MCP: An in-depth introduction
speakeasy.com
speakeasy.com
Yeah MCP is the worst documented technology I have ever encountered. I understand APIs for calling LLMs, I understand tool calling APIs. Yet I have read so much about MCP and have zero fucking clue except vague marketing speak. Or code that has zero explanation. What an amateur effort.
I've given up, I don't care about MCP. I'll use tool calling APIs as I currently do.
try it and you'll figure out
It’s just JSON RPC between a client, one or more servers. The AI agent interaction is not part of what the protocol is designed for except for re-prompting requests made by tools. It has to be AI agnostic.
For tool call workflow: (a) client requests the list of tools from the known servers, then it forwards those (possibly after translating to API calls like OpenAI toolcall API) to any AI agents it wants; when the AI then wants to call a tool (b) it returns a request that needs to be forwarded to the MCP server for handling; and (c) return the result back to the AI.
The spec is actually so simple no SDK is even necessary you could just write a script in anything with an HTTP client library.
I've spent so much time clicking through pages and reading and not understanding, but without finding the spec. Thanks so much!
I think it can be confusing in general it’s like understanding X11 where the client-server relationship is conceptually flipped. :)
Yes, that is the case, but they could have called just connections, a MCP Client (host) keeps multiple open connections to MCP servers. Everywhere in the docs and in the internet peoplE MCP Client the app connecting to mcp servers, not MCP hosts.
People often talk about web APIs, but we should also consider the integration of local tools. For me, the integration is mind-blowing.
When I tried the Playwright MCP integration [0][1] a few months ago, I really felt that after giving computers the ability to speak or communicate, we had now given them arms. I still get goosebumps thinking about it.
[0]https://youtu.be/3NWy_sxD3Vc [1]https://github.com/microsoft/playwright-mcp [EDIT]
And prople are skipping on service discovery. Making ai know what steps / operation is good.
Same. To see apps reverse engineered by LLMs with Ghidra [0] blew me away. It CTFed-out hard-coded access tokens and keys from .so's in seconds.
Pasting in a product owner's AC and and watching it browse through our test env for a few minutes before spitting out a passing - and passable - spec+test was kind of mind blowing.
This just continues to reinforce my feeling that everything around vibe coding and GenAI-first work is extremely shortsighted and poor quality.
Whether that wrapper's code runs on your laptop or a remote server changes nothing in terms of data exfiltration capabilities. If anything, it would make it more secure to have a remote server since at least you'd have full control over the code that's calling your API.
Obviously local MCP servers make sense to interact with applications that you have installed locally, but that's by far not their only use.
Will it be supplanted? Perhaps. But it's not going to die a natural death.
It's already gone so viral it's practically entrenched already, permanently. Everyone has invested too much time saying how much they love MCP. If we do find something cleaner it will still be called MCP, and it will be considered a 'variation' (new streaming approach maybe) on MCP rather than some competitor protocol replacing it. Maybe it will be called 'MCP 2.0' but it will be mostly the same and retain the MCP name for decades to come, I think.
MCP forcing JSON for tool specifications seems like a massive mistake.
Maybe Google can save us with something built on top of protobuffs.
No need to invent protocol for listing the tools or listing their schema. Just ask the gRPC server for the supported methods, look at the protobuf schema. This is mostly solved and supported out of the box. One potential improvement would be to have the server reply with original protobuf source, including comments, for even better semantic understanding.
No need for the absolute disaster of multiple HTTP requests + SSE, servers which need state to deal with session ids and all the problems that causes. It's just a gRPC channel and streaming methods.
And auth? Just shove credentials into the metadata. We can standardize that format, or have server reply what it supports.
Sigh... I feel like ten years ago garbage like this would be ignored or replaced with something actually sensible. But now nobody cares or feels the pain of the bad spec, they just vibe code some more mess on top of it and keep growing the ~ecosystem~ swampland.
Claude for example uses XML to generate mcp tool usage. At least top level strings don't need to be json encoded.
Most of the material on MCP is either too specific or too in depth.
WTF is it?! (Other than a dependency by Anthropic)
But to save you the click & read: it's OpenAPI for LLMs
Before the whole "just use OpenAPI" crowd arrives, the point is that LLMs work better with curated context. An OpenAPI server not designed for that will quickly flood an LLM context window.
Anyway, the technical merits don’t really matter. MCP (and any standard really) are only useful because they’re widely adopted. OpenAPI isn’t used for this, but MCP is. So, in practice, MCP is better for AI agents
https://en.wikipedia.org/wiki/List_of_Tron_characters#Master...
