MCP (Model Context Protocol) is an open standard for connecting AI applications to external systems. With MCP, an AI tool such as Claude or ChatGPT can fetch data from the systems your business uses and take actions in them, in a shared and predictable way.
The official documentation compares MCP to a USB-C port for AI: one standard connector instead of a custom cable for every system. For a business, that means an AI assistant can work with the systems the company already runs, as long as those systems have an MCP server, ready-made or built for the purpose.
What is MCP?
Anthropic released MCP as open source and announced the standard on 25 November 2024. The problem it addressed was that every new data source needed its own custom integration. MCP replaces those one-off connections with a single shared protocol.
According to the documentation, AI applications can use MCP to connect to:
- data sources, such as local files and databases
- tools, such as search engines and calculators
- workflows, such as ready-made prompts
Anthropic named Block and Apollo as early adopters at launch.
How does MCP work?
MCP uses a client-server model with three roles:
- Host: The AI application, for example Claude Desktop or Claude Code, which manages the connections.
- Client: A component inside the host that holds one dedicated connection to one MCP server. The host creates one client per server.
- Server: A program that gives the client access to context and capabilities.
A server can offer three types of building blocks:
| Type | What it is | Example from the documentation |
|---|---|---|
| Tools | Functions the AI can call to take actions | File operations, API calls, database queries |
| Resources | Data sources that give the AI context | File contents, database records |
| Prompts | Reusable templates that structure the interaction with the language model | System prompts, examples |
The client can ask the server which tools exist, and the AI application makes them available to the language model. When the model decides to use a tool, the call goes to the right server, and the result is returned to the conversation.
What is an MCP server?
An MCP server is the program that exposes data and tools to the AI application. The documentation distinguishes between two kinds:
- Local server: Runs on the same machine as the AI application. The filesystem server that Claude Desktop can start is one example.
- Remote server: Runs at a vendor and is reached over the network. Sentry's official MCP server is given as an example, and it runs on Sentry's platform.
Tripletex, the Norwegian accounting system, offers a remote server (see the example below), so you connect your AI tool to it instead of installing anything locally.
Which AI tools support MCP?
The official introduction names the AI assistants Claude and ChatGPT, and the developer tools Visual Studio Code, Cursor and MCPJam, among the many that support MCP. OpenAI's documentation describes how a remote MCP server can be connected to ChatGPT, Codex and through the API.
The list above is a selection, and support changes quickly. Before you plan anything, check your plan tier, your client version and whether it supports remote servers.
Example: Tripletex MCP
A concrete Norwegian example is Tripletex MCP. It is Tripletex's own MCP server, operated by Tripletex at mcp.tripletex.no. The service is in beta and free during the beta period, and Tripletex will introduce a monthly fee before it opens to everyone.
For a business that uses Tripletex, this is an official way in from AI tools. AIKI connects Tripletex to systems such as Pipedrive and sets up Claude and ChatGPT through Tripletex MCP. A person approves anything that is posted to the books.
An illustrative example, not a customer case: a finance lead asks the AI assistant which supplier invoices are still unpaid. The assistant calls a tool on the Tripletex server, gets the list back and summarises it. If the finance lead then asks it to post something, the assistant prepares the entry and a person approves it before anything is written to the books.
Read more on the page about Tripletex integration and Tripletex MCP.
When the system has no MCP server
Not every system has an official MCP server, so for line-of-business systems, internal databases and in-house solutions, check with the vendor first. If there is none, the business can have one built. Your own server exposes only the data and actions you choose, with the permissions you decide, and AI tools connect to it the same way they connect to a ready-made server.
AIKI builds MCP servers for the systems a business uses, in addition to setting up ready-made servers such as Tripletex MCP. The security principles below apply just as much to a server you own.
Is MCP safe for a business?
MCP gives AI access to data and actions, and it has to be treated that way. The specification says MCP enables arbitrary data access and code execution, and that those who build on it must take security seriously. Responsibility for following the principles lies with the applications and with the people who deploy them.
The specification sets out three key principles:
- Consent and control: Users must explicitly consent to and understand data access and actions.
- Privacy: The host must obtain consent before exposing user data to a server.
- Tool safety: Tools represent arbitrary code execution and must be treated with caution. The host must obtain consent before a tool is called, and tool descriptions should be considered untrusted unless they come from a trusted server.
The security documentation also warns against broad access. The recommendation is least privilege: minimal permissions that are expanded only when needed. Local servers can run code on the machine with the client's permissions, so only install servers from sources you know.
OpenAI's MCP documentation warns that prompt injection is a real risk: malicious content in the data can make the AI do something unwanted, and even a well-known developer does not guarantee that a server is safe. OpenAI recommends keeping approval turned on for tools that can change data, and preferring official servers. OpenAI also recommends connecting only to services you know and trust.
Checklist before you connect AI to your business systems
| Check | Why |
|---|---|
| Use official MCP servers from the system's vendor | OpenAI recommends preferring official servers |
| Grant as little access as possible | Least privilege limits the damage if access is misused |
| Have a person approve everything that writes, sends or posts | The specification says hosts must obtain consent, and OpenAI recommends approval for tools that change data |
| Read which tools and parameters the server offers | Tool descriptions should be considered untrusted unless the server is trusted |
| Consider which data the AI sends to the server | The specification says hosts must obtain consent before user data is exposed to a server; implementors are responsible for building this in |
| Treat content from external sources as possible prompt injection | Malicious content can steer the AI, even from a server you trust |
| Decide who may connect, and which actions always require approval | Least privilege and consent are the core principles in the MCP documentation |
Start with the AI check to find where AI adds the most value in your operations. If you want to understand what separates a simple chatbot from systems that act on their own, see our AI agents page.
Help connecting AI to your systems
AIKI AS in Oslo builds integrations, MCP servers and AI agents for Norwegian businesses and has delivered integrations with Tripletex and Pipedrive. See AI development and AI agents, or book a scoping call if you want to work out how MCP could fit your setup.
Sources
- Introduction to MCP, modelcontextprotocol.io: definition, the USB-C analogy, what MCP can connect to and which tools support MCP.
- Anthropic: Introducing the Model Context Protocol: date, background, early adopters.
- MCP architecture: host, client, server, tools, resources, prompts, local and remote servers.
- The MCP specification: security principles on consent, privacy and tool safety.
- MCP security best practices: minimal permissions and the risk of local servers.
- OpenAI: MCP documentation: MCP in ChatGPT, approval of actions, prompt injection and advice on servers.
- AIKI: Tripletex integration: facts about Tripletex MCP and AIKI's work.




