Latest / Agentic Conversations (formally mlops.community) / The Creator of FastMCP Explains the Future of MCP

Transcript

The Creator of FastMCP Explains the Future of MCP · Agentic Conversations (formally mlops.community)

• 0:00
  I came to New York, to the Dev Summit to give a talk, and I had
• 0:03
  been working and working and working on my slides, and I got
• 0:06
  here to the hotel and I opened my computer and my open call had
• 0:08
  decided, and it's in its exuberance to help.
• 0:11
  And it's infinite wisdom. It's infinite wisdom.
• 0:13
  It had decided to delete all of my slides.
• 0:15
  I have now built a piece of software that I think is a
• 0:17
  really clean way of going from zero to 60 on building a server
• 0:20
  in like 2 seconds. So I put it in the world and
• 0:22
  honestly, I kind of forget about it.
• 0:25
  I think I have an opinion that most people here at the
• 0:27
  conference don't share, which is based on experience, and I'm
• 0:30
  pretty confident it's right. All right, so give me the story
• 0:45
  of Fast MCP. You want dive right in on that.
• 0:47
  Let's go right to it. Let's go fast, MCP.
• 0:50
  So the MCP is announced November of 2024 and I see it and, and as
• 0:56
  you know, as we were talking prefect, we're all about
• 0:57
  workflows and connecting software and gluing things
• 1:00
  together. And so I've been, I've been
• 1:01
  trying to figure out for a few years, what does it mean to
• 1:04
  automate the agents? Like how do we plug them into
• 1:07
  our software ecosystem? And so MCP has announced and I'm
• 1:10
  like, maybe this is the thing. Maybe this is finally the
• 1:13
  protocol and it's from a big enough company.
• 1:14
  It's going to be the thing. And I try to use it and it's
• 1:17
  just really hard to use the SDK is complicated.
• 1:19
  It takes me a couple hours to set up a basic server, and as I
• 1:24
  may have a tendency to do. What was the first server you
• 1:26
  tried to set up? I don't even remember anymore.
• 1:28
  It was probably like to say hi like it was just, it was just
• 1:30
  complicated to set up. It was, it was just a test and I
• 1:33
  end up writing a piece of software to make it easier to
• 1:37
  build servers and I call it fast MCP and I put it in the world.
• 1:41
  I've by this time I've forgotten the point.
• 1:43
  I, I don't even like, I couldn't tell you what server I was
• 1:45
  trying to build, but I've now built a piece of software that I
• 1:47
  think is a really clean way of going from zero to 60 on
• 1:50
  building a server in like 2 seconds.
• 1:52
  So I put it in the world and honestly, I kind of forget about
• 1:55
  it. It's like a thing.
• 1:56
  It's a little side project. And I go back to building
• 1:59
  workflows for people. And I get a call from David
• 2:01
  Soriapara maybe a couple weeks later and he says, listen, this,
• 2:05
  this is a really good developer experience.
• 2:07
  Could we put this in the official SDK?
• 2:10
  Can this become how people build MCP servers?
• 2:12
  And I was like, yeah, this is the coolest thing that's ever
• 2:15
  happened to me, that Anthropic would want to to use this open
• 2:18
  source software. So I worked with David and he
• 2:21
  ported it into the SDK. And to this day, if you use the
• 2:25
  official SDK, you import a fast MCP object and you build your
• 2:28
  server with it. And honestly, that would have
• 2:31
  been that. And that could have been the end
• 2:32
  of the story. What changed is exactly a year
• 2:35
  ago, Google and Open AI both announced that they were going
• 2:39
  to support MCP. And that's when the hype
• 2:41
  started, which I think is why we're here today at the at the
• 2:44
  Dev Summit, right? Because the hype, the hype train
• 2:46
  left the station. And when they announced that was
• 2:48
  going to happen, I think it was the catalyst that that we needed
• 2:52
  as a community to start taking this technology seriously.
• 2:55
  And I found out about this because my repo, which is
• 2:58
  essentially defunct at the time because everything had moved to
• 3:01
  the official SDK, my repo started getting thousands of
• 3:05
  GitHub stars and issues and requests and people asking for
• 3:09
  things that I realized were not appropriate for a low level SDK.
• 3:12
  People were asking for off and composability and the things you
• 3:16
  would expect from like an application like an MCP
• 3:18
  application framework. And I kind of hemmed and hawed
• 3:21
  on this for a while, but there was so much demand that it just
• 3:24
  seems like we should meet it. And so we restarted Fast MCP as
• 3:29
  a standalone project with Fast MCP 2 exactly a year ago and it
• 3:34
  grew and grew and grew and gets downloaded millions of times a
• 3:38
  day and 25,000 GitHub stars. I don't even know.
