Op-ed · Codito Blog

MCP: the Protocol That Plugs AI Into Your Tools

The Model Context Protocol turns plugging AI into your software from custom development into a standard connection. What it changes for cost, lead time and lock-in — and the three questions to ask your provider.

EEmile Chalmé
3 min read
MCPAgents IAIntégration

MCP, short for Model Context Protocol, is an open standard published by Anthropic in late 2024 that describes how an AI connects to a tool or a data source. For a business leader, the point fits in one sentence: this protocol turns plugging AI into your software from custom development into a standard connection. What used to cost weeks per tool becomes a reusable building block.

The topic looks technical. It is not, because it determines three very concrete things: how much it costs to connect AI to your data, how quickly you can switch tools, and who owns the integration you are paying for.

The problem MCP solves

An AI assistant with no access to your data knows nothing about your company. It knows neither your prices, nor your customers, nor your documents. To become useful, it has to be connected: to the CRM, to email, to the file server, to the ERP, to the business database.

Until now, every connection was a bespoke piece of development: one model, one tool, one connector. Ten tools, ten connectors, and as many to maintain whenever an interface changes. This is the well-known M × N problem: as many integrations as models multiplied by tools.

MCP replaces that multiplication with an addition. The tool exposes its capabilities once, according to the standard; any compatible model connects to it. You move from M × N integrations to M + N.

What changes, in plain business terms

TopicBeforeWith MCP
Cost of connectingOne development per tool and per modelOne server per tool, reusable by every model
Switching modelsRebuild the connectorsConnections stay, you swap the model
Lead timeWeeks per integrationDays when a server already exists for the tool
Lock-inIntegrations proprietary to the vendorOpen standard, another team can take over
ControlHard to auditExposed capabilities are declared and listable

What it looks like in a mid-sized company

A real example, from a client of our construction division: GAEA Géotechnique, a soil engineering firm that prepares 20 to 25 quotes a week. Every quote means analysing a plot and cross-referencing public soil data. The agent we built queries those public sources through the MCP protocol, pre-fills the borehole layout and the quote template. Human validation remains systematic before anything is sent to the client.

The value is not the gadget: it is that the data connection is a separate building block from the agent. If the model changes tomorrow, the agent changes; the connection stays.

Three questions to ask your provider

  1. Are the connections to our tools MCP-compliant, or proprietary? The answer determines what you can take elsewhere. A proprietary integration ties you to the vendor who wrote it.
  2. Exactly which actions are exposed to the AI, and who approves them? An MCP server declares its capabilities: read, write, delete. A good provider shows you that list, restricts it to the strict minimum, and routes sensitive actions through human approval.
  3. Who owns the code and where do the servers run? The protocol says nothing about hosting: your contract does. Code ownership and data location remain open questions you must settle.

What MCP does not solve

Three illusions to drop. MCP does not clean your data: a half-filled CRM stays half-filled, and the AI plugged into it will answer half-way. It does not define your access rights: deciding who sees what is your job, the protocol merely executes. And it does not replace the framing work: knowing which task to automate first remains the decision that drives the return on investment — that is the purpose of our AI audit.

One sign the standard is taking hold: business software vendors are starting to publish their own MCP servers. Our CTO analysed Retool's move in this article.

Where to start

  1. List the three tools that hold the information your teams need every day.
  2. Check whether an MCP server already exists for each one — increasingly, it does.
  3. Start with a read-only connection on a single tool, and measure the gain before opening up write access.

We build this kind of connection as part of our custom development offer, from €3,000 excl. VAT, and we roll it out progressively within the AI Partnership. To discuss it, book a 30-minute call.

Back

Frequently asked questions

What is MCP (Model Context Protocol)?

MCP is an open standard published by Anthropic in late 2024 that describes how an artificial intelligence connects to a tool or data source. The tool exposes its capabilities once, following the standard, and any compatible model can connect to it, instead of a connector developed for each model-tool pair.

Why should a business leader care about MCP, not just a developer?

Because it changes three concrete things: the cost of connecting AI to company data, the lead time, and vendor lock-in. An integration built on an open standard can be taken over by another team; a proprietary integration ties you to whoever wrote it.

Does MCP create a security problem?

The protocol explicitly declares the actions exposed to the AI (read, write, delete), which makes them auditable. Security depends on what you choose to expose, on human approval for sensitive actions, and on where the servers run — matters settled by your contract, not by the protocol.

Should we wait for MCP to become universal before connecting AI to our tools?

No. The right approach is to start with a read-only connection on a single tool, measure the gain, then extend. Waiting costs more than starting small, provided you do not open write access before access rights are framed.