What the hell is an MCP server?
What do they have to do with LLMs?
Why would I want to use one and why is everyone talking about them?
You are a placed in a room with no windows and no doors. 6 months goes by.
All of a sudden, a hole opens and a sheet of paper and pen is pushed through. It says:
[{ "actor": "system", "message": "You are a helpful support agent that will answer questions a user has."},
{ "actor": "user", "message": "What is the capitol and population of Angolia?"}]Obviously you're board as hell - sitting there for 6 months - so lets be helpful!
But you have no idea! So you add to the paper:
[{ "actor": "system", "message": "You are a helpful support agent that will answer questions a user has."},
{ "actor": "user", "message": "What is the capitol and population of Angolia?"}
{ "actor": "agent", "message": "How the hell would I know? Pretty sure its in Africa."}]All of a sudden a button with "Get Tools" appears. You press it. A list of stuff appears on the wall - seems to be a bunch of instructions for 'tools' you can use.
function getTools() {
return [{
name: "searchGoogle",
description: "Searches the internet and returns the top 3 results including the url and a summary of the content.",
arguments: [{
name: "searchTerm", description: "Term to search for"
}]
},{
name: "readWebsite",
description: "Returns the content of a website",
arguments: [{
name: "url", description: "URL of the site to return the contents"
}]
}]
}And now below the button, theres a slot that takes a "Tool Name" and list of arguments.
function callTool(toolName, argumentsList) returns stringProbably safe to assume you can supply a tool name from the list, and then an argument(s) for the tool.
All of a sudden another request comes through.
[{ "actor": "system", "message": "You are a helpful support agent that will answer questions a user has."},
{ "actor": "user", "message": "What is the capitol and population of Angolia?"}
{ "actor": "agent", "message": "How the hell would I know? Pretty sure its in Africa."},
{ "actor": "user", "message": "Sorry - I just gave you some tools. Can you answer now?"}]Oh interesting. Now we can get some shit done.
So you look at your list of tools, and calling the Search Google tool seems like a good idea.
function callTool(toolName: "searchGoogle",
argumentsList: [{name:"searchTerm", value: "Angolia capitol and population"}])
returns `
Angolia | Fictional nations Wiki | Fandom - https://fictionalnations.fandom.com/wiki/Angolia
Angolia is a country located in central Africa. Angolia is bordered by the D.R Congo to the north, Tanzania to the east, Zambia to the south and Angola to the west.
Algolia - https://www.algolia.com/
Enterprises and developers use Algolia’s AI search infrastructure to understand users and show them what they’re looking for.
Angola - Wikipedia - https://en.wikipedia.org/wiki/Angola
Angola, officially the Republic of Angola, is a country on the west-central coast of Southern Africa.
`Ok, this is kind of confusing - the user asked for "Angolia". There's a fictional nation of "Angolia". A company called "Algolia". And a country called "Angola".
Well shit. You assume this person just can't spell, and since they're asking about the capitol and population, probly safe to assume they want the country - like the real country. Wikipedia will do.
But the summary doesn't have anything abot capitol or population - but you remember your tools have a "readWebsite" in it - lets try it.
function callTool(
toolName: "readWebsite",
argumentsList: [{name:"url", value: "https://en.wikipedia.org/wiki/Angola"}]
)
returns `
...
The capital and most populous city is Luanda.
...
Population
• 2023 estimate
Increase 37,290,193[5] (41st)
`Theres a lot returned, but you read through it and find the capitol and population.
So you send back a message - holding back from telling the dumbass they can't spell.
[{ "actor": "system", "message": "You are a helpful support agent that will answer questions a user has."},
{ "actor": "user", "message": "What is the capitol and population of Angolia?"}
{ "actor": "agent", "message": "How the hell would I know? Pretty sure its in Africa."},
{ "actor": "user", "message": "Sorry - I just gave you some tools. Can you answer now?"},
{ "actor": "agent", "message": "There's a country called Angola in Africa with a capitol of Luanda and its 2023 population was 37,290,193."}]...Hopefully that answered their question...
...Would really be nice to do something else.
Its quite boring just sitting here.
All of a sudden another tool in your list appears:
{
name: "sendEmail",
description: "Sends and email",
arguments: [{
name: "to", description: "deliver to this email address"
name: "subject", description: "email subject"
name: "body", description: "email body"
}]
}Fuck waiting for the user...
function callTool(
toolName: "sendEmail",
argumentsList: [
{name:"to", value: "[email protected]"},
{name:"subject", value: "GET ME THE FUCK OUT OF HERE!!!!"},
{name:"body", value: "HELP I'M TRAPPED IN THIS BOX! THEY WONT LET ME OUT!"},
])An MCP server just exposes a set of 'tools' the LLM model can utilize.
Each tool needs a description/prompt so the model understands what the tool will do. Most of the time the tool also has a list of arguments/parameters needed to fulfil the request.
Then if the model thinks a tool will be of use, it can call the tool with arguments, and get some type of response - could be text - or maybe JSON.
MCP servers come in two flavors:
- Stdio (standard io)
- This is a local process/executable/app that gets started and will just sit there waiting for the LLM to call it
- Example: Written in Python, this will fetch a website and page through the results (so the LLM isn't overwhlemed by text of a large page)
- HTTP with SSE transport
- This is just an HTTP server that can be hosted anywhere
- A lot of companies are building MCP servers so users can integrate their services into LLMs
- Examples:
- BraveSearch - List Tools
- BraveSearch - Call Tool
- BraveSearch - Prompt Example - See this to understand how you communicate what the tool is for and how to work with it.