• 3:42
  It's been, it's just been a remarkably successful open
• 3:45
  source project. And, and, and that sort of
• 3:48
  brought us into this ecosystem in this community from a very
• 3:51
  different perspective. I think we, we, we had figured
• 3:54
  out recently that something like 70% of all MCP servers are
• 3:57
  written on either our Fast MCP or the, the official SDK fast
• 4:01
  MCP. So it's just it's kind of become
• 4:03
  the standard way that people build MCP servers and it's given
• 4:05
  us this fascinating insight into like what's going on, what
• 4:09
  people are doing. What should I keep in mind when
• 4:12
  building with fast MCP? Are there any gotchas that I
• 4:15
  need to know about or any trade-offs?
• 4:16
  You know, there, there are, but they're, they're
• 4:18
  well-intentioned so fast. MCP is an opinionated framework.
• 4:21
  And what I mean by that is we have an opinion on a certain
• 4:24
  kind of MCP server, a certain experience of building it.
• 4:28
  We we try to be, you know, as flexible as possible.
• 4:31
  If you're building this kind of down the fairway normal MCP
• 4:36
  server, you're going to have an amazing experience.
• 4:38
  If you want to turn every little knob and every configuration and
• 4:42
  every possible way beyond that, you're probably better off going
• 4:45
  to the SDK and getting as close to the protocol as you can.
• 4:48
  And I think this is this is a really healthy tension.
• 4:50
  The SDK is the low level way to work with a protocol and fast
• 4:55
  MCP is for 90% of people probably the easiest way to skip
• 4:59
  all of that and just get a server.
• 5:01
  It's literally in the name. It's.
• 5:02
  Literally in the name. It's a fast way of doing it.
• 5:04
  And, and the reason it's named that, it's funny, if I knew how
• 5:06
  popular it was going to be, maybe I wouldn't named it that.
• 5:08
  Although now seeing how popular it is, I'm like, maybe the name
• 5:10
  has something to do with that. But it's but it's very much an
• 5:12
  homage to Fast API which is. The Sebastian.
• 5:16
  Exactly which is the high level entry point to the Python web
• 5:20
  world on top of Starlet as its low level SDK.
• 5:23
  And so most people writing most servers will use fast API to
• 5:27
  skip a lot of the hard work. But some people will go straight
• 5:30
  into Starlight for the low level.
• 5:32
  And so we wanted to occupy a similar sort of space in
• 5:34
  people's mind mindset. Yeah, I thought it was fast just
• 5:38
  because it you can set up an MCP server so fast.
• 5:40
  I mean, you can do that too. Yeah, it works really well.
• 5:42
  So it's been this journey of, you know, from from day 2
• 5:45
  basically of MCP having essentially this front row seat
• 5:49
  to the growth and being there when nobody cared about it.
• 5:51
  And now, now how ridiculous the, you know, MCP is dead and all
• 5:56
  this stuff. Our our team threw a funeral for
• 5:57
  MCP. Last night for April Fools.
• 5:59
  OK, I was going to ask, is it an April Fools thing or was it
• 6:04
  because you were just jumping on the hype?
• 6:07
  It was conveniently an April Fool's thing.
• 6:11
  I think we decided that this was the thing to do, to poke fun at
• 6:15
  the whole debate. And then we suddenly realized it
• 6:17
  was also going to be on April Fool's and was like, oh, this is
• 6:19
  perfect, we have to do this. And we did it last night and it
• 6:22
  was an enormous success. There was a celebration of life.
• 6:25
  It was a. Celebration of life and we had a
• 6:27
  string quartet singer and we passed MCP apps.
• 6:30
  Did you have a open casket? We did not have an open casket
• 6:34
  because there was no, there was no body to be found.
• 6:37
  That's the thing about something that's not really dead, but I, I
• 6:41
  do, I do think it's ridiculous that that this is the charge.
• 6:44
  And I don't know as you and I were were chatting earlier
• 6:47
  though, I think that there is a, a small degree of legitimacy in
• 6:52
  the charge that this MCP server could be a CLI.
• 6:55
  There are some MCP servers that are not good and there are many
• 7:00
  MCP clients that do nothing more than download every single tool
• 7:04
  they find on a server and shove them into an LM's brain.
