AI Explained: MCP — the USB of LLMs and the key to Agentic AI
What if you could ask an AI to schedule your holidays and it would look for destinations with good weather, find cheap flights, reserve an AirBnB, Uber to and from the airport, restaurant reservations, a local tour guide, tickets to museums & events, a public transport card for a week, excursions, request your boss to approve your holidays, …? For all of that to happen, AI needs to be able to talk to:
- a personal profile [where you store what type of holidays you like, where you live, personal information needs like passport data,…]
- a weather API to find the weather in different locations
- a flights reservation system like Amadeus to find and book flights
- the AirBnB or any other hotel API
- the Uber or local taxi & transport APIs
- an API to book restaurants like OpenTable
- an API to reserve excursions like TripAdvisor
- an API like TicketMaster for museums and events
- an API to public transport, potentially Uber
- an API to request holidays and add them to your Calendar
Now all of this and more can be done with a new open standard called Model Context Protocol or MCP. MCP allows the integration between LLMs like Anthropic’s Claude [the original creators of MCP] or ChatGPT [via for instance Goose] as well as your own hosted LLM like Llama via running oLlama on your desktop. Below you can see how llama3.2:1b is running locally with ollama.
Adding an MCP server is easy as well:
So provided you would have all the required MCP routers installed, then you could ask your local LLM to start reserving your perfect trip.
Now the cool thing is that MCP allows this server to run on an enterprise level. This way all employees could share access to any APIs for which the company has a license. Other tools could be integrated as well, e.g. ask your AI to go into your Salesforce or Monday and get the top 100 customers in Europe, to SAP or Xero to get your latest revenue numbers, to your HR system to book holidays or approve expenses,…
An MCP server has two main protocols:
- Stdio — which is used for instance to run local programmes, e.g. get me all the files on my hard disk which …
- SSE — server side events, which can be used to run a remote server.
MCP supports three types of main concepts:
- Tools — allow any API, database, local command line interface tool, custom developed tool,… to be discovered and integrated.
- Resources — any text or binary files to be discovered and read, e.g. provide access to all the internal specifications of the products or services you sell so your support agents can get answers.
- Prompts — structured workflows of questions to obtain something from the users.
The latest draft of the MCP standard talks about oAuth becoming the standard for SSE authentication and as such opening the doors for enterprise usage. At this moment most products still download arbitrary Javascript and Python code that gets executed locally which is a CISO’s worst nightmare. However we have been working on an open source MCP SSE server which will allow in the future to package new “MCP routers” as secure Wasm containers which can run even hostile and badly written code without endangering any enterprise infrastructure. Our open source server is built in Rust and uses Actors. Given it is focused on enterprise adoption we have named it mcp-ectors from MCP + Enterprise + Actors. Nothing to do with the fact that the last name of the creator is ectors ;-)
If you are looking into creating our own AI that gets embedded with lots of enterprise functionality or exposing your internal or external tools to employees and customers, you should check out: https://github.com/mcp-ectors/mcp-ectors
If you need help, please reach out to https://linkedin.com/in/mectors.
See other posts in the AI explained series:
- Deploying AI automatically via MLOps
- Retrieval Augmented Generation explained
- LLMs for content classification
- Integrating external tools with LangChain
- Image to text and computer vision
- LLM performance
- GPT, Operator, Responses API, Agents SDK vs MCP
If you are more interested in understanding in the business side of AI:
