What MCP actually is, and why your business should care
A plain explanation of the Model Context Protocol, what it lets AI assistants do with your own systems, and when building an MCP server is worth it.
Every AI assistant has the same limitation: it only knows what it was trained on and what you paste into it. It cannot see your inventory, your tickets, your schedule or your database. So people copy and paste, which works until the data is large, private, or changes hourly.
The Model Context Protocol solves that. It is an open standard for connecting AI assistants to real systems, and it is worth understanding because it changes what “using AI at work” can mean.
The plain version
MCP is a common language between an AI assistant and your systems. Instead of every assistant needing custom integration code for every tool, you build one MCP server that describes what your system can do, and any MCP-capable client can use it.
The comparison people find useful is a USB port. Before a standard, every device needed its own connector. After one, anything plugs into anything. MCP is that, for AI and your software.
An MCP server exposes three kinds of thing:
- Tools, which are actions the assistant can take. Look up an order, create a ticket, run a query.
- Resources, which are data it can read. A document, a record, a file.
- Prompts, which are pre-built workflows for common requests.
What this looks like in practice
Concretely, for a business:
Someone asks the assistant “which jobs are behind schedule this week?” Without MCP, the assistant guesses or asks them to paste a spreadsheet. With an MCP server connected to your job tracker, it queries the actual data and answers with real numbers, this morning’s numbers.
Or a support agent asks “what did we tell this customer last time?” The assistant reads the ticket history through MCP and answers, with references, rather than inventing something plausible.
That last distinction is the whole point. An assistant with MCP access is answering from your data. An assistant without it is guessing. For anything that matters, that difference is not subtle.
When it is worth building
An MCP server earns its cost when all of these are true:
- Your team already uses an AI assistant daily. If nobody is using one, solve that first, or you are building a bridge to an empty road.
- The information lives in a system with an API or database. MCP is a connector, not a data-entry project. If the answers live in someone’s head or a filing cabinet, that is a different problem.
- People ask the same questions repeatedly. Recurring questions are what justify the build, the same way recurring manual work justifies automation.
- Being wrong is survivable, or a human approves the risky steps. Read-only access to reports is low risk. Anything that sends an invoice or emails a customer should have a person in the loop.
When it is not
Skip it if your data fits in a document you could paste. Skip it if only one person would use it. Skip it if the underlying system has no API and would need building first, in which case build the system properly and add MCP later.
And be sceptical of anyone selling an MCP server as a product without asking what questions your team actually asks. The protocol is the easy part. Knowing which tools to expose, and which to deliberately withhold, is the work.
The part most people get wrong
The instinct is to expose everything. Full database access, every action, all of it, and let the assistant work out what to use.
That is a bad idea for the same reason you do not give every employee production database credentials. An MCP server should expose the narrowest set of capabilities that answers the real questions, with:
- Read and write clearly separated, and writes restricted to what genuinely needs them.
- Approval gates on anything destructive or outward-facing.
- Logging of every call, so you can answer “what did it do and when.”
- Its own credentials, scoped to exactly what it needs, never a shared admin account.
Treat the assistant as a capable new employee on their first week: useful, worth trusting with real work, and not yet given the keys to everything.
Cost and effort
A focused MCP server over an existing API is a small project, usually one to three weeks. It can often run serverless for a few dollars a month, because it only does work when someone asks a question.
The cost grows when the underlying system has no usable API, when data needs cleaning first, or when write access means designing approval flows. Those are real projects, and worth scoping honestly rather than discovering midway.
Where to start
Pick one question your team asks constantly and cannot answer quickly. Build a read-only MCP server that answers exactly that. It will take days, it cannot break anything, and it tells you whether the rest is worth doing.
We build MCP servers and agent workflows for Canadian businesses, including public servers that run on a single serverless function for a few dollars a month. If you want a straight read on whether MCP fits your situation, tell us what your team keeps asking.
Related: when your business actually needs AI and our AI agents and MCP work.
Further reading
Topics in this article
- MCP 4
- AI agents 4
- claude
- Integration 2
Get new articles by email
One email when something new goes up, roughly twice a month. Plain writing on what software costs and what is worth building. No sequences, no sales calls, and one click to leave.
Keep reading on the same subject
-
What building an MCP server for your business involves
A practical walkthrough of scoping, building and running an MCP server: which tools to expose, what to withhold, and what it costs each month.
See what a build involves -
The guardrails an AI agent actually needs
Agents that take real actions need limits, not enthusiasm. The seven controls worth building, why each exists, and which ones you can skip.
Review the guardrails -
AI agents that do real work, not demos
What separates an AI agent that survives contact with a real business from an impressive demo, and the guardrails that make one safe to deploy.
See what agents can run