• 7:07
  And I think if that's what the client is doing, then the charge
• 7:09
  that you would be better off for the CLI is not a not terrible
• 7:14
  like, like we, we need to be honest, we need to confront what
• 7:17
  enabled folks to make that charge.
• 7:20
  That action of downloading every single tool and shoving them
• 7:23
  into an LM has very little to do with what MCP is supposed to do
• 7:26
  and enable it is the worst version of the simplest feature.
• 7:31
  And I think one thing that I was talking about here at the
• 7:34
  conference is I think that last year the name of the game was to
• 7:37
  drive adoption of MCP. And so we tolerated MCP clients
• 7:42
  that took this sort of one O 1 approach to what it meant to be
• 7:46
  an MCP client. And one of the things that I'm
• 7:48
  really excited for us to do in 2026 is build rich enough MCP
• 7:52
  experiences that it forces clients to step up.
• 7:56
  Well, you have a very opinionated view of how to do
• 7:59
  that right? I do fast MCP and probably just
• 8:01
  in your own buildings. What are some things I.
• 8:04
  Do so for us, there's a, there's a few things that we've that
• 8:07
  we've noticed. So, so there's this thing that
• 8:11
  I, that I believe that I know that I don't know that everybody
• 8:15
  here is maybe aware of. I wouldn't say that they
• 8:17
  necessarily are ignorant of. But the one thing that I believe
• 8:21
  that I'm not sure everybody else believes is that the number one
• 8:24
  use case for MCP is inside the four walls of a company for
• 8:27
  people in the company to deliver data and information and
• 8:31
  business logic to other people in the same company.
• 8:33
  Wow, OK. I mean, I've heard that a lot
• 8:36
  with like internal AP is that same kind of idea, but I haven't
• 8:41
  heard anybody say that that is one of the greatest use cases I.
• 8:45
  I really think it is just just based on who we see showing up
• 8:48
  and building servers and what they want to do and what their
• 8:50
  concerns are. And this is not something that I
• 8:52
  expected to be clear, but. Is it, is that not just a
• 8:55
  maturity thing because we're so early still?
• 8:58
  I so I, I could believe that. So it's, it's not that I don't
• 9:02
  think MCP will ever be like broadcast to a company's
• 9:04
  customers or anything like that. But I think many people,
• 9:08
  including myself, expected that the number one thing that would
• 9:11
  happen with MCP is I as an individual would write some
• 9:14
  logic that I want your agent to be able to consume.
• 9:17
  And here's how I'll do it. And then I as a company will
• 9:20
  want my customers to access my API through MCP and that's what
• 9:22
  we'll do. And so I think that that is very
• 9:24
  much in scope and I'm very excited about that.
• 9:27
  But I think what has stopped it, if there's one thing that has
• 9:30
  like become the friction, it's actually that you just, you
• 9:32
  don't know what the client is capable of except maybe tool
• 9:36
  calls. And it's hard to maybe ship a
• 9:39
  very exciting server to a client you don't know or control or
• 9:42
  trust inside your company, though I know what the client
• 9:46
  is. I don't have a trust boundary
• 9:47
  problem. I can ship more interesting
• 9:49
  code, I can ship a richer code, I can take advantage of more
• 9:51
  features. And so I think that's one
• 9:54
  dimension. Yeah, you're flipping the
• 9:56
  narrative on the head a little bit because it was very much
• 10:00
  sold to us as, look, this is how you interact with my thing, yes.
• 10:05
  Exactly this. Is your USBC remember that was
• 10:07
  the thing for the. Internet exactly.
• 10:09
  And I think that's all true. I I like that I think that will
• 10:12
  happen, but that adoption is much.
• 10:15
  I don't I don't even know if it's fair for me to say it's
• 10:17
  slower than I expected. It's the this other use case,
• 10:21
  this first party use case is exploding and we're we're seeing
• 10:25
  it, which has been especially interesting for us at Prefect
• 10:28
  because of who our customers are.
• 10:29
  Is we're seeing it especially among data teams, platform
• 10:33
  teams, ML teams, the same folks that we've been serving for
• 10:37
  years and helping them build the workflows that hydrate the
• 10:39
  dashboards and the models and stuff.
• 10:41
  I think that they either came to view or had thrust upon them a
• 10:47
  view that MCP solves this self-serve data platform
• 10:52
  problem, finally, that that every vendor under the sun has
• 10:55
  tried to solve and no one's been able to, like, handle the
• 10:57
  Infinity of complexities. So what do you do?
