Guide · AI agents Model Context Protocol (MCP) for Business: Uses and Limitations

What MCP is

MCP stands for Model Context Protocol. According to the provider's website, as of October 2026, it is an open-source standard for connecting AI applications to external systems. These include data sources such as local files and databases, tools such as search engines and calculators, and workflows made available through specific prompts. The official documentation compares MCP to a USB-C port for AI applications. Check these details with the provider.

For business leaders, the function matters more than the technical name: an AI application can use information or tools from another system if a suitable connection has been set up. Whether it is allowed to do so, and for which task, must be defined for each project. MCP describes the interface. It does not decide which systems should be accessible.

Anthropic introduced MCP as an open standard in November 2024, according to its website as of October 2026. The documentation lists version 2026-07-28. Versions can change, so check the current specification with the provider before implementation. Planning involves more than confirming that a connection is possible. You also need to establish which version the systems involved will use.

How MCP works: host, client and server

The host is the AI application you work in. It creates a separate client for each connected MCP server. This client is the part of the application that communicates with that server. The MCP server provides access to a data source, tool or workflow. These roles are described in the official architecture, according to the provider's website as of October 2026. Check these details with the provider.

A server can run locally on the same device or be accessed over a network. According to the provider's website, as of October 2026, local servers generally use STDIO transport and serve one client. Remote servers use Streamable HTTP and often serve several clients. Messages use JSON-RPC 2.0, a defined format for requests and responses. These terms matter for technical implementation. For your decision, the main questions are where the server runs, which systems it connects to and who is allowed to use it. Check the technical details with the provider.

According to the provider's website, as of October 2026, MCP is supported by AI assistants including Claude and ChatGPT, and development tools including Visual Studio Code and Cursor. This does not mean every connection you need already exists or will work without modification. Check the current capabilities with the respective providers and assess the specific integration your project needs. L&L Synergy FlexCo is neither a partner nor a reseller of the providers named here.

Where MCP can help a business

The official documentation gives two examples, according to the provider's website as of October 2026: an agent accesses Google Calendar and Notion; and a business chatbot connects to several of an organisation's databases so users can analyse data through chat. Check these details and the requirements of the respective services with their providers. The examples illustrate two different tasks: connecting existing applications and making information from several data sources available to answer a question.

For your business, this leads to a practical question: does the proposed AI application need information or functions from systems you already use? If so, MCP may be one way to connect them. You should first define which data will actually be available. A chatbot that answers questions about business data does not automatically need access to every database. Likewise, an agent should only be given the functions it needs for its task.

We develop AI agents and automations that work with your systems. MCP can provide the connection, but it is not needed for every project. The decision depends on the task, the existing systems and the permissions required.

Security and permissions

An MCP server can provide access to data and functions. Reviewing permissions therefore belongs at the start of planning. According to the provider's website, as of October 2026, its Security Best Practices page covers confused deputy problems with MCP proxy servers, token passthrough, server-side request forgery (SSRF), compromised local MCP servers and scope minimisation, among other issues. In simpler terms, these concern misdirected access, the handling of credentials, unwanted network requests, tampered local servers and permissions that are too broad. Check the provider's current security guidance.

Our recommendations for a project are to use only servers from trusted sources, grant only the permissions needed for the task and log access so actions can be traced. Require approval for actions that change data. Test the connection in a test environment first. These are our own recommendations, not quotations from the documentation. For further considerations on handling business data, see AI and data protection.

What MCP is not

MCP is neither a language model nor a ready-made AI agent. According to the provider's website, as of October 2026, the standard defines how context is exchanged. It does not determine how an AI application uses language models or manages the context it receives. Check this distinction in the provider's current documentation. A working connection alone therefore says nothing about whether answers are correct, complete or suitable for a business decision.

MCP is not blanket permission to access all business data, either. Permissions, processes and checks must fit the specific use case. It should be clear which application uses which server and whether it can only read information or can also make changes.

When MCP is suitable and when it is not

MCP is worth considering when an AI application needs targeted access to existing data sources or tools and that access can be clearly limited. It is particularly worth assessing when a task involves several systems. Its value depends on the specific workflow, not on using a standard for its own sake.

MCP is less suitable when a task can be completed without an external connection or when access rights and responsibilities have not yet been clarified. A single, narrowly defined integration does not automatically need an MCP server either. Start with the business task, then decide on the technical interface.

How to proceed

1. Define the task: Describe the question or workflow the AI should support. Identify which systems it needs and whether read-only access is sufficient or changes are required.
2. Review the connection and permissions: Establish whether MCP makes sense for these systems. Check the source of the server, permissions, approvals and the current specification. Decide who is responsible for access.
3. Test on a limited basis: Try the workflow in a test environment. Check which data is accessible, whether approvals work and whether access can be traced. Then decide whether to proceed.

Questions about MCP in business