Imagine you have just hired a highly capable assistant. They understand everything you explain, reason well and write without errors. The problem is that they cannot open your CRM, they do not know how to query the ERP and they have no access to the order history. For them to be truly useful, someone has to build a bridge to each system, one by one. That has been the bottleneck of enterprise AI until very recently. The Model Context Protocol (MCP) is the solution the industry agreed on to fix it: an open standard that acts as a universal connector between artificial intelligence models and the tools, databases and APIs you already use every day.
What is the Model Context Protocol (MCP) and where does it come from?
The MCP is an open-source communication protocol published by Anthropic in November 2024. Its goal is to standardise the way a language model (LLM) connects to external data sources and executes actions in real systems. Before MCP, every AI vendor and every company that wanted to integrate an assistant with its tools had to build its own connectors: specific code for the CRM, specific code for the ERP, specific code for the document management system. Every model or tool change meant rewriting those integrations.
The analogy that circulates most widely in the technical community is that of USB-C: before that standard existed, every manufacturer had its own connector. Today any USB-C cable works with any compatible device. MCP does the same for AI: an MCP server that exposes your CRM works with any model that speaks MCP, whether that is the model powering your internal copilot, the agent that processes invoices or the support chatbot.
The protocol is built on JSON-RPC 2.0 and defines three types of capabilities that a server can offer to an AI client:
- Tools: functions the model can invoke, such as querying a CRM record, creating a ticket, running a SQL query or sending an email.
- Resources: data the model can read to enrich its context, such as documents, database records or API responses.
- Prompts (templates): predefined instructions the server suggests to the model to guide its responses in specific contexts.
The architecture is client-server: the MCP host (the environment where the model runs, for example your copilot or your n8n agent) acts as the client; the MCP servers are the adapters installed alongside each tool or system you want to expose. The model never accesses your database directly; it always does so through the corresponding MCP server, which enforces the access rules you define.
Industry adoption: no longer Anthropic-only
What makes MCP a genuine standard rather than a single-vendor proposal is the speed at which the rest of the ecosystem adopted it. OpenAI added native MCP support in March 2025, just four months after the protocol was published. Google followed in May 2025. By mid-2026, all the major model providers — Anthropic, OpenAI, Google, Microsoft, Meta — support MCP in their agent platforms.
Adoption numbers are significant: by March 2026, the MCP SDK had accumulated more than 97 million monthly downloads across its Python and TypeScript versions, with more than 10,000 public MCP servers available in the ecosystem and more than 950 dedicated to business applications. Companies such as HubSpot launched their MCP server in production in April 2026 with more than 100 tools available to their users.
Automation frameworks also embraced the standard: n8n incorporated native MCP nodes in 2025, LangGraph and LlamaIndex have first-class integration, and the most widely used enterprise agent platforms in Spain include it as a default tool layer.
How MCP works in practice: the step-by-step flow
To understand the protocol without getting lost in technicalities, it is worth following the flow of a concrete interaction. Suppose your company has deployed an AI agent for the sales team. A salesperson asks: «What pending orders does customer García Hermanos have?»
- The MCP host (the agent) receives the question and decides it needs to query the CRM.
- It sends the MCP server of the CRM a JSON-RPC call with the tool name (
get_customer_orders) and the relevant parameters (customer name, status «pending»). - The MCP server authenticates the request, verifies that the agent has permission to read orders and executes the query against the CRM.
- It returns the structured result to the agent, which incorporates it into its context and formulates a natural-language response.
- The salesperson sees: «García Hermanos has 3 orders pending delivery: ref. PED-2241 (€15,400, estimated delivery 10 Jun), ref. PED-2267 (€8,200, delivery 18 Jun) and ref. PED-2289 (€21,700, no confirmed date).»
The entire cycle happens in seconds. The model never had direct access to the CRM database; the MCP server acted as a controlled intermediary. If the company changes its CRM tomorrow, only the MCP server for that system needs updating; the agent and all other integrations remain untouched.
Comparison: MCP versus earlier approaches
| Criterion | Ad hoc integration (pre-MCP) | Classic function calling | MCP |
|---|---|---|---|
| Reusability | Zero: unique code per model + tool | Low: functions must be redefined per model | High: one MCP server works with any compatible client |
| Maintenance | High: any tool change breaks the integration | Medium: function schemas must be updated | Low: the MCP server absorbs changes internally |
| Access control | Manual in each integration | Depends on the implementation | Centralised in the MCP server (OAuth, SSO on 2026 roadmap) |
| Auditability | Very difficult to unify | Partial | Unified log per MCP server |
| Multi-model support | No (LLM-specific code) | Limited (API differences between providers) | Yes (standard adopted by all major providers) |
| Time-to-integration | Weeks to months | Days to weeks | Hours to days (pre-built servers for the most common CRMs and ERPs) |
Why MCP matters to your business, not just your technical team
When an operations director or a general manager hears the word «protocol», the usual reaction is to delegate the decision to the IT department. With MCP that would be a mistake, because the decision to adopt it has direct implications for cost, deployment speed and data control.
First, development cost. Enterprise adoption analyses published in 2026 indicate that companies that used MCP to connect their systems to AI agents have reported significant reductions in integration costs compared with earlier approaches, and have cut time-to-deployment from months to weeks or days, depending on the tool.
