Latest / Agentic Conversations (formally mlops.community) / 10 Cities. 4 Countries. One Unexpected MCP Lesson.

Transcript

10 Cities. 4 Countries. One Unexpected MCP Lesson. · Agentic Conversations (formally mlops.community)

• 0:00
  All right, so at the beginning of April I went to New York for
• 0:03
  the MCP Dev Summit, and while I was there I got the chance to
• 0:07
  record a few podcasts with attendees that were at the
• 0:11
  event. This is one of those
• 0:12
  conversations. Hope you enjoy.
• 0:14
  I think the main thing is that people don't quite understand
• 0:16
  what it is that easily. People think it's some way of
• 0:18
  mixing AI into your API in some way.
• 0:20
  Like API is going to change what it does and magically it's going
• 0:23
  to sort of do new things, which of course isn't what it is.
• 0:26
  It's AI using AP is there's a lot of questions you get that I
• 0:30
  can't answer that well because they're about other ways to
• 0:32
  actually like inject prompts back into the LM or you know,
• 0:35
  other things. But I think part of it is just
• 0:37
  that if you're using MCP to use your API, it's good to not
• 0:41
  exposed to much of it, and it's good to tell us if API be very
• 0:43
  secure in the first place. Hi, I'm Alex Saltkeever and I'm
• 0:57
  here at MCP Dev Summit North America in New York City.
• 1:01
  I'm here with Ben Morse of Deep Bell, which is a Tell us what it
• 1:06
  is, please. What is Deepel?
• 1:08
  Deepel is AAI language translation company based in
• 1:12
  Germany. We've got about 1000 employees,
• 1:15
  so it's like a popular like phone app and web app and such
• 1:18
  thing in countries like Germany and other European countries and
• 1:22
  also Japan and even Korea not as well known in North America.
• 1:26
  So hire people like me to be here in North America and say
• 1:28
  hey, we exist. Hello.
• 1:30
  Hi, here we. Are thank you.
• 1:33
  The reason I wanted to talk to you though is your recent MCP
• 1:36
  roadshow. 10 countries in 10 days I believe it was.
• 1:39
  It wasn't quite 10 countries, but there were 10 days.
• 1:42
  But there were. How did?
• 1:42
  This How did this come about? This mad plan to spread MCP to
• 1:46
  the world. Yeah, it was kind of, it just
• 1:48
  kind of grew organically. So when I got to Deep L and
• 1:51
  became the developer advocate, a person had this remit to make it
• 1:55
  to people knew what Deep L was in North America, and the
• 1:59
  developers knew we existed. So I had to go out there in
• 2:01
  communities and just find people.
• 2:04
  And so I was looking for ways to do that.
• 2:06
  I thought, here's a hackathon in San Francisco.
• 2:08
  Like this was like May 2nd. In May, Sorry. 10 cities, yeah,
• 2:13
  it was 10 planets and 10 light years.
• 2:15
  That was, it was 10 cities. I think 10 talks in four
• 2:19
  countries. I think, yeah, we're going to go
• 2:22
  into the US of course, Canada, France, Germany.
• 2:24
  I think those are the countries this demand in more places too,
• 2:27
  certainly for this. So anyway, so I thought, OK,
• 2:30
  we'll be bold. We'll go to San Francisco and
• 2:31
  sponsor hackathon there. And I was, you know, I was
• 2:35
  following AI, following agentic AI, listening to lots of
• 2:38
  podcasts and trying things out. But I did what MCP was at all.
• 2:41
  I just kept hearing about it. I thought, what are we going to
• 2:43
  do to go to San Francisco all these cool, like I hipsters, you
• 2:47
  know, with all their cool AI companies based on the Area 75
• 2:50
  hackathon. So exactly.
• 2:52
  Yeah. And like, I noticed you bell at
• 2:54
  all we're going to do. But, you know, I'm pretty brave
• 2:56
  sometimes. So I figured we would just go
• 2:58
  there. And then I heard about MCP and I
• 3:01
  thought we should build an MCP server.
• 3:03
  This is like the latest hotness, just on the fly.
• 3:05
  Like Yolo MCP. That was the idea.
• 3:07
  And then there was an engineer who knew about it too and wanted
• 3:09
  to try it out out and his manager thought was a good idea.
• 3:12
  So we just built an MCP server, mostly him, his name is Akash
• 3:15
  and he did most of the work and we just kind of helped him like
• 3:18
  debug a little bit and then so you.