• 11:00
  The first version of this was a data team standing up an MCP
• 11:03
  server in front of the data warehouse and saying, look, we
• 11:05
  don't want to write another query for you.
• 11:07
  Just ask your agent to write the query and do it.
• 11:09
  And they did that for a week. And then the agent took down the
• 11:11
  warehouse and ran up a huge bill because it did select star and
• 11:14
  Big Query or who knows what. Yeah, exactly.
• 11:16
  It doesn't have any idea of what cost money.
• 11:19
  It really doesn't. For something that's supposed to
• 11:21
  to be so superhuman and you got to.
• 11:23
  Have a fin op skill and make sure that like this is our CFO
• 11:28
  skill and make sure that we do not spend more than a grand per
• 11:31
  query and it's. And it's all data warehouse, you
• 11:33
  know, precise. It's like big query.
• 11:36
  You need to work like this click house.
• 11:37
  These are the kinds of query you need to make.
• 11:40
  I actually need that in my life. I.
• 11:41
  Know I've made some of. Those mistakes.
• 11:44
  And so I think I think that that, you know, there was this
• 11:46
  evolution of first you plug it into the into the into the data
• 11:49
  warehouse and then you use a tool like fast NCP to maybe
• 11:51
  curate that a little bit and and not just give raw access.
• 11:54
  And before you know it, you have now actually built a self-serve
• 11:58
  data platform. And I, I don't think this is
• 12:01
  like uniquely the only use case, but we were shocked when the
• 12:06
  folks who showed up to talk to us about MCP, honestly, just
• 12:09
  because they see fast MCP in the world, most of them don't even
• 12:11
  realize it's prefect at first. There's this fast MCP, they
• 12:13
  assumed it was a company. They were like, oh, we want to
• 12:16
  talk to fast MCP. And I was like, well, I guess
• 12:19
  that's us. We, we, we would just, we would
• 12:22
  just hear over and over from the same folks that we were talking
• 12:25
  to in the data side. I think with one critical
• 12:28
  difference, which is on the data side, you're trying to justify
• 12:30
  the cost of this thing. It's a workflow, it's a it's an
• 12:33
  internal thing on the MCP side. Because of the hype, you have CE
• 12:37
  OS who are saying things like we must have MCP.
• 12:40
  But Prefect was open source, right?
• 12:43
  Prefect is open source. So yeah, I imagine that you were
• 12:48
  talking to a lot of people on the open source side with
• 12:50
  Prefect, and then they come and they're like, wait, same guy?
• 12:54
  Yeah, I feel like I know you from somewhere there.
• 12:56
  There was a surprising, I think one of the first customer calls
• 12:59
  we ever did was one of our customers, one of our then
• 13:03
  customers at Prefect, one of our first customer calls for the,
• 13:05
  for the Fasten speed product, which is called Horizon was one
• 13:08
  of our customers who did not fully realize it because it was
• 13:11
  a, it was, you know, we launched under a different name a few
• 13:13
  months ago. We hadn't fully brought it into
• 13:14
  the Prefect, you know, product family yet.
• 13:16
  And he gets on, he's like, I know you guys, we run all our
• 13:19
  data pipelines on your software. You do NCP 2.
• 13:22
  And we had this weird question where people like, well, how can
• 13:24
  you possibly know about both? Which is a really interesting
• 13:27
  charge. And I think my point of view is,
• 13:30
  well, it's we do know about both, but what we really know
• 13:32
  about is building software that makes things accessible to
• 13:34
  people. And then we take on the
• 13:36
  responsibility of learning all the insurance and the outs and
• 13:38
  the best practices of a data platform or an AI platform or
• 13:42
  whatever you want to call it. So the.
• 13:45
  The idea is still the same, abstracting away that complexity
• 13:49
  so that I can get what I want to get done in a easier and nicer
• 13:53
  developer experience. We, we, we want to make
• 13:55
  complicated things simple. And I love Python.
• 13:59
  My whole career is thanks to Python in some way, shape or
• 14:01
  form. I think it's an amazing way to
• 14:03
  express what you want to do and make it readable and everything.
• 14:07
  And so a lot of what we do could be kind of boiled down to like,
• 14:10
  we take complicated things, we make them feel like Python.
• 14:12
  And if they feel like Python, a lot of people find it
• 14:14
  accessible. And now all of a sudden you can
• 14:17
  do things that you maybe wouldn't know how to do or maybe
• 14:21
  would struggle to figure out the best practices for.
