Skip to content
dhii.ai

MCP

Model Context Protocol (MCP) Explained for Business and Engineering Leaders

What the Model Context Protocol is, why it matters, how MCP servers work, real use cases, security considerations and how to decide whether your organisation or product needs one.

By Amit Valecha · · 6 min read

The Model Context Protocol (MCP) has quickly become one of the most important pieces of infrastructure in applied AI. If you lead a product or engineering team, you’ve probably heard it described as “USB-C for AI”. That analogy is roughly right, but it doesn’t explain why MCP matters for your organisation or what to do about it. This article does.

The problem MCP solves

AI assistants and agents become far more useful when they can access real data and take real actions: query a database, read a document store, create a ticket, check an order. Before MCP, each AI application needed its own custom integration for each system. Five AI tools and ten internal systems meant up to fifty integrations, each built and maintained separately.

MCP standardises the connection. A system exposes its capabilities once, through an MCP server. Any MCP client — a chat application, an IDE, an agent framework — can discover and use those capabilities. Build once, use from many places.

How MCP works, in plain terms

MCP defines a protocol between:

  • Hosts and clients — the AI application the user interacts with (for example, a desktop chat app, a coding assistant or your own agent), which contains an MCP client.
  • Servers — lightweight programs that expose capabilities from a system: your CRM, your file storage, your internal API, a database.

A server can offer three main kinds of capability:

  • Tools — actions the model can invoke, such as search_tickets or create_invoice_draft, each with a described input schema.
  • Resources — data the application can read and provide as context, such as files, records or documents.
  • Prompts — reusable, parameterised prompt templates for common tasks.

When a user asks the assistant something, the client tells the model which tools are available. The model decides to call one; the client sends the request to the server; the server executes it against the underlying system and returns the result; the model uses it to continue. Servers can run locally, communicating over standard input/output, or remotely over HTTP with authentication.

Why it matters

For internal teams, MCP means your staff can use AI assistants that actually know your systems. An analyst can ask for “accounts that churned last quarter with more than three support tickets” and get an answer drawn from your warehouse and helpdesk — with their own permissions.

For engineering teams, MCP reduces integration work and lock-in. One well-designed server works across multiple AI clients, and you can switch or add clients without rebuilding integrations.

For SaaS companies, MCP is becoming an expected integration surface, much like a public API. Customers increasingly want to use your product from their AI tools. An MCP server makes your product accessible to those assistants and agents.

For agent builders, MCP provides a standard way to give agents tools, including tools maintained by other teams or vendors.

Use cases we see

  • Internal knowledge and data access: assistants querying documentation, wikis, warehouses and CRMs.
  • Operations: creating and updating tickets, checking order and shipment status, drafting records for review.
  • Engineering productivity: giving coding assistants access to internal APIs, documentation, issue trackers and deployment status.
  • Customer-facing SaaS integrations: letting customers manage projects, query analytics or automate workflows from their AI assistant of choice.
  • Agent tooling: a shared tool layer for multiple internal agents.

Security: the part that deserves the most attention

Connecting AI to systems that hold real data and can take real actions raises real risks. A responsible MCP deployment addresses:

Authentication and authorisation. Remote MCP servers should authenticate users — the protocol’s specification describes OAuth-based authorisation for HTTP transports — and act with that user’s permissions, never with a shared all-powerful account. If a user can’t see a record in the CRM, the AI acting for them shouldn’t either.

Least privilege. Expose only the tools needed. Prefer read-only tools by default. Make write tools specific and bounded rather than offering generic “run any query” capabilities.

Confirmation for consequential actions. Clients can and should ask the user to confirm before tools with side effects run. Design write tools so they can be previewed.

Prompt injection. When a model reads content from documents, emails or web pages, that content may contain instructions designed to hijack it. Combined with powerful tools, this is the most serious practical risk. Mitigations include limiting which tools are available alongside untrusted content, requiring confirmation for data-exfiltrating actions and treating retrieved text strictly as data.

Supply chain. Third-party MCP servers run code and receive data. Vet them as you would any dependency: source, maintainer, permissions requested and update practices.

Audit logging. Log every tool call with user, arguments, results and timing, so you can investigate incidents and demonstrate control.

Designing a good MCP server

Technical correctness is the easy part. What makes a server genuinely useful is tool design:

  • Model-friendly tool names and descriptions that say what the tool does, when to use it and when not to.
  • Task-oriented tools that map to how a person would work, rather than mirroring every API endpoint.
  • Strict schemas and helpful errors that let the model correct its own mistakes.
  • Compact responses that return what’s relevant, not entire API payloads.
  • Testing with multiple clients, because models and clients behave differently.

We cover these patterns in depth in designing tools for AI agents.

Do you need an MCP server?

Consider building one if:

  • Your staff use AI assistants and would benefit from access to internal systems.
  • You’re building multiple agents that need the same tools.
  • You run a SaaS product and customers ask how to use it with AI tools.
  • You want to reduce lock-in to any single AI vendor’s plugin ecosystem.

You may not need one yet if your AI use is limited to a single application with a couple of integrations; a direct integration may be simpler. But as soon as more than one AI client or agent needs the same capabilities, MCP pays for itself.

Build, buy or adopt?

Many popular SaaS tools now publish official MCP servers, and open-source community servers exist for common systems. Before building, check whether a trustworthy server already exists for the system you want to connect. Official servers from the vendor are usually the best choice, because they track API changes and handle authentication properly.

Build your own when the system is internal, when existing servers expose far more capability than you’re comfortable granting, when you need tools shaped around your business processes rather than raw API endpoints, or when you need specific security, logging or data-residency behaviour. A thin custom server can also wrap an official one, exposing only a curated subset of its tools.

Whichever route you take, apply the same review: what can this server read, what can it change, whose permissions does it use, and where are its actions logged?

Measuring value

As with any integration, measure whether it is used and whether it helps. Track which tools are called, how often calls fail, which clients they come from and which tasks they support. Combine that with a few qualitative check-ins with users. Tools that are never called can be removed; tools that fail often usually need better descriptions or schemas. That feedback loop keeps the server small, safe and genuinely useful.

Getting started

A sensible first project is a read-only MCP server for one high-value system — your knowledge base, ticketing system or analytics warehouse — with proper authentication and logging. Measure how it’s used, gather feedback and add carefully designed write tools later.

Our MCP server and tool integration service covers design, security review, implementation and client rollout.

Services related to this article

Keep reading

Start a project

Have a workflow that should run itself?

Tell us what you want automated or built. You get a written assessment of feasibility, risks and a fixed quote — usually within two business days.