• 3:20
  Did this in work essentially like a week or something right?
• 3:23
  It was pretty quick yeah, but you can do with MCP, as we've
• 3:25
  learned isn't that hard to actually use if you know how to
• 3:28
  use it which we didn't of course.
• 3:30
  So there were some confusing things that happens and
• 3:32
  confusing bugs but basically finished it.
• 3:35
  I went down to this hackathon give me a talk of the bidding
• 3:37
  like hey you get MCP server. Everyone just figured one of the
• 3:40
  many AI companies that existed people accepted us as one of
• 3:42
  their own and it was actually pretty successful People didn't
• 3:45
  use it much in hackathon honestly they kind of used API
• 3:47
  instead but having having it meant that we were like a serve
• 3:50
  a player in the scene. It was kind of like wearing the
• 3:52
  right shirt or the right classes we should be knew what was going
• 3:55
  on with MCP. That's how it all began, just
• 3:58
  kind of as a spontaneous idea to fit into hackathon.
• 4:01
  And so how did it progress? Because it sounded like you
• 4:03
  learned a lot from being essentially 10 MCP hackathons in
• 4:08
  four countries over a very short span of time.
• 4:11
  I'm kind of curious A, what you learned B what you saw that was
• 4:13
  different in the audiences, the, or even the tenor of how people
• 4:17
  were approaching MCP, how it was different, you know, like what,
• 4:19
  what did you learn? I.
• 4:20
  Just thought it was really good topic and I just got really
• 4:22
  interested in the topic myself. I mean, tech these days is
• 4:25
  really fascinating. It's changing all the time and
• 4:27
  AI is so interesting. I got really deep into actually
• 4:29
  how it worked. I thought how could this
• 4:31
  possibly happen? Like this is an LM.
• 4:33
  It just like says stuff that talks to us.
• 4:35
  We know really what it is. It takes in tokens as input, put
• 4:38
  that tokens as output. How does it like use tools?
• 4:41
  And none of the explanations I saw talked about how it actually
• 4:43
  worked. They're all kind of similar and
• 4:45
  talked about the USBC for AI and didn't talk about what actually
• 4:48
  happened. I did a lot of thinking about
• 4:50
  it, investigating, asking friends who are more deeply
• 4:52
  involved than I was, and so I really understood it well.
• 4:55
  And then I thought, well, I have to give a talk in June in New
• 4:57
  York. I'll give a workshop at MCP
• 4:59
  because I think it should be a thing everyone understands how
• 5:01
  to do. So this would be a good reason
• 5:02
  for me to figure out like in a couple weeks how to really works
• 5:05
  well enough to actually teach. People had to use it, which is a
• 5:07
  big challenge to actually teach the thing you don't really know
• 5:09
  at all. But again, you try to go for
• 5:11
  challenges in life and this makes us learn things.
• 5:13
  So I realized it wasn't that hard to actually make a server.
• 5:16
  Like you could actually make a server and seven lines of
• 5:18
  Python, you could explain it actually pretty easily.
• 5:21
  I found like the inspirations for it, like the tool former
• 5:24
  paper from a couple years ago that talked about how to
• 5:26
  actually just train the model with the data that wasn't just
• 5:31
  data, but also these faux function calls that then could
• 5:34
  be spat out by the model later on as actual function calls
• 5:37
  intercepted and then uses actual function calls.
• 5:40
  So I gave this workshop and it was surprisingly successful.
• 5:44
  Like the room was packed. People had a lot of questions.
• 5:47
  They actually listened. I think for the most part they
• 5:48
  didn't ignore me as much as they might normally ignore me
• 5:51
  afterwards. People were like, the next day I
• 5:53
  built an MCP server, I did it, here it is.
• 5:55
  So I realized. There was demand.
• 5:56
  This is great. Then you're like, oh.
• 5:59
  Yeah, people really were interested in the topic and it's
• 6:01
  a nice entry point to learning about how to use AI because I
• 6:05
  realized you could build a server pretty quickly.
• 6:07
  I made a sample server that was just basically it told jokes.
• 6:10
  So the the the simplest tool just returned the same joke
• 6:13
  every time. So very easy to understand.
• 6:15
  You know, just return the same string every time and the joke
• 6:17
  the API found is very easy to use The API cube is necessary.
• 6:20
  You can just call the API as a REST API.
