TWed Talk: "Model Context Protocol (MCP): Standardizing Tool Use for LLM Systems" (18 Feb 2026)

Key Takeaways

It's not a problem. >> Oh, but you're recording now because we we want the recording. >> I'm recording. >> Okay. Thanks. >> Okay. Uh, [sighs] okay. So, MCP. So MCP has been showing up like recently as the kind of the de facto way of making large language model talk to external tools and u many um LLM's uh servers are already coming with that builtin um so it's kind of pretty easy uh to integrate these into uh your own LM workflow. Um and many new models, if not most of them are already uh trained for true calling, which makes those things even easier. Um I'm going to be very pragmatic uh in this talk. So I'm going to show the motivation and then we're going right jumping into how it's done. [laughter] Okay. So the motivation is how can we make um those uh large language model actually talk to external systems. I think the the biggest idea here is we all know that LM's outputs are statistical. Sometimes they might reason in the way they want them to do it. Sometimes they might not. So when you call an external tool, most likely that tool will be um deterministic. So the results obtained uh can be used like in a reusable or trustworthy manner and present to the LM as a source of uh authoritative sources or data. So I think that's the um the biggest motivation that is not in the books. uh but it's uh for me it's the it's the most important one in the books they are oh it's a it's a way of LMS to call systems systems right so again LMS are no longer standalone chat systems they need to check FPIs and other tools uh get data from databases and knowledge graphs access documents external data and be part of workflows and services and calling those services as part of workflows. Uh the two integration can be messy. So uh services exposing capabilities before MCP could be many ways that web services, APIs and so forth. Um and uh this can lead to problems such as how to decide which tool to to use, how to match data types and um and being hard to debug when you have workflows as well. So the way you can easily see MCP is the analog of the good old web services stack, right? So not original but in the in the web world you have HTTP and open API is a probably the most used API for web services. So if you have some implementation you have you have an HTTP client trying to access that uh and that uses the that can use the OpenAI spec to call a web service right and MCP is kind of similar so instead of your software traditional software you might have AI based >> or NLM itself. Um and then you you instead of calling through HTTP you'll be calling through MCP right and you have an schema for that and that schema is composed by three things I'm going to talk next tools resources and prompts so those three things can be exposed the MCP interface those are the core concepts tools resources and prompts So tools are actions. So basically do something. So there are functions. Uh this maps directly to any like web servers exposed function. Uh they are functions. They can do deterministic computation. Uh but they can involve of course not. Uh so example I query data run some analysis. Uh resources are just like read only things. Um it's supposed to be something that is um just retrieval uh as is. But traditionally actions and tools and resources are bundled together when we have services. We don't differentiate with them. They're just a function we call and retrieve data or execute some actions. So there's no distin distinction. Um and there are prompts as well. So we can think about as a kind of a reasoning template. So our pre-created prompts that the server is offering the tool if they want to use a specific prompting that might make it easier to work with whatever those things return. Right? But let's say this is in the JSON. I might have a template let's say this is JSON like using some schema and I can maybe create a prompt to that the client can use to parse that JSON better right so I don't need to create my own prompt um depending on the client uh they might have implemented this this or this or just this the one I'm going to show in the demo is very simple I think they just implemented tools Right. Okay. So, I'm going to going to jump straight into the demo. Let's go ahead. So if you do Python, the easiest way to create an server is using fast MCP library. You can s that with pip, right? And uh it's very simple uh your code you can define tools by annotating them with this MCP.2 annotation and you basically just describe create your function. So this one is a very simple function that takes some string as parameter and returns another string. Right? This one is like this. Uh you said this. So I returned the basically the same thing that got what is important here is this thing. So this is actually not documentation. This is actually uh instructions for the MCP client, right? So I'm describing the function and when to use it as natural language, right? I use this tool when the user asks you to form a short piece of text in a simple deterministic way. The description is used by this to decide whether to call the tool. I may or may not inform the inputs and output types. Uh some servers that I have seen they describe that some they do not. Uh and they still work well. But this is how you do it. Then inside your function could do basically your anything you want to do. query something, do some computation, anything, and then return the value, right? >> But you always have to have that description. >> I don't think you always need to have that, but if you don't, I