I like this succinct explanation.
You write a wrapper ("MCP server") over your docs/apis/databases/sites/scripts that exposes certain commands ("tools"), and you can instruct models to query your wrapper with these commands ("calling/invoking tools") and expect responses in a certain format that they can then use.
That is it.
Why vibe-coded? Because instead of bi-directional websockets the protocol uses unidirectional server-side events, so you need to send requests to a separate endpoint and then listen to the SSE hoping for an answer. There's also non-existent authentication.
Awful case of "not invented here" syndrome
I'm personally interested in if WebTransport could be the basis for something better
The protocol could easily be transported over websockets. Heck, since stdio is one transport, you could simply pipe that over websockets. Of course, that leaves a massive gap around authn and authz.
The Streamable HTTP transport includes an authentication workflow using OAuth. Of course, that only addresses part of the issue.
There are many flaws that need improvement in MCP, but railing against the current transports by using a presumably denigratory term ("vibe-coded") isn't helpful.
Your "that is it" stops at talking about one single aspect of the protocol. On the server side you left out resources and prompts. On the client side you left out sampling, which I find to be a very interesting possibility.
I think MCP has many warts that need addressing. I also think it's a good start on a way to standardize connections between tools and agents.
but that doesnt have to be necessarily negative
Could these commands be executed over websockets? Yes, they could. Will they? No, because the specification literally only defines two transports, and all of the clients only support those.
As with any hype, the authors drink their own coolaid, invent their own terminology, and ignore literally everything that came before them.
Even reading through explanations on the once again vibe-coded https://modelcontextprotocol.io/ you can't help to wonder why.
"tools" are nothing but RPC calls (that's why the base of this is JSON RPC)
"resources"? PHP could do an fopen on remote URLs in the 90s. It literally is just that: "Each resource is identified by a unique URI and can contain either text or binary data." You don't say.
"sampling"? It literally is just bi-directional communication. "servers request data from the client by sending commands". What a novel idea, must have a new name and marketing blurb about "powerful MCP feature, enabling sophisticated agentic behaviors while maintaining security and privacy."
As for auth, again, MCP doesn't have it, and expects you to just figure it out yourself. The entirety of the "spec" on it is just "MCP provides an Authorization framework for use with HTTP and your expected to conform to this spec". There's no spec. Edit: to be clear. At the point of writing all mentions of "MCP Auth Spec" on the internet link to https://modelcontextprotocol.io/specification/2025-03-26 which at the time of writing contains zero mentions of OAuth and says nothing about auth (and is not a spec to begin with) [1]
And so on.
It's hype-driven vibe-coded development at its finest.
[1] The auth spec is here: https://modelcontextprotocol.io/specification/2025-03-26/bas... I don't think anything on the site links to this directly. I found the link from some github discussion. See issues with it here: https://blog.christianposta.com/the-updated-mcp-oauth-spec-i...
You clearly have no desire to objectively evaluate what the specification is trying to do and are simply disregarding all aspects of the specification as trite or pointless. As such, this will be my last response of the subject.
I encourage you to take a breath and maybe try to understand why the specification was created in the first place before dismissing it fully.
Not on mobile. At least I couldn't see any obvious link to a very important part of the spec. There are circular links everywhere, none to auth.
> You clearly have no desire to objectively evaluate what the specification is trying to
I'm not questioning what it is trying to do. I'm questioning how it's doing that.
> try to understand why the specification was created in the first place before dismissing it fully.
I'm not questioning why it is trying to do. I'm questioning how it's doing that.
Stop buying into hype and marketing wholesale, and actually read what it is, and actually understand what your opponents are talking about.
Conversely, it allows many different LLMs to get context via many different Applications using a standard prodocol.
It addresses an m*n problem.
that's the missing piece in most of these description.
You send off a description of the tools, the model decides if it wants to use one, then you run it with the args, send it back to the context and loop.
Unless I'm missing something major, it's just marginally more convenient than just hooking up tool calls for, say, OpenAPI. The power is probably in the hype around it more than it's on technical merits.
The reality is that the space is still really young and people are figuring things out as they go.
The number of people that have no real clue what they are doing that are jumping in is shocking. Relatedly, the number of people that can't see the value in a protocol specifically designed to work with LLM Tool Calling is equally shocking. Can you write code that glues an OpenAPI Server to an LLM-based Tool Calling Agent? 100%! Will that setup flood the context window of the LLM? Almost certainly. You need to write code to distill those OpenAPI responses down to some context the LLM can work with, respecting the limited space for context. Great, now you've written a wrapper on that OpenAPI server that does exactly that. And you've written, in essence, a basic MCP Server.