• 6:23
  So I made more tools to get various kinds of jokes and I
• 6:26
  thought it was pretty easy to understand how it worked.
• 6:27
  It didn't look like complicated code or anything.
• 6:30
  Once you install a fast MZP or your favorite client library,
• 6:33
  you can just kind of do it. So I made some sample servers
• 6:35
  like this and I started using these and more and more
• 6:37
  workshops because people always wanted to hear about MCP.
• 6:41
  It was a good topic for talks. Obviously.
• 6:43
  So like over the next 7-8, we like what changed, what
• 6:46
  progressed is was there like, yeah, sort of a, you know,
• 6:50
  things that at the end that were very different or.
• 6:52
  I did a bunch of Google Dough Fest, actually, because I used
• 6:55
  to work at Google. I had friends there, so they had
• 6:57
  me come in October, actually. My, my, my, my indie rock band
• 7:00
  played, opened the show. Then I gave MCP Talk after that,
• 7:04
  which is kind of a funny combination.
• 7:06
  Possibly what changed there were mostly it was that there were
• 7:11
  cases of people who were more interested or less interested.
• 7:13
  People needed to know about it more in some cases, less in some
• 7:16
  cases. But I think the demand is still
• 7:18
  as tied as it was before. I did my last workshop on Pure
• 7:21
  MCP in February in Philadelphia, and it was for an audience of
• 7:26
  the eye specialist who really understood the train models more
• 7:28
  than I do. And they were so interested in
• 7:30
  the talk and didn't know these things about how it worked.
• 7:32
  And I mean, I think the main thing is that people don't quite
• 7:35
  understand what it is that easily.
• 7:37
  People think it's a way of mixing AI into your API in some
• 7:40
  way. Like API is going to change what
• 7:41
  it does and magically it's going to sort of do new things, which
• 7:44
  of course isn't what it is. It's AI using AP is.
• 7:48
  So I think they like the idea that you can actually get the AI
• 7:51
  to act in a rational way, like use one of these tools over
• 7:54
  here, but use API other kinds of ways.
• 7:57
  Yes, restricted, you know, it's, it's discreet, it's not as
• 8:01
  stochastic. But to your question, what
• 8:03
  actually has changed since then? Not much actually has changed,
• 8:06
  to be honest. That's cool.
• 8:08
  So essentially the the like the protocol people are starting to
• 8:11
  get in and there's the questions are better.
• 8:13
  And the desire to learn is still as big as it was before.
• 8:16
  The questions are kind of the same questions like So what?
• 8:18
  Are like the top five questions that people get or the
• 8:20
  misconceptions or the I mean, if you were going to redesign the
• 8:22
  course from the ground up, what, what would you do differently?
• 8:25
  I would like to show how it worked by actually showing an
• 8:28
  example. Actually, it's not that hard to
• 8:29
  do. You actually can go to Claude or
• 8:31
  your LM and just tell us, OK, now you can use tools with this
• 8:34
  protocol over here and put in some fake XML or Jason and it
• 8:38
  will say, OK, I'll do that. I'll begin spitting out things
• 8:41
  using those tools. You could then say, what's 5 * 8
• 8:44
  use your new math tool and uses the math tool it already knows
• 8:46
  about. And instead of playing out 40,
• 8:48
  it says, you know, like angle brackets math, 5 * 8 close angle
• 8:53
  brackets math. So now the LMS are good at
• 8:56
  actually using tools and smart about it.
• 8:57
  I think I tried to talk about how it actually works.
• 8:59
  You know, it actually generates a prompt and the prompt is the
• 9:03
  thing which gets in the system prompt.
• 9:04
  And this explains how the tools work is why your contacts went
• 9:07
  dull and get you, not because you have a lot of stuff in
• 9:08
  there. I would explain how it worked
• 9:10
  and what it is. People wouldn't be as surprised
• 9:13
  by what it actually is. So I still get questions about
• 9:17
  that, like how it actually works and questions that reflect what
• 9:20
  really is going on. But more common questions I get
• 9:23
  are about security. I was going to set.
• 9:26
  So what do you tell them? I tell them be careful because
• 9:29
  it's kind of like shareware. I mean, if you're downloading
• 9:31
  someone else's code to your laptop and letting it run, then
• 9:34
  be careful about that kind of thing on the design.
