Model Context Protocol (MCP) Explained: How to Connect AI Agents to Your ERP, CRM and Internal Systems

Short answer: the Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024 and now supported across the industry, that defines how AI applications and agents connect to external tools and data. Instead of writing a custom integration for every model and every system, you expose a system once as an MCP server, and any MCP-compatible assistant or agent can use it, within the permissions you define.

For companies building agentic AI, MCP has become the practical way to give agents safe, structured access to ERP, CRM, document repositories and internal APIs. This article explains how it works and how we deploy it at Code4Nord.

Key takeaways

  • MCP standardises the connection between AI agents and business systems, replacing one-off integrations.
  • An MCP server exposes tools (actions), resources (data) and prompts (templates) to AI clients.
  • The protocol does not make an integration secure by itself: scopes, authentication, approvals and logging are your responsibility.
  • Start with read-only servers, then add write actions with human approval.

What problem does MCP solve?

Before MCP, connecting an AI assistant to a business system meant writing glue code for a specific model provider, a specific framework and a specific API. Ten AI applications and twenty systems could mean dozens of fragile integrations. MCP turns this into a standard plug: build the connector once, reuse it everywhere.

This matters even more for agents. An agent that plans and executes multi-step tasks needs many tools, consistent descriptions of what each tool does, and predictable error handling. MCP provides that structure.

How MCP works

MCP uses a client-server model:

  • MCP host and client: the AI application, for example an assistant, an IDE, a custom agent built with an agent SDK, or a gateway such as OpenClaw. The client discovers what a server offers and calls it on behalf of the model.
  • MCP server: a small service in front of a system such as your ERP. It declares the available tools, their inputs and outputs, and the data resources it can share.
  • Transport: local servers usually communicate over standard input and output, while remote servers use HTTP-based transport with authentication, typically OAuth.

When a user asks an agent to “check stock for this product and prepare an order”, the model sees the tool descriptions, decides which tools to call, and the client executes the calls through the MCP server. The server enforces what is allowed and returns structured results.

MCP vs. classic API integration vs. plugins

Classic API integrationVendor-specific pluginsMCP
Reuse across AI applicationsLowLimited to one platformHigh
Tool discovery by the modelManualPlatform-specificBuilt into the protocol
Effort to add a new systemHighMediumMedium, once per system
Security modelCustomPlatform-definedDefined by you, on the server

Typical enterprise MCP servers

  • ERP: read stock, prices and order status; create draft orders that a person approves.
  • CRM: search accounts, log activities, create leads.
  • Knowledge: search SharePoint, Confluence or file shares with the user’s own permissions.
  • Database: read-only analytical queries on approved views.
  • Ticketing: create and update issues in Jira or ServiceNow.

Security: what MCP does not do for you

MCP defines how to connect, not what is safe to allow. In our projects, every MCP deployment follows a few rules:

  1. Least privilege: each server exposes only the tools a use case needs, with scopes per agent.
  2. User-bound access: where possible, the agent acts with the permissions of the user it serves, not with an all-powerful service account.
  3. Approval for writes: creating, changing or deleting data requires human approval until the workflow is proven.
  4. Prompt injection awareness: content returned from documents, e-mails or web pages is treated as data, never as instructions.
  5. Full audit trail: every tool call is logged with the agent, the user, the inputs and the result.
  6. Vetted servers only: third-party MCP servers are reviewed like any other dependency before they are installed.

A pragmatic rollout plan

  1. Pick one workflow with clear value, for example answering sales questions about stock and delivery times.
  2. Build a read-only MCP server for the systems involved and test it with real questions.
  3. Add evaluation: a set of test questions with expected answers, run before every change.
  4. Introduce write actions behind approval, measure accuracy and time saved.
  5. Reuse the same servers for the next agent instead of building new integrations.

How Code4Nord helps

Our Agentic AI & Cloud team designs and builds MCP servers, agent workflows and the security and observability around them. In the OnOptic case study we used MCP to connect AI agents to ERP, CRM, booking and catalogue systems with human approval for sensitive actions. If your data and systems still need groundwork, our Data & AI Engineering team can help prepare them.

Frequently asked questions

What is the Model Context Protocol (MCP)?

MCP is an open protocol that standardises how AI applications and agents connect to external tools, data sources and prompts, so that one connector can be reused by many AI clients.

Who created MCP?

MCP was introduced by Anthropic in November 2024 as an open standard and has since been adopted by many AI platforms, developer tools and agent frameworks.

Is MCP secure?

MCP supports authentication and structured permissions, but security depends on how servers are designed and deployed. Use least-privilege scopes, user-bound access, approval for write actions and full logging.

Can MCP work with open-source and local models?

Yes. MCP is model-agnostic. Any client that supports the protocol can use MCP servers, including agents that run on open-weight models hosted in your own environment.

How long does it take to build an MCP server for an ERP?

A focused, read-only server for a few well-defined operations can often be built and tested in one to three weeks, depending on the quality of the ERP’s APIs and the security requirements.