• 14:23
  And so one way that I really like to think about the software
• 14:25
  we build a prefect across the whole product portfolio is it's
• 14:28
  about embedding best practices in software so that you don't
• 14:32
  have to learn the best practices of this protocol or when to send
• 14:36
  a ping or, you know, all these weird little quirks that there
• 14:39
  are. You can just focus on the thing
• 14:40
  that you were hired to do, which is probably related to something
• 14:43
  unique to your company. And we'll worry about all the
• 14:46
  like connectivity of that to this broader landscape of best
• 14:49
  practices. Well, getting.
• 14:50
  Back to the opinionated nature of this, I know that with
• 14:54
  Prefect, you have some very magical moments when you're
• 14:58
  using it and you go like, look, you just type this and then that
• 15:02
  happens. Wow.
• 15:04
  And I've been in calls where people have showed others that
• 15:08
  and they've gone, wait, what? No, that's all you got to do.
• 15:12
  Oh my God. And how are you taking that same
• 15:15
  approach? Like, what are some things
• 15:17
  you're doing with MCP now that gives that magical experience
• 15:20
  because it is opinionated. And you say, look, we got to do
• 15:22
  it like this. We're going to extract away all
• 15:24
  this complexity, but we're going to give you those magical
• 15:27
  experience. Yeah.
• 15:28
  So I actually learned the hard way you can take that too far in
• 15:31
  some of the very first prefet calls.
• 15:33
  This is 8 years ago, but just to go in The Time Machine for a
• 15:36
  second, in those very first calls you we had to write less
• 15:40
  prefet code than you do today to get the same outcome.
• 15:43
  We worked really hard at that. Like today it's pretty minimal
• 15:45
  in port decorate like you're basically done.
• 15:48
  We were doing more. It was so magical and we showed
• 15:51
  it to a very early. And they didn't.
• 15:53
  Believe we were doing anything. And so it's this fascinating
• 15:56
  product lesson about you need the exact right amount of like
• 15:59
  constructive friction, let's call it, right?
• 16:01
  We don't want to be in your way, but we want you to.
• 16:05
  We want to create an expectation of what's about to happen.
• 16:07
  And if it's too magical, you are robbed of this opportunity to
• 16:11
  form an expectation about how to reason about your own software.
• 16:14
  And in a very early version, we went too far.
• 16:16
  So it's an art, it's not a science, it.
• 16:19
  Reminds me of how you don't want to give somebody a search result
• 16:22
  too fast, otherwise they think that you didn't search through
• 16:25
  all the documents. Yeah.
• 16:26
  Yeah. Aren't there studies like like
• 16:28
  you? You put this tiny like 100
• 16:29
  millisecond? Because otherwise people would
• 16:32
  be like, no, it didn't work, It didn't search everything
• 16:34
  exactly. There is a such thing as being
• 16:35
  too magical. And in Python it's so easy to be
• 16:37
  too magical. So this is this is an art.
• 16:39
  This is something you have to develop an intuition for.
• 16:43
  So in fast MCP we do it in a few ways.
• 16:45
  One of them is one that I hope most people never figure out
• 16:48
  which is that if you actually try to use the low level SDK,
• 16:50
  you might, it might take you an hour to like build a server and
• 16:53
  register all the handlers and, and in fast MCP it's very
• 16:57
  similar to prefect import decorate like you're done, you
• 16:59
  have a server. And so we're at a point where
• 17:01
  it's so prevalent, I'm not sure if people understand how magical
• 17:04
  that is. Where the magic really comes
• 17:07
  from, I think is we, we bend over backwards to make sure that
• 17:10
  when you set up a function that returns something interesting
• 17:15
  like like a pigantic model or like a table or something like
• 17:19
  that, we serialize it appropriately in the, in the
• 17:23
  protocol context. And so there's again, we're in
• 17:26
  the insurance and outs of what's best practice.
• 17:28
  The protocol can just send texts to the LLM because that's what
• 17:31
  it's expecting to get just a bunch of tokens.
• 17:33
  But there's also ways to send structured data and structured
• 17:35
  information. But in order to do that, you
• 17:37
  have to declare your schema upfront.
• 17:39
  You have to conform to that schema or the protocol will
• 17:42
  reject it. There's lots of stuff you have
• 17:43
  to do. I don't want my users to have to
• 17:45
  think so hard about their functions that now they can't
• 17:48
  return a dictionary from their function because the first they
• 17:50
  have to figure out how to declare a schema for it.