• 9:36
  Side what are the because I mean you know the project has a long
• 9:40
  security best practices doc. What do you intend to starve
• 9:42
  with? And helping people.
• 9:43
  There's a lot of questions you get that I can't answer that
• 9:46
  well because they're about are there ways to actually like
• 9:48
  inject prompts back into the LLM or you know, other things.
• 9:52
  But I think part of it is just that if you're using an MCP to
• 9:54
  use your API, it's good to not expose too much of it and it's
• 9:57
  good to tell us if the API be very secure in the first place.
• 10:00
  And then it comes down to API security.
• 10:02
  You can use the same kinds of things, I think, to secure API
• 10:05
  and secure API against LMS using your API.
• 10:09
  And then, yeah, look for those best practices and
• 10:10
  authorization. Make sure that people can't just
• 10:13
  like easy spin up gigantic queries that will, you know,
• 10:17
  slow down your your servers and things like that.
• 10:20
  Honestly, I don't have the best advice for security.
• 10:23
  I think it's still a thing people have to deal with
• 10:24
  themselves and looked into themselves.
• 10:27
  My best advice is still just make sure your stuff is secure,
• 10:31
  you know, be careful about it. Make sure you do the usual
• 10:33
  things, sanitizing, input, all those kinds of things usually
• 10:36
  would do. Look out for authorization
• 10:38
  problems, those kinds of things. And so as you kind of traveled
• 10:41
  from one of these to the next, did you see differences in the
• 10:45
  way people thought about it, you know, like the AI experts versus
• 10:49
  the noobs or people that are sort of full stack guys?
• 10:51
  Or, you know, what were the difference in the questions you
• 10:54
  got from the different audiences?
• 10:56
  I couldn't always tell, honestly.
• 10:57
  That's a good question. I should be able to tell, but
• 10:59
  it's just different places and different kinds of questions and
• 11:02
  different kinds of levels of interest.
• 11:04
  I gave one talk actually at a it was actually at a Business
• 11:07
  School in Canada. It was another Google Dev Fest,
• 11:10
  but the students weren't that interested.
• 11:11
  The professionals were more interested.
• 11:14
  This talk I gave, I think at all was API days in Paris.
• 11:17
  The room was really full. People were outside the room and
• 11:20
  they're really just curious about what this thing was, how
• 11:22
  they could use it themselves. I think we've estimate how
• 11:24
  people actually might know things because people that are
• 11:27
  really involved in a technology really don't think everyone
• 11:29
  knows what they're doing. But everyone else out there is
• 11:31
  just kind of like, I saw this thing, I saw an article about
• 11:34
  this thing somewhere, but what is this thing?
• 11:36
  I don't quite know what it is. Unless you actually use it, you
• 11:38
  don't really get what it is. So if you were going to try to
• 11:41
  summarize like, you know, the the quick start guide in at all
• 11:48
  5 minutes or whatever, what what, what are the principles
• 11:50
  you'd go over? I would just show how to build a
• 11:53
  really simple MCP server. And yeah, I talked about the the
• 11:55
  principles first of like basically what it is.
• 11:58
  It's a way you can get an LM to actually make these faux
• 12:02
  function calls. And you then control the
• 12:05
  function calls yourself. It calls your server.
• 12:07
  Your server is working over here.
• 12:08
  It's just basically functions. It can call your API or not, but
• 12:11
  it's all it is, is basically you write functions of them.
• 12:13
  MCP server can go ahead and call if it wants to and then show how
• 12:18
  to build some simple ones. You know, simple simple things.
• 12:21
  Once you get simple ideas down, the risk can follow from there.
• 12:23
  But I found harder those MCP apps and other kinds of new
• 12:26
  things, they don't always work for me quite yet.
• 12:29
  And the harder did you explain like the protocol for MCP apps
• 12:31
  is more complicated, MCP UI, this sort of thing.
• 12:34
  Yeah, I have not had a simple way to actually explain how that
• 12:36
  works or to even show how that works.
• 12:38
  So were you involved with the design and construction at DPL
• 12:41
  of the MCP server? I mean, did you kind of watch
• 12:43
  the process? Yeah.
• 12:44
  I kind of maintained myself now, so.
• 12:46
  So, So what, what were the conventions around that?
• 12:49
  Because I mean, it's kind of interesting that I mean, the
• 12:51
  idea of a language server over MCP, yeah, you know, like sort
• 12:55
  of why or what's different or I mean, how do you design it, you
• 12:59
  know, so it's different than a straight up API or something
• 13:01
  like that. Yeah, it's not that different.
