What is MCP (Model Context Protocol) and when does it matter for your company

Sources: anthropic.com/news/model-context-protocol and the Agentic AI Foundation donation announcement (2025-12-09).

In this article

How MCP works

1. 1UserAsks in the chat
2. 2AI appMCP client
3. 3MCP connectorMCP server: exposes data and actions
4. 4Your systemCore, CRM, documents
5. 5AnswerWithin permissions

10,000+

active public MCP servers as of December 2025. That month Anthropic donated the standard to the Agentic AI Foundation, under the Linux Foundation.

Example at a bank

Example: a bank with 3 systems (core, CRM and document manager) and 3 AI apps (internal chat, collections assistant and complaints assistant). Custom-built: 3 × 3 = 9 integrations. With MCP: 3 connectors and 3 apps = 6 pieces.

Example: adding the third AI app

Custom

1. 1New appComplaints assistant
2. 23 new integrations+24 days (3 × 8, example)
3. 39 integrationsEach system change touches 3

With MCP

1. 1New appComplaints assistant
2. 2Plugs into the 3 connectors+1 day (example)
3. 36 piecesOne connector per system

Custom integrations or MCP: days of work and cost

Scope

AI apps or assistants

3

Systems to connect (core, CRM, documents…)

3

Effort (example)

Days per custom integration (example)

8 days

Days per MCP server (example)

10 days

Maintenance days per piece per year (example)

3 days

Technical team cost per day (example)

USD 200

Custom: 99 daysWith MCP: 51 days

First year: build + maintainestimated

• Custom99 days · USD 19,8009 pieces · build 72 days + maintain 27 days per year
• With MCPless work51 days · USD 10,2006 pieces · build 33 days + maintain 18 days per year

Savings in year one48 days · USD 9,600

Savings each following year9 days · USD 1,800

MCP recovers its extra build cost infrom the start: building with MCP already costs less

MCP pays off: each system is connected once and every app reuses it.

Method: custom = apps × systems integrations; with MCP = one connector per system + one connection per app (1 day each, example). Maintenance = pieces × days per piece per year. Days and day cost are examples: ask your team or provider for their figures. With the starting values (3 apps, 3 systems): custom 72 + 27 = 99 days (USD 19,800); with MCP 33 + 18 = 51 days (USD 10,200); savings in year one 48 days (USD 9,600), and MCP has no extra cost to recover: building it already takes 39 fewer days. Payback = extra MCP build days ÷ maintenance days it saves per month; with 1 app and 1 system, MCP costs more to build and to maintain, so it never pays back.

MCP or a direct connection

Criterion	Direct connection (API)	MCP
Fits when	One AI app and one system	Several apps or several systems
Pieces to maintain	apps × systems	apps + systems
If you change app or model	Rebuild the integration	The new app reuses the same connectors
Permissions	Defined in each integration	Defined once on the connector
Governed by	Your team	Open standard at the Linux Foundation
Upfront effort	Lower	One connector per system

Fits when

Direct connection (API)

One AI app and one system

MCP

Several apps or several systems

Pieces to maintain

Direct connection (API)

apps × systems

MCP

apps + systems

If you change app or model

Direct connection (API)

Rebuild the integration

MCP

The new app reuses the same connectors

Permissions

Direct connection (API)

Defined in each integration

MCP

Defined once on the connector

Governed by

Direct connection (API)

Your team

MCP

Open standard at the Linux Foundation

Upfront effort

Direct connection (API)

Lower

MCP

One connector per system

Security before you connect

Before connecting a system through MCP

Copy checklist

• Start with a read-only connector over a single data source.
• Use a service account with minimum permissions, not a person's account.
• Every action that writes (payments, data changes) requires a person's confirmation.
• Log every call: who asked, which tool was used and what it returned.
• Mask ID (cédula) and account numbers when the task does not need them: it is the data-minimization principle of Law 7593/2025 (PY) and the LOPDP (EC); for banks, confirm it against what the BCP's SIB or the Superintendencia de Bancos requires.
• Review the code of any third-party MCP connector before installing it.
• Confirm with compliance where the connector runs and which data leaves your network or the country.

The rules behind the checklist:

• Paraguay: Law 7593/2025: only strictly necessary data (art. 4, minimization) and limits on sending it to another country (art. 19).
• Ecuador: LOPDP (Registro Oficial 459, 2021-05-26): also requires minimization and regulates sending data abroad.
• Banks: the BCP's SIB (PY) or the Superintendencia de Bancos (EC).
• By country: AI for banks and fintech in Paraguay · AI for banking and cooperatives in Ecuador.

Example MCP pilot, with deliveries every 2 weeks

1. Week 1-2Read-only connectorOne source the team already uses, minimum permissions and logging.KPI: queries solved without escalation
2. Week 3-4Connected to the internal assistantOne team uses it with real questions.KPI: % correct answers
3. Week 5-6DecisionAdd a second system or one action with confirmation.KPI: time per query vs. today