What Is an MCP Server? A Practical Guide for Businesses
The short answer to what is an MCP server: a small service that exposes selected data and actions from a business system to an AI assistant through one agreed protocol, with access controlled by permissions you set. For a Malaysian company running a legacy ERP, it is the controlled doorway between an assistant and the systems your staff query every day, added without replacing what already runs.
The real problem is not that AI cannot think. It is that it cannot reach your invoice tables, stock levels or customer records. Staff paste those into a chat window by hand instead, which is a data protection problem as much as a lost afternoon. An MCP server gives compatible assistants a standard way in, so you are not building a separate connection for every model and every vendor.
Key takeaways
MCP stands for Model Context Protocol, an open standard for connecting AI applications to external systems.
An MCP server controls which tools and data an assistant can access; it does not give an AI blanket access to your database.
A shared protocol can reduce repeated integration work, but compatibility and available features still vary between assistants.
Legacy ERPs can often be connected through an existing interface, a read replica or a thin service layer.
Start with read-only use cases, then add write actions only with identity checks, approval steps and an audit trail.
Table of contents
The problem MCP servers solve
AI models are trained on broad information, not your live business data. Without a connection, an assistant does not know what is in your ERP, warehouse system or internal procedures. Staff may copy records into a chat window to get an answer. That creates a data protection concern as well as wasted effort.
The older integration pattern is bespoke: build one connection between one system and one assistant, then repeat when you add another assistant or change vendors. Each integration creates maintenance work and another point where permissions and updates can be missed.
The Model Context Protocol (MCP) is an open standard for connecting AI applications with external systems. Its business value is reuse: one well-designed connection can serve multiple MCP-capable assistants, subject to each product’s supported features and configuration. You still need to test each client; “build once” does not mean every assistant behaves identically.

What an MCP server actually is
At its core it is a program that makes defined capabilities available to an AI application over the Model Context Protocol. It can run locally or remotely; common transports include stdio for a locally launched process and Streamable HTTP for remote connections. The server is not the assistant. It is the thing that handles the connection to your system, and decides what gets through.
It can expose three kinds of capability:
Primitive | Plain-language meaning | Example |
Tools | Specific actions an assistant can request | Look up a purchase order |
Resources | Data the application can retrieve as context | An approved SOP or a defined report |
Prompts | Reusable instruction templates a user can select | Summarise a supplier’s document history |
Think of it as a service counter with a published list of requests. The counter does not let a visitor wander through the stockroom. Likewise, a well-designed MCP server exposes a deliberate set of tools and resources, rather than a direct connection to everything in the database.
The server defines that capability surface, while the host application and identity policies determine who can connect. The server must also enforce permissions when a request arrives.
MCP server vs API vs custom plugin
MCP is not a replacement for APIs. It is a standard way for AI applications to discover and use capabilities that a server provides. A business can use an existing API behind an MCP server, or use a service layer to connect to a system with no suitable API.
Approach | What it takes to build | Maintenance | Assistant compatibility | Access control |
Direct API integration | A tailored connection for each assistant and system | Changes can require updates in multiple integrations | Usually tied to the specific client or application | Designed into each integration |
Custom plugin or connector | A client-specific extension and system connection | Must follow that client’s interface and changes | Usually designed for a particular platform | Depends on the plugin and platform |
MCP server | A server implementing the standard and connecting to the system | Shared connection can reduce duplicated work; the server still needs support | Can work with MCP-capable clients, subject to their implementation | Enforced by the server and the client’s identity setup |
MCP may not be the right choice if you need a single, simple, fixed integration and the existing API already meets the requirement. It is also not a shortcut around authentication, data modelling or integration testing. Our guide to connecting ChatGPT to internal systems explains the wider integration choices.
How an MCP server connects to a legacy ERP
For a company using SAP ECC, Dynamics NAV/AX, JDE or IBM i/AS400, the ERP can remain the system of record. The MCP server sits between the assistant and the ERP: it receives an allowed request, retrieves or processes the relevant information through a controlled connection, and returns only the required result.
The integration route depends on the system and its configuration. Options may include ODBC/JDBC views, SOAP services, SAP RFC/BAPI interfaces, stored procedures, a read replica or a thin service layer. We first check what is supported, what must remain stable, and whether the connection should be read-only.

A sound MCP server for ERP does not expose the whole database. It offers narrow, business-named operations with validated inputs and outputs. That keeps the assistant away from raw tables, limits the blast radius of a mistake and gives engineers a clear place to apply business rules.
This is where legacy system integration with AI can be practical: add a controlled AI interface to a system you need to keep, rather than re-platforming just to make it accessible. If SAP ECC is in your estate, SAP’s official maintenance strategy describes mainstream maintenance for the latest ERP 6.0 enhancement packages through the end of 2027. An AI connection does not replace your ERP support or migration planning.
What to expose first
Start with low-risk questions that take staff time to answer and have a clear source of truth:
Look up outstanding invoices, return a customer’s open invoices and due dates for Finance.
Check purchase order receipt status: show ordered, received and outstanding quantities for Procurement or Operations.
Search approved procedures, retrieve the current SOP or policy for a staff question.
Check stock across warehouses, give planners and sales teams a read-only stock view.
Find a supplier’s document history: retrieve approved, relevant records for Procurement or Compliance.
Keep these tools read-only at first. It is easier to confirm accuracy, permissions and business value before allowing an assistant to change records. Add a write action only when its rules, exception handling and approval step are clear.