• 13:03
  I mean, it's mostly about trying to find the most commonly used
• 13:05
  functions of the API and make them really easy to use, like a
• 13:09
  translation texts, you know, sending texts, getting text back
• 13:13
  and just trying to make sure you can look for errors like our
• 13:16
  language codes used correctly. Did you require a country code
• 13:19
  over here or not? And give helpful information to
• 13:22
  the LM so it knows what to actually do.
• 13:23
  And when I first did this, I remember it would have failed
• 13:25
  because we use language codes incorrectly.
• 13:27
  It would use lowercase and uppercase sometimes and things
• 13:29
  like that. So it's not that forgiving yet,
• 13:32
  but now it's much more forgiving because one, we made it more
• 13:34
  forgiving by giving more information in the annotations,
• 13:37
  annotations and so on. And two, LM service has gotten
• 13:40
  smarter of what they try. So when you started, what were
• 13:42
  what were you working on? I mean what was the LLM in the
• 13:45
  back what you were piping? Into I usually use cloud because
• 13:47
  cloud desktop was the most fun thing to demo in and it's kind
• 13:50
  of make make conversation with you and it would try things out
• 13:52
  and try to be helpful. But it's gotten much better at
• 13:54
  this stuff. Like last time I gave him this,
• 13:56
  I, I did the talk at Drupacon last week and I wrote an MCP
• 13:59
  server that would look for translations existing in your
• 14:04
  Drupal site. If there weren't any, it was
• 14:05
  just making new ones for you and send those over for you.
• 14:08
  And it began to do all this stuff by itself.
• 14:10
  I gave it a couple of tools and it would, it would use DPL's MCP
• 14:13
  server. It's already using my clawed
• 14:15
  desktop and use my, when I made for Drupal.
• 14:17
  And it was just all kinds of things.
• 14:19
  Do you want me to translate this into Danish for you?
• 14:21
  And here it is. And it also would comment on
• 14:23
  translations because it also knows translations itself being,
• 14:26
  you know, also clawed behind the scenes to comments on the
• 14:29
  translations and other kinds of ways did.
• 14:30
  You I mean, so essentially you had it almost like Co browsing
• 14:33
  and like, Hey, look at this Drupal site and you know, and
• 14:36
  support of the CMS and look at sort of all the meta text or
• 14:39
  whatever and and tell me what you think it I mean, and then it
• 14:42
  went on its own. It's like, Oh, by the way.
• 14:44
  Yeah, I give access to look at the nodes and see what was in
• 14:46
  the nodes, the article nodes, other kinds of those that we
• 14:48
  just look at these things and start making suggestions.
• 14:50
  But I didn't expect at all. I thought require a lot more
• 14:53
  human intervention but really tries to be helpful.
• 14:56
  And and that's like something that in the past weren't there
• 14:59
  specialized tools that you would, you would, you know, use
• 15:01
  for translation versus, you know, sort of generic, you know,
• 15:05
  LLM. That was the idea exactly
• 15:07
  because people use these things called cat tools.
• 15:09
  Whether these large things are often very old looking
• 15:12
  interfaces that are kind of like a scary looking version of Excel
• 15:15
  where the send off stuff from machine translation has a first
• 15:17
  draft, Then the translator will then modify them as they want to
• 15:20
  and it can go and find with their text you've translated
• 15:23
  before is present in new text and they can try to suggest more
• 15:26
  things for you. So I thought if you're actually
• 15:28
  a translator working for a Drupal site, you're not going to
• 15:30
  want to go and probably walk in and like start typing things as
• 15:32
  a Drupal because it could be a little confusing.
• 15:34
  Maybe you're going to want to use your favorite LM and just
• 15:37
  talk to it, you know, use claw or ChatGPT or something.
• 15:40
  And it was surprisingly effective at that.
• 15:42
  Like it was really pretty friendly.
• 15:44
  So you almost discovered a new product category for deep L
• 15:47
  without realizing what you're. Doing yeah, maybe we'll actually
• 15:49
  make this into a product people could use.
• 15:51
  I mean, I think in these MCP apps you could actually.
• 15:53
  What it can't do now that what server I made like in a couple
• 15:55
  of days is you, you can't really modify things yourself that
• 16:00
  easily. You can, but it's not.