Second, data control. With MCP, your company's sensitive data never leaves your infrastructure without explicit authorisation. The MCP server acts as a gatekeeper: it defines what the model can see and what it cannot. This is relevant in the context of the European AI Act (EU Regulation 2024/1689), which entered into force on 1 August 2024 and whose articles on high-risk AI systems require documented traceability and access control. A well-configured MCP server facilitates compliance by centralising the audit trail.
Third, vendor independence. If you build your integrations on MCP today, you can switch AI models without rewriting anything. That flexibility has real value in a market where models evolve every few months.
If you want to see how we implement MCP in real SME and mid-market projects, take a look at our MCP integration service for businesses, where we detail the process from tool analysis to production deployment.
Real-world use cases in SMEs and mid-sized companies
MCP is not only for large corporations with AI engineering teams. The most common use cases we encounter in Spanish SMEs in 2025-2026 are the following:
Sales agent connected to the CRM and ERP
The agent queries the purchase history, credit status and open orders without the salesperson having to switch screens. It reduces meeting and call preparation time by between 30 and 50 % across the projects we have supported.
Support assistant with access to the knowledge base and ticketing system
The assistant queries the internal knowledge base (RAG), automatically opens tickets when it cannot resolve the query and escalates to a human agent with the full conversation context. The result: fewer repeated tickets and shorter resolution times.
Finance copilot with access to accounting data
An agent with an MCP server pointing at the accounting software can answer questions such as «How much did we invoice to the public sector in the first quarter?» or «Which supplier invoices have been outstanding for more than 60 days?» directly from the chat, without exporting to Excel.
Document process automation
The agent extracts data from incoming invoices (via OCR or document processing), validates them against the purchase order in the ERP through an MCP server and records the approval or the discrepancy without human intervention. This type of flow can be built on n8n automation, where native MCP nodes facilitate the connection with management systems.
Security and governance: what you need to know before deploying
MCP solves the interoperability problem, but it does not solve the security problem on its own. There are several aspects every company must consider when deploying MCP servers in production:
- Permissions per tool and per role: define what each agent can invoke and on behalf of which users. Not all agents need write access; many use cases only require read access.
- Call auditing: log every tool invocation with a timestamp, parameters and result. In the context of the AI Act and the GDPR, traceability of automated decisions is a requirement, not an option.
- Input validation: MCP servers must validate the parameters they receive from the model before executing any action. A manipulated model or a prompt injection could attempt to pass malicious parameters.
- Robust authentication: the 2026 MCP roadmap points to OAuth 2.0 and SSO as standard authentication mechanisms between host and server. In the meantime, use service tokens with limited scope and periodic rotation.
The good news is that MCP's own design — with the server as a control layer interposed between the model and the data — facilitates applying these measures in a centralised way, without having to implement them in every integration separately.
The ecosystem in 2026: which MCP servers already exist
One of the advantages of MCP having become a standard is that a library of pre-built servers already exists for the most common tools. This dramatically reduces implementation time. Among the MCP servers available in 2026 for enterprise use, the following stand out:
- CRM: Salesforce, HubSpot (100+ tools in production), Pipedrive, Zoho CRM.
- ERP: SAP (selective modules), Odoo, Microsoft Dynamics 365.
- Project management: Jira, Asana, Linear, Notion.
- Communication: Slack, Microsoft Teams, Gmail.
- Databases: PostgreSQL, MySQL, MongoDB, Snowflake, BigQuery.
- Document storage: SharePoint, Google Drive, Confluence.
- Automation: n8n (native nodes since 2025), Zapier, Make.
For vertical or proprietary systems that have no published MCP server, the alternative is to build a custom server, which is still significantly faster and more maintainable than the ad hoc integrations that preceded the protocol.
Frequently asked questions
Do I need programming skills to use MCP in my company?
Not directly. If you use standard tools (HubSpot, Salesforce, Notion, Jira, Google Drive…) there is likely a pre-built MCP server you can configure without writing code. For integrations with proprietary systems or vertical ERPs, development is needed, which is typically handled by the technology partner. What you do need to understand as a business owner or manager is what data you are exposing, to which agents and with what permissions: that is a governance decision, not just a technical one.
Is MCP the same as OpenAI's function calling or other models' tool_use?
Not exactly. Function calling is a mechanism internal to the model for invoking functions defined in the API call. MCP is a layer above that: a standardised protocol that allows those functions (tools) to be defined in a separate, reusable server that can be shared between different models and agents. In practice, models that support MCP still use function calling internally, but the tool definitions live in the MCP server rather than in each application's code.
How does MCP affect compliance with the AI Act and the GDPR?
MCP facilitates compliance because it centralises access control and auditing. The EU AI Regulation (EU Regulation 2024/1689, in gradual application since August 2024) requires traceability in high-risk AI systems, and the GDPR requires that access to personal data be documented and justified. A well-configured MCP server logs every tool call with its parameters and result, making it easier to demonstrate to an auditor what the system did with which data and at what point in time. It is not a compliance solution in itself, but it is a piece that fits well within a responsible AI architecture.
How long does it take to integrate MCP in a mid-sized company?
It depends on the number of systems to connect and whether pre-built MCP servers exist for them. A typical MCP integration project for an SME with three or four standard tools (CRM, ERP, document management and ticketing) can be in production in two to six weeks, including permission configuration, security testing and team training. For proprietary systems or more complex integrations, timelines are longer. At Summum IA we accompany this process from the initial analysis through to deployment and ongoing monitoring.