Before MCP, every AI application had to build every integration itself: one connection to the CRM, one to the database, one to the ticket system, each in its own format. The Model Context Protocol solves this the way USB did for peripherals: a system exposes its functions once in a standard way, and any compatible AI application can use them.
The architecture
MCP distinguishes three roles. The host is the AI application a person works with, such as a chat client or a development environment. The client within the host maintains the connection to a server. The server exposes a system’s capabilities. Servers can run locally or as remote services over HTTP; for remote servers, authentication via OAuth is provided for.
What a server provides
Tools are functions the model can call, such as a search or creating an entry — technically the same principle as tool calling. Resources are data that can be read as context, such as documents or records. Prompts are prepared templates for recurring tasks. A well-built server offers a few clearly described tools rather than a one-to-one mapping of every API route.
MCP for SaaS products
For SaaS providers, an MCP server is a new kind of interface: customers can use the product directly from their AI assistants, query data and trigger actions. We also run MCP servers for products in our own portfolio. Experience shows the effort lies less in the protocol than in shaping the tools, permissions per user and tenant, and clean error messages.
Security questions
MCP spreads the risks of tool calling widely. A server with broad rights is an attractive target; prompt injection through content that is read can trigger tool calls; manipulated tool descriptions can lead a model to unwanted actions. Countermeasures are least privilege per tool, server-side checks of tenant and user, confirmation of consequential actions, logging and a deliberate choice of which servers an organisation allows.
Perspective
MCP is not a replacement for a well-built API but a layer on top, tailored to models. Those with a clean API including authentication and tenant separation can add an MCP server quickly. Those who still need such an API should start there; our API integration sprint helps.