• 16:01
  It's mostly read only. It's like, you know, it'll say,
• 16:04
  OK, here's the translation for you.
• 16:05
  Do you want to use this or not? You can type a new translation,
• 16:08
  you can copy and paste it and change things.
• 16:10
  It should be more like an interface, more like a cat to
• 16:11
  where you have little boxes and here's the translations.
• 16:14
  Do you like these things or not? Feel free to modify these
• 16:16
  things. Send them all off as a batch to
• 16:18
  the Drupal site. That'd be a better way to do it.
• 16:22
  From my understanding to the audience is understanding how is
• 16:26
  Deep L's translation layer differentiated or or you know
• 16:30
  better than what you might get out of the frontier LMS right
• 16:33
  now. Yeah, that's an interesting
• 16:35
  question. Because Deep L first came out in
• 16:37
  2017. They had the best translation
• 16:39
  anywhere because they had their own models.
• 16:41
  They trained to do these things with linguist help, and it's
• 16:43
  better than Google Translate, better than anything else.
• 16:45
  LMS came out with a pretty good translation, just kind of
• 16:47
  out-of-the-box. So Deep L has now trying to stay
• 16:50
  ahead by just taking models, building its own models, fine
• 16:53
  tuning models, and just doing more and more testing with
• 16:56
  linguists. More and more fine tuning and
• 16:59
  studies and tests do show that Deep L is better for most
• 17:01
  language pairs in most cases, often substantially better.
• 17:05
  But it's more of a race now because LMS can do pretty good
• 17:07
  translation. If you want very good
• 17:09
  translation, you should use though, like a company like Deep
• 17:11
  L that does it. You know, only this kind of
• 17:13
  stuff. People often say they prefer it
• 17:15
  and people will use it because it's just better.
• 17:17
  You're translating stuff kind of casually for fun or just a few
• 17:20
  strings on your site. You could probably use an ALM
• 17:22
  out there and do it that way. So if anybody wants to use the
• 17:26
  Dibel MTP server, they can basically integrate it into like
• 17:30
  as a front layer in front of anything that they're doing.
• 17:33
  Yeah, just stick it in there, try it out, and your LM will
• 17:36
  know to use DIBEL usually instead of its own translation
• 17:39
  these days, because it will know there's a tool for that.
• 17:41
  They're kind of biased towards using tools as opposed to not
• 17:43
  using tools. Yeah, and just gives you a
• 17:45
  higher quality of translation. If you have things where your
• 17:48
  translation quality really matters to you, You know it's
• 17:50
  pretty cheap actually to DPL. So were people using, were
• 17:53
  people actually building things, MCP apps or whatever with the
• 17:57
  DPL MCP during this tour? I mean, they just start to see
• 18:00
  people try to do translation stuff.
• 18:02
  I. Can't tell if they're
• 18:03
  coordinated or not. We've seen growth quite a bit in
• 18:05
  our DPL and CV server usage is going growing.
• 18:08
  It's still very small compared to the actual API, which is very
• 18:10
  large, of course, but is growing very fast.
• 18:13
  Even folks in that as you were building things with it and
• 18:15
  they're making various things with it.
• 18:16
  I think they just find it more convenient than using the API
• 18:19
  sometimes because our tools are more carefully defined.
• 18:22
  You know, like here's this translation text, here's
• 18:24
  translating a document. And there are more more
• 18:27
  guardrails around these things so that I think it's just easier
• 18:29
  to use. So like essentially the
• 18:31
  flexibility of MCP was something that as you were traveling
• 18:34
  around people were it's just much easier for or people that
• 18:37
  don't want to have to think in terms of the structures of an
• 18:39
  API or. API is the reverse because API
• 18:41
  may be a whole new thing, like what is this API?
• 18:43
  How do I use this thing? MCP?
• 18:45
  I think it's just sort of simpler.
• 18:46
  Here's the tool you know. You give this tool these two or
• 18:49
  three parameters and then it gives you the results you're
• 18:51
  looking for. Very cool.
• 18:54
  So when's the next tour? That's a good question.
• 18:58
  The last talk was just last week, but actually trying to
• 19:00
  focus on making some scale content like writing more of the
• 19:03
  traditional devrel things like more documentation, make some
• 19:06
  videos, things like that. So I'm taking a little break
• 19:08
  from traveling out to do those things and I want to go back out
• 19:12
  there and talk again with you. More magic AI and more of those
• 19:15
  sorts of topics that are also of course hot topics, you know,
• 19:17
  building agents, which is super hot as you know.
