The Model Context Protocol (MCP) is an open standard launched by Anthropic in November 2024 that lets artificial intelligence models connect to external tools, data and systems in a standardised way. It is the "USB-C port" of enterprise AI: a single protocol for integrating any data source, removing the need for custom integrations per service.
What MCP is, explained simply
If traditional AI is a brilliant consultant working in a room with no windows, MCP is the door that gives them access to the whole building. Before MCP, every connection between an AI model and an external system required a bespoke integration: dedicated code for each CRM, each database, each internal API.
MCP standardises that connection. An MCP-compatible AI model can dynamically discover the capabilities of any MCP Server and use them, in the same way a web browser can display any website without the site having been built for that specific browser.
Why MCP exists
The problem MCP solves is known as the M×N integration problem: with M AI models and N business systems, and no common standard, you need M×N separate integrations. With MCP each system exposes a single MCP interface and each AI model speaks a single protocol: integrations become M+N instead of M×N.
Anthropic published MCP as an open, non-proprietary standard precisely to encourage adoption across the ecosystem. The major AI vendors - OpenAI, Google, Microsoft - are converging on MCP as the de facto standard.
How the MCP architecture works
The MCP architecture rests on three components that work in sequence:
MCP Host (where the AI lives)
The MCP Host is the AI application hosting the language model: it might be Claude Desktop, a custom application, or an enterprise system. The Host manages several MCP clients at once and coordinates interactions between the AI and external systems. It is the "brain" of the setup.
MCP Client (the mediator)
The MCP Client lives inside the Host and acts as a mediator. It translates the AI model's requests into structured messages compatible with the protocol (based on JSON-RPC), and returns the responses in a format the model can process. Each Client keeps a dedicated connection to a single Server.
MCP Server (access to the data)
The MCP Server is the external service exposing specific capabilities: a GitHub connector, access to a PostgreSQL database, a Slack integration, or a custom ERP connector. Every Server exposes three kinds of capability: Tools, Resources and Prompts.
For a technical deep dive on how an MCP Server works, how to build one and which servers exist today, see the complete guide to MCP Servers →
The three MCP pillars: Tools, Resources, Prompts
Tools: when the AI takes action
Tools are functions the AI model can invoke on its own. They are operations with effects: updating a CRM record, sending an email, creating an order in the ERP, running an SQL query. The AI decides when and how to use them based on the context of the conversation or workflow.
A practical example: while processing a customer order, the AI uses one Tool to check product availability in the warehouse, a second to create the order in the ERP and a third to notify the sales team by email - all in one automatic sequence.
Resources: controlled access to data
Resources are read-only data sources. Similar to the GET endpoints of a REST API but optimised for AI consumption: they provide context without side effects. When a user asks about a customer, the AI automatically reaches the relevant Resources - order history, active contracts, open tickets - and includes them in the context before answering.
Prompts: reusable templates
Prompts are predefined templates that optimise the use of tools and resources for specific scenarios. They are the evolution of prompt engineering: tested, tuned patterns for recurring use cases ("process this delivery note", "reconcile this invoice against the order").
MCP vs traditional APIs: 5 key differences
| Aspect | Traditional API | MCP |
|---|---|---|
| Integration cost | High (custom code per service) | Low (one protocol) |
| Maintenance | High (every API changes) | Low (stable standard) |
| Flexibility | Fixed (predefined schema) | Dynamic (discoverable) |
| Discoverability | None (manual documentation) | Automatic (the model discovers capabilities) |
| Standard | Proprietary per service | Open and universal |
Business use cases for MCP
CRM and ERP integration
With MCP, a corporate AI agent can reach the ERP in real time to check prices and availability, the CRM to retrieve customer history, and the document system to consult active contracts - all within a single conversation or automated transaction.
Automated document handling
MCP enables fully automated document flows: an email with a delivery note attached lands in the inbox, an AI agent receives it through an Email MCP Server, extracts the data with intelligent OCR, validates it against an ERP Resource, and creates the warehouse movement through an ERP Tool. With no human intervention for standard cases.
Multi-tool AI agents
The power of MCP shows in agents that orchestrate several tools in sequence. A procurement agent can receive an order request, check the supplier in the CRM, verify the budget in the ERP, generate the order and send confirmation by email - all from a single user instruction.
MCP and TypeLens: a concrete example
Workflow Builder: no-code automation, professional logic
The TypeLens Workflow Builder lets you assemble document flows by dragging intelligent blocks: intake → extraction → validation → ERP entry, with no code. MCP is the infrastructure that lets those blocks talk to business systems in a standardised way.
The available blocks include AI Agent (models specialised for each step), Router (automatic routing based on document content) and HITL - Human in the Loop - for manual validation on ambiguous cases.
Context means cleaner data
Every block in the workflow receives the full context: the original document, customer history, up-to-date product inventory. That removes ambiguity and improves extraction accuracy from the very first cycle, reducing false positives and the need for manual review.
How to get started with MCP
Existing MCP Servers
Dozens of ready-to-use MCP Servers already exist for the most common systems: GitHub, PostgreSQL, Slack, Filesystem, Brave Search and others. For the full list with installation guides, see the page dedicated to MCP Servers →
Building your own MCP Server
To integrate a proprietary system or a company ERP not covered by existing servers, you can build a custom MCP Server using the official SDK in Python or TypeScript. The process involves defining the Tools, Resources and Prompts to expose, and configuring the transport (stdio for local use, SSE or WebSocket for remote use).
Security considerations
Before exposing business systems through MCP, it is important to decide: which Tools can have destructive effects (and therefore require explicit confirmation), which Resources expose sensitive data (and require authentication), and how accesses are logged for the audit trail.
Frequently asked questions about MCP
Who created MCP?
MCP was created by Anthropic and announced as an open standard in November 2024. It is designed to be adopted by the whole AI ecosystem, not only by Claude models.
Is MCP compatible with OpenAI and Gemini?
Yes. MCP is an open, model-independent standard. OpenAI, Google and other AI providers are adopting or planning to adopt MCP in their systems. The goal is a universal integration ecosystem.
Do you need to be a developer to use MCP?
Not necessarily. Using existing MCP servers in applications such as Claude Desktop only requires basic configuration skills. Building a custom MCP Server or integrating one with a company ERP does require technical expertise.
What is the difference between MCP and a traditional API?
With a traditional API every integration requires custom code: authentication, error handling, response parsing. MCP standardises all of that: an AI model discovers the server's capabilities dynamically and uses them without service-specific integration code.
What is an MCP Server?
An MCP Server is a service that exposes tools, resources and prompt templates to an AI model through the Model Context Protocol. It lets the AI perform concrete actions such as reading a database, sending an email or updating an ERP.
MCP: the future of enterprise AI
MCP is the infrastructural foundation on which the next generation of enterprise AI agents is being built. Just as HTTP made the web possible by standardising communication between browser and server, MCP is standardising communication between AI and business systems.
Organisations implementing MCP today are not just adopting a technology: they are building infrastructure that will support the entire future evolution of AI agents in their stack.
For the technical side, see the guide to MCP Servers. To see how MCP-based document automation works in practice, look at how Typelens automates delivery notes with this architecture.