What is the Model Context Protocol?

Why MCP was created

Before MCP, every AI application had to build its own integration for every tool – a project tool here, a calendar there, an internal database somewhere else. With many applications and many tools, the number of integrations multiplied. MCP was published as an open protocol in 2024 and is now supported by numerous AI clients and providers. The specification is publicly available at modelcontextprotocol.io.

How MCP works

MCP distinguishes three roles:

• Host: the AI application the user works with, for example a chat or an agent platform.
• Client: the component within the host that maintains the connection to a server.
• Server: a service that makes a system accessible – for example a CRM, a ticket system or a file storage.

An MCP server offers mainly three kinds of capabilities: tools (actions the model can call, such as "search records"), resources (data such as files or entries that can be loaded as context) and prompts (predefined templates). Communication runs via JSON-RPC, either locally or over HTTP; for remote servers, the specification provides authorisation based on OAuth.

What MCP means for companies

• Less integration effort: a system connected once via MCP can be used by different AI clients.
• Freedom of choice: tools and AI platforms can be combined and exchanged more independently.
• Basis for agents: AI agents need tools; MCP is a common way to provide them.

Risks and points to check

• Permissions: an MCP server acts with the rights of the account or token used. Grant only the minimum rights necessary.
• Trustworthiness: only connect servers from trusted sources; a server can return manipulated content that misleads the model.
• Write actions: tools that send, change or delete data need clear rules and, where appropriate, confirmation.
• Data flows: check which data reaches which system and whether a data processing agreement is needed – see GDPR and AI.

MCP in AIPE

AIPE supports MCP in both directions. First, AIPE provides an MCP endpoint through which external AI clients can access AIPE functions – secured via the automation API with API keys. Second, users can connect their own or external MCP servers from the connectors area, authenticated via OAuth or a token. Agents then use the configured MCP servers as tools. Built-in connectors and MCP servers are described on the connectors page; how agents use tools is explained in the guide AI agents.