• 19:20
  But MCP, I mean, people still don't know what it is in most
• 19:23
  cases. Outside of in New York, San
• 19:25
  Francisco. Or people that are just
• 19:27
  following these things and who are working with it and try it
• 19:29
  themselves. But people are still like, what
• 19:31
  is this thing outside the demand bubble?
• 19:32
  We always forget the demand is still there for just, you know,
• 19:35
  introductory content. It's always there for
• 19:37
  everything. So the next time you do a tour,
• 19:39
  what? What are you going to do
• 19:40
  differently when you design your MCP curriculum?
• 19:43
  I'm going to try to make MCP app this time to see if I can get
• 19:46
  all the way to making apps and using the new fancier features
• 19:49
  and seeing how that goes. It's a good question.
• 19:51
  I feel like what I had did before, how can I improve it?
• 19:55
  I feel like actually if I could get people to do more coding
• 19:57
  themselves, that'd be nice, because when they learn things
• 20:00
  for real, I'd have workshops. People often would come in
• 20:02
  without a computer and wanted to listen to a talk.
• 20:05
  Right. So they didn't do hands on as
• 20:06
  much actually. I first did.
• 20:07
  I thought I'd do a workshop and people just came and they were
• 20:09
  looking at me like, OK, talk to us.
• 20:11
  They made it into a talk spontaneously and they did it
• 20:14
  later in some cases, try it themselves.
• 20:17
  I think it's probably better. So I think if I was going to do
• 20:19
  it again, if I do do it again, I'll try to get folks to
• 20:22
  actually really write some code and know that you can use
• 20:25
  Python, you can use TypeScript, you can use your favorite
• 20:27
  language. You can do it pretty quickly.
• 20:30
  As long as you've got something installed on your computer,
• 20:32
  which is an LLM, which is an AI client, which maybe you build
• 20:34
  people don't haven't got yet. If they have an AI client, they
• 20:36
  can do it. You can install your library,
• 20:39
  you can just go and do it. And it feels, it feels like
• 20:42
  magic actually, an LMS are actually like using your tool
• 20:45
  and you see it doing that and then responding with your
• 20:47
  information you've given it is really kind of cool.
• 20:51
  I, I mean, I, I personally, I'm not a coder.
• 20:54
  I've been, well, I'm a dangerous coder, right?
• 20:56
  Like I could code stuff you would not.
• 20:57
  If one's a coder. You would not want me to put it
• 20:59
  in prod, but yeah, it's just blowing me away.
• 21:01
  How you know, you can essentially go into Claude code,
• 21:03
  give give it a spec. It'll recommend the tool.
• 21:06
  Or if there's MCP servers out there that'll make sense, it'll
• 21:09
  recommend them. You know, it's a, it's, it's,
• 21:11
  it's really, really different. Yeah, watching like you get like
• 21:13
  Claude or Taji and they actually use your tools in the call to
• 21:16
  three different tools. Like I could do a thing, I could
• 21:19
  give a demo, which is like, OK, I'm going to have it use the
• 21:22
  translation tool now. I'm going to ask it to translate
• 21:24
  this thing, like find a joke and then translate the joke.
• 21:27
  We'll call 2 tools. People are on the edge of their
• 21:29
  seats. They're going to have actually
• 21:31
  like tell, tell a joke, translate it into French and put
• 21:33
  it into a spreadsheet. Three tools he watches Claude
• 21:37
  spins. That actually usually does it
• 21:38
  right. So it's just, it's exciting,
• 21:40
  it's oddly suspenseful and gratifying because.
• 21:43
  Usually it works. Thank you very much for coming
• 21:45
  out on the. Podcast.
• 21:46
  Thank you. All the audience one last time,
• 21:48
  you know, like who you are and what you do again in case you're
• 21:51
  catching a tail end. In case you forgot or missed the
• 21:54
  beginning because whatever reason you're doing something
• 21:56
  else, I'm Ben Morris. I'm from DPL Developer Advocate
• 21:59
  and we're talking about MCP servers, how they work and
• 22:02
  Deepel is a language translation company that uses AI makes on
• 22:05
  models to make language translation as good as possible.
• 22:09
  Thanks for coming. Thanks for having.
• 22:11
  Me enjoy MCP dev solar. I will, thanks.