Now, if someone were to write an MCP Server that used an LLM (via the LLM Client 'sampling' feature) to consume an OpenAPI Server Spec and convert it into MCP Tools dynamically, THAT would be cool. Basically a dynamic self-coding MCP Server.
We are already off to a wrong start, context has a meaning specific to LLMs, everyone who works with LLMs knows what it means: the context is the text that is fed as input at runtime to LLM, including the current message (user prompt) as well as the previous messages and responses by the LLM.
So we don't need to read any further and we can ignore this article, and MCPs by extension, YAGNI
As you yourself say, the context is the text that is fed as input at runtime to an LLM. This text could just always come from the user as a prompt, but that's a pretty lousy interface to try to cram everything that you might want the model to know about, and it puts the onus entirely on the user to figure out what might be relevant context. The premise of the Model Context Protocol (MCP) is overall sound: how do we give the "Model" access to load arbitrary details into "Context" from many different sources?
This is a real problem worth solving and it has everything to do with the technical meaning of the word "context" in this context. I'm not sure why you dismiss it so abruptly.
Also providing data through function calls/tool use is not context, you are overloading the term. Context is LLM context, if you fetch from a db it's something else
Given that your only contributions in this thread are to acknowledge your ignorance of MCP [0] and to post the summary dismissal upthread that shows your ignorance of it, it would probably behoove you to actually learn about MCP more before confidently making assertions about it. Suffice it to say that this is inaccurate and others have already explained in your "WTF is it" thread what MCP actually is.
> Also providing data through function calls/tool use is not context, you are overloading the term. Context is LLM context, if you fetch from a db it's something else
If you believe this then you don't understand how tool use is implemented. It's literally accomplished by injecting a tool's response into the context [1].
As a general life tip: most pedants are wrong most of the time. If you find yourself being pedantic, take a few steps back and double check that you're not just wrong.
[0] https://news.ycombinator.com/item?id=44011320
[1] https://platform.openai.com/docs/guides/function-calling
I was doing tool use before chatgpt released an official API for function calls. You literally give ChatGPT API specs and ask it to generate call parameters.
The API is fed into the LLM as context, the response is part of the output. Whether you pass that output through another layer of LLM is trivial. And even if you do, the "context" in that case would be only the response, not the whole database. You are confusing even yourself, you accepting the overloading of the word 'context' (pushed by a company for commercial purposes) and you are now unable to distinguish between LLM context in terms of tokens, an external data source, and a response fetched by the tool.
It's not that I am ignorant of what Anhtropic claims Context means, I'm contesting it. If Microsoft releases a new product and claims that Intelligence is the parameters of their Microsoft Product, then it pays to be a bit cynical instead of parroting whatever they say like some unpaid adman
> The API is fed into the LLM as context, the response is part of the output.
So you implement tool use by feeding an API into the LLM as context in order to get it to produce call parameters. Got it.
> Whether you pass that output through another layer of LLM is trivial. And even if you do, the "context" in that case would be only the response
So the output of the tool when called with those parameters can be fed back into the LLM as further context. Got it.
Given the above, it seems that we agree that tool use is implemented entirely by giving selected bits of context to the model.
With that in mind, if one were to design a protocol that makes tool use plug-and-play instead of something that has to be coded by hand for each tool—a protocol designed to allow a model to discover tool APIs that it might want to bring into context and then use those APIs to bring their outputs into context—it would be reasonable to call said protocol the Model Context Protocol, because it's all about getting specific bits of Context into a Model.
I'm not sure why the word "context" is the hill you decided to die on here when there is so much else to pick on with MCP, but it's time to get off the hill.
It pays to be precise when speaking and studying, and it pays to develop a precise language on nascent technologies when we comunicate about them.
This reminds me when I was studying chemistry and I thought they were pedantic for the way they used the word salt. Or when I studied chess and I called every bishop and knight attack to the f6 pawn the fried liver, instead of the specific sequence of moves that we call the fried liver. Or when I thought that the arm forearm distinction was pedantic in medicine
Science demands precision in communication, feel free to steal a well defined term and use it to mean something else that already has a different sign to denote it. But I'm not playing
Agent LLMs are able to retrieve additional context and MCP servers give them specific, targeted tools to do so.