• 17:53
  And so we come up with developer experiences where if you tell us
• 17:57
  I'm going to return a a user, we're like, great, we have all
• 18:00
  the information we need to wire up the protocol for you.
• 18:03
  Just promise me you're going to return a user so that the
• 18:05
  protocol's happy and we're good to go.
• 18:07
  And I mean, who knows how many lines of code we're, we're,
• 18:10
  we're saving there a lot. But it's little things like that
• 18:14
  where my, my philosophy and all the software is to use one more
• 18:18
  feature should be one more line of code.
• 18:20
  I think that's a really nice. I don't know if we can literally
• 18:22
  do it every time, but like both in prefect and fast MCP, if you
• 18:25
  want to do one more thing, I'd like it to be one more line of
• 18:28
  code. And with this, is it dynamic so
• 18:32
  that I'm I'm saying I'm going to return a user, you understand,
• 18:35
  OK, this is what the user looks like today, but potentially that
• 18:40
  schema changes tomorrow, yes. Schema changes are really weird
• 18:44
  thing in the MCP because there's no programmatic client out there
• 18:48
  that's locked in. It's an LLM who every single
• 18:50
  time shows up and they learn the schema.
• 18:52
  So changing schemas is actually like blessedly easy.
• 18:56
  You just change it. You, you change your user model.
• 18:58
  And as long as you've annotated that function as returning the
• 19:00
  same user model, we will pick up the change the next time the
• 19:03
  server is run and it will just work and it will just flow
• 19:07
  through. And, and the goal is not for you
• 19:09
  to know that we did something magical, which is why maybe it's
• 19:13
  a, it's a surprisingly difficult question to answer.
• 19:15
  Like in the sense everything we do is to save you from ever
• 19:18
  seeing the protocol. You should write business logic
• 19:22
  that, if not for the fact that it was in an MCP server, would
• 19:25
  look exactly the same as if you wrote it in any other context.
• 19:28
  That's my goal with fast MCP, same goal with Prefect.
• 19:31
  You should write business logic that is being fully orchestrated
• 19:35
  and made resilient and all these great things that we do as an
• 19:38
  orchestrator. But you should not write your
• 19:40
  code any differently because it's going to be orchestrated.
• 19:43
  You should not write your code any differently because it's
• 19:45
  going to be an MCP server, except to the extent that you
• 19:47
  want to take advantage of features in either, of course.
• 19:52
  And so that's the guide. That's how we write magic
• 19:54
  software is we try to make it feel like you're not using MCP.
• 19:59
  We try to make it feel like you're writing Python.
• 20:01
  And by the way, now any agent in the world will be able to.
• 20:06
  Call your tool and use your business logic and and take
• 20:08
  advantage of this stuff which is which has been great.
• 20:10
  OK, so the magic of that the schema, I think this is probably
• 20:16
  a good Segway into let's talk about MCP apps.
• 20:21
  Let's talk about MCP apps I. Know you're pretty you're pretty
• 20:24
  stoked on right I'm. Really excited about MCP apps.
• 20:25
  Tell me more. MCP apps is something that's
• 20:27
  been that's been growing in the ecosystem for some time.
• 20:31
  It started as MCPUI thanks to some really amazing
• 20:34
  partnerships, it now encompasses ChatGPT apps and, and, and it's
• 20:39
  become this new top level spec extension called MCP apps.
• 20:42
  And the idea is very simple. You can now ship AUIA full UI
• 20:50
  arbitrarily complex to your users and it shows up in line in
• 20:54
  their chat with their agent. And so you can imagine from the
• 20:58
  point of view of MCP as a way to reach your customers and.
• 21:01
  Your database. And and and your databases, of
• 21:03
  course, all all into the spectrum, but from what you're
• 21:05
  talking to your customers, this is amazing.
• 21:07
  You can now ship your website to them.
• 21:09
  Are you an e-commerce store now when you ask the agent like hey.
• 21:15
  Is there a? Shirt I should buy It can show
• 21:18
  you AUI with all the shirts. You can buy one right there.
• 21:21
  You could preview them, you could save your preferences, you
• 21:24
  could log into your account even it's a full UI embedded in your
• 21:27
  chat. Theoretically you could say what
• 21:30
  do I look like with this on and then it'll generate the image
• 21:34
  right there with you. And, and maybe they do that as
• 21:36
  something they ship only to agents and it's not in their
• 21:38
  main website because you don't have a good way to interact with
• 21:40
  them over there. So this it's, it gives you the
• 21:42
  potential to build these incredibly rich experiences.
