing. Every decision inside it still belongs to the person who always made it. Nothing was built and nothing was bought. So it’s worth being precise about what you did get.
What you got
One person’s assistant got further. The business didn’t get anything.
Nothing new was bought and nobody else was involved. One person’s AI can now reach the tools they already use, and everything it does still lands with them alone.
What one person’s AI can reach
ContextWhat the AI can read right now
One person, their AI
Meeting notes
Calendar
Email
Support desk
Accounting
Budget
Maps
Files
Eight doors around one person’s AI. Each one that opens is more it can read, for them alone.One person and their AI, drawn together inside a ring labelled context: what the AI can read right now. Eight tools connect to that ring, each through its own door: meeting notes, a calendar, email, the support desk, accounting, a budget tool, maps and files. When a door is open its connection is drawn solid, the tool lights in its own colour and a segment of the ring lights to match, so the ring shows how much the AI can read at that moment, and the pair in the middle turns greener with every door that opens. Doors open one at a time and some close again, and the ring changes with them. Nobody else appears in the drawing: whatever the AI reads, it reads for this one person.
The half hour comes back.
The list arrives drafted. The lead edits and ticks instead of reading five messages and retyping three.
Two tools, no project.
Neither tool changed and nothing sits between them. Both already had the door, and one prompt used both.
Nobody learns a new screen.
It happens in the window they already had open. Nothing to install and nothing to be trained on, which is why it survived the first busy week.
For work that starts and ends with one person, that’s enough. It stops being enough the first time somebody else needs to know what happened.
Over time
Nine tools later, a chat window has become a process. Nobody wrote it down.
Two tools plugged into a chat window are a convenience. Nobody decides to have nine. One line reached two tools, then seven more arrived over two years, a door each, every one of them somebody solving a small problem on a Tuesday afternoon.
One chat window, two years apart
Two years agoThe prompt, one line at a time
Read yesterday’s customer messages, and open a job in the tracker for each one that is a bug.
Then check the billing system for anything unpaid on that account.
The shared drive
The calendar
The sales system
The finance spreadsheet
The team’s chat
The reporting dashboard
NowNine doors open
Where the decisions are keptNothing is recorded here
One line two years ago, eight lines now, nine doors in the margin, and nothing in the slot where the decisions were meant to be kept.One chat window. It opens with the prompt as it was written two years ago, one sentence long: read yesterday’s customer messages, and open a job in the tracker for each one that is a bug. It carries two door marks, one for each tool it reached. Seven lines have been added since, one at a time, and each one brought another door with it: check the billing system for anything unpaid, save the list to the shared drive, put a reminder in the calendar, look each customer up in the sales system, add refund amounts to the finance spreadsheet, post a summary in the team’s chat, and update the count on the reporting dashboard. Only the billing line is still written out in words. The other six are named by the tool they opened, and the rest of each instruction is drawn as grey bars, there to be seen rather than read. The margin ends with nine doors and the prompt ends at eight lines. A slot labelled where the decisions are kept stands empty, reading nothing is recorded here.
Nobody will be saying MCP in two years. They’ll still be asking who owns this and where it’s written down. Once more than one person depends on it, it has outgrown the chat window, and that’s a different kind of AI.
Common myths
Four myths about MCP.
You might think
“MCP is a product we can buy.”
Actually
There’s no invoice with MCP on it. You buy the tool, and its door either comes built the agreed way or it doesn’t. If it doesn’t, somebody can build one over the door that tool already has for software. That’s a job of days, not a purchase.
You might think
“Connecting a tool means the AI can see all of it.”
Actually
It gets the access you hand it, and you choose whose access that is. What it reads goes where a pasted email already goes: to the AI your team uses, on the terms you already have with them.
You might think
“This replaces the software we already run.”
Actually
It’s a second way into the same software. Every record stays where it is, the screens don’t change, and the people using them never see a difference.
You might think
“Our tools have it, so we’re connected.”
Actually
A door nobody has walked through for a particular job changes nothing. It’s the job that gets connected, not the tool.
Keeping it safe
Open one door at a time, and know who opened it.
Opening a door into a system you pay for isn’t the risky part. Losing track of how many are open is. None of these four is technical. Each is somebody being able to say what a door is for, whose permission it uses, and who reads the result before anything leaves.
01
One door, one job, one person who can say what it’s for.
If nobody can name the job in a sentence, the door is open for no reason. Ask that before you ask anything technical.
02
The door gets the asker's permission, not the builder's.
The AI reads as whoever is asking and sees what they’re allowed to see, not what the person who set it up could see. How that works is on the AI and your data page.
03
Nothing leaves the business without a person reading it first.
Drafting is the AI’s half. Sending, opening and approving stay with people, and anyone involved should be able to say which half they’re doing.
04
Keep a list of the doors that are open, and close the ones nobody uses.
One page, one owner, read once a quarter. The ninth door is never the one anybody remembers opening, and a door the vendor quietly rebuilt is still on the list and no longer working.
How to start
Name the job first. The connection is the last decision.
Start from the list nobody keeps: the things somebody on your team does by hand every morning that cross two tools. It will be longer than you expect. One job off it is enough to find out whether any of this is worth your time.
Day one
Write the job on one line.
What goes in, what has to come out, and who checks the result. If it takes a paragraph, it’s more than one job.
Week one
Ask whether both tools already have a second door.
Then open one, for that single job and nothing else. For whoever looks after your systems, that’s usually an afternoon, not a project.
Weeks two and three
Run it on real records, with a person approving every list.
Then count what came back, in minutes a morning. That’s the shape of an AI Jumpstart.
Next step
Stop guessing where AI belongs. We’ll show you.
The jobs worth connecting are rarely the ones anybody put on a plan. An AI Jumpstart starts there instead: which two tools somebody opens in the same hour, what has to come out, and who approves it. Two to four weeks, fixed scope, proven on your real records.