don't think the model will be ever called that tool. [laughter] >> Uh resource, it's uh very simple as well. So we annotate with MCP.resource. this one in the clients that I'm using it the clients do not implement resources or prompts but uh so I never actually tested this but the idea is there is no like computations just retrieval right so you might just return a a document it might just return some file some anything like a a knowledge graph entity something that is very straightforward So I never actually uh tried that uh in a pipeline because the client that I'm using does not implement this yet. Um in prompt it's very similar. It's even simpler. So you just put the prompt in. You also describe the prompt. Use this prompt when this you should use this prompt when doing that. Right? And then you return the prompt which is like this. You are using a demo prompt template and this is the following this this this. So the idea here is the model will decide which things to call based on the natural language instructions you give them. So that's there's the where comes the decision. I know it's weird. >> But what if they're super duper similar? How is it going to decide? >> Yeah, if you have like lots of tools, >> yeah, >> it might get weird. So, I think the uh the co-pilot uh plugin for visual studio code, I think it there's a warning. If you have more than 175 tools, it's not recommended to have that many. [laughter] But it seems like it's more important to have tools that that description of when to use don't don't overlap. >> Yes. Yes. >> I mean, I could have 175 tools, but I use them in very different ways. >> But you're going to see how it works. So, it might try one doesn't work. Oh, I I'll try the other one. The modern sallall many times, right? Also, takes it take more time to to produce the answer. Um okay so that's the uh the basics it's very simple expose them as prompt tools or or sources and uh your main functions as simple as this MCP run and you you define the transport uh which I'll be so if you have actual server, then you put this. If you if you're trying like locally, you might do stdio, but I'll do like this because I'm going to actually use this in my server. So, I run this assuming that I have Python with MCP fast already installed. I just do Python server.py and that's it. And then it's it's running. But that's exposed on port 800 8,000. And it's running. It's very very simple. I won't even test this because this one is very naive. But we're going to show you this one. This one is another server. It has two functions, two tools. One is add two numbers. So, a and b as parameters as integers and return an integer. And this one is I think I I I got from this idea from you John to actually have a OS thing getting the current time stamp. >> This is brilliant function I tell you. >> I'm going to show you that work. It's very funny. Okay. So, just to get the gr time it's a a little bit more than that. So uh if you say like yesterday you'll get like you understand that you're talking about that this one actual numbers and uh this one can take a time zone as parameter if you don't send it um New York by default. So let me actually run this first server by it's running. Okay. So I am going to go to my client. Let me delete this. So this is LM studio. Um it's a local uh LLM host you can install. I think it's available for anything. Um, and you can download local models and run and has a chat interface for you to talk to them. Uh, so I'll create a new chat. Let's see. I will use GPT open source 20 billion and I'm going to uh let me reload that. Let's restart. Okay, so that's my server. This MCP2 add that one that I'm running right now. So it shows the ad function and the get time context function. I think if I click here, [clears throat] not showing for some reason. Okay, sometimes it shows the the actual string that I put in there like add numbers. So I'm going to do this. I'm going to put on medium reasoning. This is a model that supports that. So, actually, let me just disable this first. So, I'm going to um add three I don't know 13 plus I don't know 5,000. So, it's not using the fang yet. So, like we can see the thought process. That's the thought process. The user says this, they want the sum probably a simple addition. The answer should be this. Maybe they just want the result. So respond this. That's the thought process. So let me create a new chat now. Turn this on. Right. So the model know this thing exists. Now I'll do the same thing. So we need to use the I function. So it calls that you see the parameters A 13 B 5,000 and get the results deterministically computed and just answers that right. So it's it decided to call the ad because there was like add to numbers. So for from the context uh I try to understand what this client actually does when injecting those function descriptions. Uh I haven't found like a final answer but they say they have an internal system prompt that they tweak as they know those functions. Um and that's it. So another thing that I would like to show is this one. Oops. Sorry again. And I'll turn this off. What day is today? This one is funny. [laughter] >> Write write a dissertation on what the day is. [laughter] >> Wow, there's lots of thinking. Yes, it's too hard to know what >> it's funnel index fly >> speaking which calendar [laughter] >> tell us congrance is this floor compute this [laughter] >> I have no idea >> host minus Steve Jobs birthday is this divide [laughter] it's today a It >> is like >> three, four, five, six, seven, eight. [laughter] [laughter] >> Algorithm. Oh, there's that mention. I'll