Security, governance and Malaysian requirements
MCP is a protocol, not a security guarantee. A governed deployment needs controls around the server, the assistant and the connection to business systems.
Authenticate users individually. Pass or verify user identity where the client supports it; avoid relying on one shared service account that grants everyone the same access.
Apply least privilege. Scope each tool to necessary functions and fields. Deny unneeded actions by default.
Gate writes. Require a human approval step for consequential updates, and validate requests against business rules before they reach the ERP.
Keep an audit trail. Record which user and assistant requested what, the tool called, the result, and whether an approval was given. Protect logs from unauthorised access.
Treat returned content as untrusted. A document can contain prompt-injection text that attempts to influence an assistant. Retrieval should not grant the document authority to change system rules or permissions.
Set rate limits and monitor failures. Track usage, unusual requests, errors and changes in accuracy or performance (drift).
For Malaysian organisations, connect the controls to actual obligations. The Personal Data Protection Department (JPDP) publishes information on the 2024 PDPA amendments, with requirements including DPO appointment, breach notification and cross-border transfer considerations. A data protection impact assessment and privacy-by-design review can help establish how personal data is handled in the proposed workflow.
The Inland Revenue Board of Malaysia (LHDN) publishes e-Invoice guidance. E-invoicing does not make an MCP integration compliant by itself; keep the underlying records, supporting documents and access logs needed for your own controls and tax processes.
For financial institutions, Bank Negara Malaysia’s RMiT policy and its discussion paper on AI in the financial sector are relevant references for technology risk and AI oversight. The Cyber Security Act 2024, administered by NACSA, has specific obligations for designated National Critical Information Infrastructure entities; it does not apply identically to every business. For a fuller overview, see our post on AI regulations in Malaysia and data governance for AI.
What it costs and how long it takes
The effort depends on the ERP interface, data quality, identity setup, number of tools, security requirements and testing. A first server can be modest if the system has a supported interface and the scope is read-only. Work grows when data is spread across multiple systems or business rules are undocumented.
For context, businesses often compare this work with one or two reporting analysts at RM60,000–90,000 each per year, fully loaded, or a data-warehouse/BI rebuild that may cost RM300,000–1 million. A broader first workflow build with connectors and 12 months of managed service may fall around RM150,000–200,000 in year one, with RM100,000–120,000 for setup and connectors. These are indicative ranges, not a quote or a price for an MCP server alone.
How Sentient Concepts approaches the work
We start by choosing a measurable workflow and checking the system interfaces, data access and risks. Then we design the server’s tools, identity controls, approval steps and audit trail; build and test the connection; and monitor it in production.
Our MCP server development is part of our broader Integration & Architecture work. The full services offering spans AI Strategy & Roadmap, Readiness & Data Diligence, AI & GenAI Solutions, Deployment & MLOps, Monitoring & Evaluation and Managed AI Operations. Our Advise–Build–Run model keeps the whole arc with one team, reducing handoffs between planning, engineering and operations.
For sector context, see our pages on Manufacturing and Supply Chain & Logistics. You can also read our guide to AI agents for business. If you have a workflow in mind, start a conversation about what it would take to connect it safely.
Our perspective
“An MCP server is useful when it gives an assistant a narrow, well-governed path to a real business task. The protocol matters, but so do the permissions, evidence and operating controls around it. We start with the workflow and build the connection to fit.” Thomas Samuel
FAQ
What is an MCP server, and is it the same as an API?
An MCP server exposes defined tools, resources and prompts through the Model Context Protocol. An API is an interface a system can provide; an MCP server may use that API behind the scenes, but it is not the same thing.
Do we need an MCP server if our ERP already has APIs?
Not always. If a direct integration is simple and will serve one fixed application, it may be enough. MCP becomes more useful when you want a standard connection that multiple compatible AI clients can use.
Which AI assistants can use an MCP server?
Several AI tools support MCP, including Claude, ChatGPT and Microsoft Copilot-related products. The exact connection method and supported MCP features differ by product, plan and deployment, so confirm compatibility before building.
Is it safe to let an AI read from our ERP?
It can be designed with appropriate controls, but access should be limited to specific data and users. Use authentication, least privilege, read-only access initially, monitoring and an audit trail; do not expose direct database access to an assistant.
Sources
Model Context Protocol: What is MCP?, Model Context Protocol project
Understanding MCP servers, Model Context Protocol project
Model Context Protocol introduction: Anthropic
Building MCP servers for plugins and API integrations: OpenAI
Extend your agent with Model Context Protocol: Microsoft Learn
Personal Data Protection (Amendment) Act 2024: Personal Data Protection Department, Malaysia
e-Invoice guidelines: Inland Revenue Board of Malaysia (LHDN)
Revised Policy Document on Risk Management in Technology (RMiT): Bank Negara Malaysia
Cyber Security Act 2024: National Cyber Security Agency (NACSA)