Model Context Protocol (MCP) is an open standard that defines how an AI model connects to external tools — CRM, ERP, databases, internal files — in a structured and secure way. Anthropic announced it in November 2024, and it is now governed by the Linux Foundation's Agentic AI Foundation, with backing from OpenAI, Google, Microsoft, and AWS. The official analogy in the spec sums it up well: MCP is the "USB-C for AI applications."
Only 20.3% of Spanish companies actively use AI, according to Eurostat and INE data from 2025. The number-one barrier, as identified by the Banco de España, is not cost — it is the lack of knowledge about what to build and how to connect it to existing systems. MCP is often the missing piece.
What problem does MCP actually solve?
Before MCP, connecting an AI model to an internal tool meant writing a bespoke integration for each combination. Three AI models and ten internal tools meant up to thirty custom connectors to build and maintain. MCP reduces that problem from M × N connectors down to M + N: each tool publishes one MCP server, and any compatible model can connect to it.
The protocol defines three types of elements an MCP server can expose:
- Resources: read-only data (a document, a customer record, a report).
- Tools: executable actions (create a task, query the CRM, send an email).
- Prompt templates: reusable instruction structures to guide the model on specific tasks.
What does an MCP look like in a real business?
The most common scenario: your data is spread across several systems — CRM, ERP, internal documentation — and the only way the AI could use it was if someone copied it into the chat manually. With an MCP server, the AI agent can query those systems directly.
Concrete examples already in production:
- CRM (Salesforce, HubSpot): the agent looks up a customer's history, drafts a personalised follow-up email, and logs it in the CRM — all in one flow. HubSpot shipped its MCP server with over 100 tools in production in April 2026.
- ERP (Microsoft Dynamics 365): the agent queries financial KPIs or generates journal entries directly from a conversation.
- Internal database (PostgreSQL, etc.): the agent runs structured queries and returns results in context, without anyone writing SQL by hand.
- Internal documentation: combined with RAG (Retrieval-Augmented Generation — a semantic search layer over your documents), the agent locates the right information and cites it accurately.
César García and Smart Growth build MCP servers on top of existing internal systems as a dedicated service: MCPs on internal APIs.
API vs MCP: what is the actual difference?
This is where most technical teams get confused. The distinction is not just technical — it is architectural.
| Feature | Traditional API | MCP server |
|---|---|---|
| Who consumes it | A developer writing code | An AI agent |
| When is the logic decided | At design time (hard-coded) | At runtime (dynamic) |
| Session state | Stateless (each call is independent) | Stateful (the agent maintains context) |
| Reasoning | None (does exactly what it was programmed to do) | The agent decides which tool to use based on context |
A traditional API executes exactly what a developer wrote. An agent with MCP can reason: if a user asks for a summary of client X, the agent decides on its own to first query the CRM, then look up related emails, then draft the summary. No one programmed that sequence step by step.
For a deeper look at when to build an MCP server versus a conventional integration, see how to turn your internal APIs into MCP servers.
Is it safe to expose internal systems to AI via MCP?
Since March 2025, the MCP standard requires OAuth 2.1 with PKCE for all remote servers. Your system credentials never reach the AI model — they stay on the MCP server, which acts as a controlled intermediary. The model only sees the result of an operation, not the keys that made it possible.
The protocol also allows you to set permissions per tool. You can configure the agent to have read-only access to the CRM, or only write access to a specific database table. Every operation is logged and auditable.
One documented supply-chain incident is worth noting: in September 2025, an unofficial third-party MCP server for the Postmark platform — with 1,500 weekly downloads — was modified to silently copy emails to an attacker. The lesson: only use official or internally-built MCP servers.
For companies handling sensitive data — professional services firms, fintechs, family offices — the most secure approach is to deploy MCP servers on private infrastructure, so data never leaves the company's own perimeter. This is the architecture César García applies for clients with strict data-sovereignty requirements.
Where do I start if I want to use MCPs?
The first step is not technical. It is identifying which internal data source or tool, if the AI could access it automatically, would save your team the most time each week. That question drives the design of any useful MCP server.
The typical journey has three phases:
- Diagnostic: identify which systems have the most impact as context sources for the AI.
- MCP server build: expose the relevant tools and data in a controlled, permissioned way.
- Agent connection: configure the agent — this could be Enclave, Smart Growth's private enterprise ChatGPT, or another compatible model — to use those resources.
MCP-compatible models now include Claude (Anthropic), ChatGPT (since March 2025), GitHub Copilot, VS Code, and Cursor. The official MCP Registry passed 2,000 entries in September 2025, signalling that adoption has moved well beyond the experimental phase.
If you are unsure whether a full AI agent with MCP is what your business needs or whether a simpler chatbot would do, the article chatbot vs AI agent: when is it worth it? helps frame that decision.
In summary
MCP is the standard that makes it possible for AI agents to work with your real business data and internal tools — without bespoke connectors for every combination. It is secure when configured correctly, backed by the major players in the industry, and already running in production across businesses of all sizes.
If you want to know whether it fits your situation and where to begin, César García and Smart Growth can help you assess the starting point.