open your chat. [laughter] Turn this on and ask the same thing. Imagine and give some philosophy. [laughter] Hey, let's finish. Oh, yeah. >> The other one's still writing a dissertation. [laughter] >> Yeah. So, I call decided to call the tool and the tool did its thing. Returned. That's it. Right. So, yeah, it's that simple. Uh let me turn this off again. I want to show the wiki data. So in chatb you always use the uh the um uh what is the oldest >> technical university >> technical university in the United States. >> That's funny. >> It's a test thing, right? I turn on the wikid data one. What was your name? >> India. >> Note that I never asked this to actually match um any wiki data thing, right? I just enabled that. So the model itself already reason something. So found MIT was this maybe this this is older in the end of this size but I better confirm with wiki data >> that was actually worse than 3.2 Right. That's that's worse than anything we've seen. >> Yeah. And then let's see the query for this oldest university in United States. Some results. Nothing. Still calling. So it's calling like until it thinks it has the answer. As if this is still going >> bear machine. >> I like that. But is it considered a technical university like that? It looked like it got uh RPI and RIT confused. the the same with my parents' friends when I got admitted to [laughter] >> the same with my invited speaker when I was running a conference here. He booked a flight to where RIT is. >> [snorts] >> Let's see if it does something because when we first tried this last week, I guess it was between MIT and RPI, the same model, >> but RPI never showed up on this reasoning today. >> Yeah. No, let's see if it discovers that RPI is the oldest somehow. >> Isn't that >> Well, it is contemplating what is a technical university. >> Oh, it said was instead of RVIPS. So what happens here when I send a message, the client intercepts the message and uh works with the tool descriptions before actually producing the final answer to decide whether to just respond or to call something and um and um and do the right thing most likely. >> Well, and new to look at RPI there. I don't know if you saw that. Is RPI 1180 just that? >> I don't think it >> What is Q 1180? >> It's the entity >> code beta. >> Oh, I think RPI is 1180. [laughter] That's actually one of the impressive things about what he's showing is it seems to not have a problem with the Wikid data funky uh ID. >> Oh yeah, we got the wrong key ID. RPI is actually [laughter] >> and then it query again for RPI. Does it solely search wiki data like can it? >> Yeah, it's and I just enable wik data mcp server so it's just wiki data >> right now. So the biggest thing I'm getting from this is that OSS the GPT OSS 20B is kind of crappy. >> Yeah, for for things that might not be the best choice to code. >> RPI is Q49211. Q49211. So it got right this time. Let's see if it's able to tell if RPI is missed. Notice if you restrict [laughter] the purple 8,000 time. Okay. >> Okay. Hey >> the line shows inception 1824 founded 1824 early then roster >> it's just a rooter >> okay holding my last graph final answer is being so different >> all right >> so yeah so it's that simple um and I have a last example here so I think this is before the effect people >> like I asked chat GPG to create like a conversation log [laughter] >> in um I think it's a workshop organization setting >> I see >> like create a CFP recruit reviewers stuff like that and uh I'm using that to start the server first. So this is not a graph. This is just like a series of JSON conversation I mean but um but just pause but you got data from somewhere and uh inputters over here. So that me something already running that server with that conversation in the back end. So what are the milestones in the in this conversation log? So it's calling the search log function which goes into the conversation log and retrieves milestones. So workshop idea confirmed. Venue can pass the share. Call for proposal release. Sub deadline for proposal. Deadline extended. Deadline extended. Review process completed. Final program schedule confirm. Confirmed. Workshop day. >> And does the log use the word milestone or is it >> uh it does use this one users. Yeah. But I can ask stuff like this. What happens in the last week of January? So the search log functions receive like a se here you go uh a several of optional parameters. So we can query by text by by speaker or agent >> tags start and end time linear of results and so forth. So those are options. >> Um so >> key happenings in the last week of January. So those things happen milestones summaries ideas one keynote can they confirmed action item updates action updates blockers and so forth. Ricky did did it show search logs interface there right before I I miss it. I did it I felt like it showed something. I'm just curious what if it showed >> Oh you mean this? Yeah, that's so this is this is essentially the the parameters that could be used in the search log function. >> Oh yeah yeah yeah >> to play with. Okay. >> Yeah. So the model decides which one they know the parameters they know that it's optional and it based on the user question the model decides whether to use any of them based on the question. So it can write regist that's interesting. >> Yeah. So the wikid data has execute sparkle. So if the model decides to create a sparkle and send that we data will take it. >> That's cool. >> Yeah. Yeah. So I'm not using any prompts. So the model is deciding to create this table and present as a table. But if I had a prompt like display results of the search in