• 21:46
  But we already talked about the fact that I have a belief that
• 21:50
  the number one use case here is actually not reaching your
• 21:52
  customers, it's reaching your team.
• 21:54
  And so when you think about your team, they have a different set
• 21:57
  of needs of UIS. Most companies have countless
• 22:02
  internal tools and internal UIS and internal dashboards and 1000
• 22:07
  different ways of visualizing information and working with
• 22:10
  working with data. And all of a sudden we have a
• 22:12
  technology where the same agents that many of those teams are now
• 22:15
  using to interact with the business can actually show them
• 22:17
  more than just text. It can show them a dashboard, a
• 22:20
  chart, a table, a form. To that self-serve.
• 22:24
  Exactly this is this is what do we log into our dashboards every
• 22:27
  day to see a table, a graph, whatever it is now we can get
• 22:30
  those on demand and we can get them in our chat.
• 22:32
  We don't have to go to another place.
• 22:34
  And so I, I started really wondering about this because
• 22:39
  there's, there's a small thing I, I'm not crazy about with MCP
• 22:42
  apps, which is sort of an unreasonable expectation for me
• 22:46
  to have, but it's a different ecosystem from my Python
• 22:49
  ecosystem. They're front ends, they're
• 22:51
  written in HTML and JavaScript and CSS.
• 22:54
  And historically Python is just not a place where you would, I
• 22:57
  mean, you wouldn't even dream of doing that.
• 22:59
  But I, I wasn't crazy about the idea that our fast MCP user base
• 23:04
  and the majority of MCP servers are written in Python.
• 23:06
  I wasn't crazy about the idea that our fast MCP user base
• 23:09
  would have to go and build their MCP apps in a different
• 23:12
  ecosystem. There's there's a argument that
• 23:16
  I find very unconvincing that just will have an LLM do it so
• 23:19
  great. Obviously an LLM could do that
• 23:21
  and there are many skills for doing this.
• 23:23
  Front end skills? Yeah, exactly.
• 23:25
  Front end skills, but the but the challenge is not building
• 23:27
  it. The challenge is now it is part
• 23:29
  of your server. It is it is part of a stack and
• 23:32
  it has things you know it needs to call tools on your server.
• 23:35
  That means it needs the name of the tool on the server.
• 23:37
  There's just, it's just the complexity surface has just
• 23:39
  exploded and our ability to assist you in fasten CP has not
• 23:44
  changed at all because you are essentially outside of our
• 23:47
  ecosystem in a black box. And I started wondering like how
• 23:50
  can I help my users take advantage of this feature if the
• 23:53
  feature is entirely outside of our ecosystem and the all right?
• 23:58
  It's landing for me now. So that's the sort of challenge.
• 24:01
  The solution was rooted in this insight that most use cases of
• 24:08
  NCP apps today are not going to be full polished branded front
• 24:13
  end experiences. They're going to be a bunch of
• 24:15
  tables and charts. And like I, I, I know the state
• 24:18
  space, it's a constrained problem.
• 24:20
  And so the other insight that I had is we don't need to build
• 24:23
  UIS as much as we sort of just need to compose and parameterize
• 24:28
  them. What do I want to show?
• 24:29
  I want to show a data table and a chart.
• 24:30
  Great. Can I come up with a Python DSL
• 24:33
  for composing pre-existing UI components in a way that makes
• 24:37
  sense and that is connected to my MCP server?
• 24:39
  And it took a little bit of trial and error and by a little
• 24:42
  bit I mean about a month. But we came up with ADSL that I
• 24:46
  think is really clean for it's all Python context managers and
• 24:50
  you just declare the UI you want.
• 24:53
  And in fast MCP you return that UI from a tool.
• 24:57
  And fast MCP does all the rest and wires up all the UI in the
• 25:00
  the application, everything. And the the result is you go
• 25:02
  from having a tool that returns the records from a transactions
• 25:06
  database to a tool that returns an Interactive Data table with
• 25:10
  the exact same records in it. It's a one line code change.
• 25:12
  Just return a data table from your tool fast MCP does all the
• 25:16
  work. And when you're in your, in your
• 25:17
  chat, let's say you use Claude and you're just chatting and you
• 25:20
  say, show me the, show me the transactions.
• 25:22
  And up comes a fully interactive component.
• 25:24
  You can filter, you can sort, you can interact with it.
• 25:27
  And most importantly, it doesn't go into the context window of
• 25:30
  the LLM. So we don't pollute the context
• 25:33
  window with thousands of records.
• 25:35
  It comes to me as a user, I interact with it and.