some specific way right as a summary or as narrative or as topics I don't know so the model started to do that uh >> y >> but the key thing here is this client already has MCP support so we don't have to deal with that depending on the project is using specific chatbot to run this if it does not have MCP support you need to implement that somehow >> so what does it mean to implement MCT MCP support >> so you need to yourself I don't know if there's like plugins sir I don't know if you have know that or not but um it means that you need to intercept the user message to the available tools in a separate loop to prompt the model with those tools and get the intermediate response to decide to call it or not. Uh just then you would return the exom. So you don't let the model go without injecting the the two options. So the whole thing as the name suggests is to go off and use one of these these um servers to pull back context specific context and bring that in and then uh continue on present the the model with to to the model this enhanced that's prompt with the enhanced context >> and do we do we have insight into I mean you don't have to worry about this right enough. We got more to show, but what the enhancement actually looks like. I mean, what actually goes to the model? >> Um, so the model sees like this uh this result JSON, >> right? >> Okay. >> Okay. >> So, this is basically the type of the returning variable uh in the function and then you get the value like this. But but in addition to this, there must be some sort of standard prompt language. >> No, there's one that's not. There's one there's one that the model sees >> I think that this is a JSON and tries to understand a JSON. But if if we had a prompt and if if this client would have like the prompt support because it doesn't have we could like oh my result my JSON will be formatted this way so you should parse it that way. So I think it's parsing freely without guar rails >> right now. Yeah. >> So so it's it send off what's in arguments to that search log function. It got back the result. it turns around back as context >> as additional context to the model. >> Correct. >> So, we've got like six minutes. >> Yeah. And I think that that's all what I wanted to show. If you have questions, uh I may not have all the answers, but >> I just want to show that it's it's not hard. >> What if you reasoner? H >> what if you wanted it to do some reasoning like maybe connecting an external reasoner or reason somehow? >> Yeah, I think it's like it's it's a Python code. You can do anything right even call another from there. >> So uh or run or go to a reasoner do anything. So you can basically do anything you want Python code. So but the model will be waiting for the answer right. So it's I don't know how long depends on the implementation but it will probably eventually time out if it takes too long to to respond function. So if you're bolting something on a triple store, is the strategy to build a bunch of essentially stored queries? Um what what's kind of the the strategy there? I guess >> so I I think it works best when you have like generic enough kind you need to be kind in the middle. Uh you need to be generic enough to not create like specific tools, small tools. So the tools need to be uh kind of in a level that is not very specific but not too generic. So we data is I find it interesting because has search items and search property. So search items is search entity. I don't know why they call items but so anything that the user searches uh they try to find an an entity or a property and then given you have either this or this you get the claims which are the triples out of that entity. >> Oh okay. >> Uh >> so claims are statements triples. >> Yeah. I don't know why they call claims or items. That doesn't sound like wikid data. >> No. uh and claim values uh >> but you could get the whole instance and class hierarchy. >> You could that as well. Yeah. And what's it give it to you in >> as one you you you send the uh the entity uh and returns the I think the types of the entity >> like as a hierarchy human animal >> but you could have you could encapsulate much more powerful functionality. I mean this is pretty atomic. Yeah, because it's interesting because if if you have like the right balance between like specifics and generic, the model has enough freedom to call again if it hasn't found an an answer in the first call, right? So for instance, if I ask what's the difference between Deborah McIz and Jando's careers, it will call that search items twice. one for one for Jambler, one fabra and then you compare u if I have something very specific I might so you need to experiment with I think uh and see what level of abstraction for those tools works best. But if you if you've got a triple store that you want to build um you know it you've built it for certain specific tasks um where they're you're going to be making certain kinds of queries that are essentially like reports. You could have more your functions could be more complex. >> Yes. >> Yes. Yes. um you know and and provide more information that's more ready to go. Um whereas this is a more much more general purpose needs to be more general purpose so it's it's lower level. Um so I think we probably should let Enrique get on his way. So um >> yeah I think it's a matter of experimenting with which level of abstraction you need for a specific case. Yeah. >> So for for those working on TMS, is this giving you some ideas? >> It is. Oh, it is. [laughter] >> So I'm gonna I'm gonna stop the record. Thank you very much, Enrique. Let's let's thank Enrique. Sure.