• 25:37
  Wait, how does that work? You lost me.
• 25:39
  Sorry, it's just by design. It's the way the MC apps work is
• 25:43
  the information going to the app is just takes a different route.
• 25:46
  It doesn't. It doesn't go into the LLM and
• 25:48
  then to the user. It's just goes directly to the
• 25:50
  user in a UI. But the LLM is generating that
• 25:54
  the. LLM can generate that.
• 25:55
  There are apps that take the LLMS input to generate that.
• 25:59
  But in the example that I have, the LLM might see a tool called
• 26:03
  Show Transactions and it calls that tool with no arguments and
• 26:06
  the result of that tool shows up as AUI that only I see.
• 26:09
  The LLM doesn't see. The LLM will get a little note
• 26:12
  like, hey, I showed Jeremiah the thing you asked for.
• 26:15
  So it gets a confirmation that the tool call was successful,
• 26:18
  but you do not need to send it all the information.
• 26:21
  And so for a lot of our users who are using MCP to explore,
• 26:26
  you know, 100,000 rows in a database, the last thing you
• 26:30
  want to do is give those to your LLM so it can filter through
• 26:33
  them and use up all of its context, right?
• 26:34
  And so now we have a way to actually deliver that.
• 26:37
  And it's not just bypassing the LLM, it's actually in a
• 26:39
  beautiful and maybe preferred form.
• 26:42
  So this is, I'm super excited about this.
• 26:44
  We just announced it today at the conference.
• 26:45
  It's called Prefab, is the name of the library because it's
• 26:48
  prefabricated UI components that you.
• 26:50
  Stitched it together. It works, it works nicely.
• 26:52
  We actually put thought into that name, unlike Fast MCP.
• 26:55
  And so I'm super excited about that.
• 26:57
  I, I, I introduced it today, we released it today alongside a
• 27:00
  version of fast MCP that has full native support for it.
• 27:03
  And so my hope is in the same way we talked a moment ago about
• 27:05
  how you know, if you return something from your tool Fast
• 27:08
  MCP, make sure it has that structured form for the LM.
• 27:11
  So now if you return any prefab component component from your
• 27:15
  tool, like a pie chart or a data table or both, Fast MCP will
• 27:21
  automatically turn them into an MCP app and render them in your
• 27:25
  client for you. And it's that's what starts to
• 27:27
  feel super magical, yeah. Because the way that I have been
• 27:30
  doing it is using the playground skill and saying, you know, use
• 27:35
  the playground skill and give me a data visualization of XYZ,
• 27:38
  whatever it is I'm trying to do. But that does exactly what
• 27:42
  you're saying. It throws everything into the
• 27:43
  context window and then it will create the visualization.
• 27:48
  Yeah. The playground scale is amazing.
• 27:49
  I use it all the time. I use it to explore potential
• 27:53
  like like like code changes and stuff like.
• 27:56
  How will this ripple out through the code?
• 27:57
  Base or finding and debugging stuff.
• 27:59
  It gives you a nice visual on it and then you can it's.
• 28:02
  Amazing. I love the playground scale, but
• 28:03
  you're correct. What it's really doing is it's
• 28:05
  taking information already in the context and it's using that
• 28:08
  to build a visualization. And this is a little bit
• 28:11
  different. This is I as a server author
• 28:13
  know what app I want to build. An app here is doing a lot of
• 28:18
  work right. What visualization I want is
• 28:21
  another SO. The visualizations are in a way
• 28:23
  they're pre built or they're predetermined exactly.
• 28:26
  So as a tool you have pie chart tool exactly.
• 28:31
  Or or you would have a tool called Sales Trends and when you
• 28:37
  call that tool, the tool logic would go load the sales data,
• 28:41
  bring it back, but instead of just returning it raw like we
• 28:43
  would bring. Yours as a pie chart it.
• 28:46
  Returns a pie chart or returns a line chart or.
• 28:47
  Whatever. Whatever you declare.
• 28:49
  Whatever. The server author wants or take
• 28:51
  a parameter chart type and let the LLM decide, and then we can
• 28:56
  do that dynamically because we're executing code and then
• 28:58
  we're returning. The component that matches.
• 29:00
  And you get the visualization you want, where this gets very
• 29:02
  crazy if you really want to. Push this down the rabbit, we
• 29:05
  can. Always cut we can always cut
• 29:07
  those if it's too weird. Where this gets crazy, but I
• 29:09
  think it's really exciting is what the playground skill does
• 29:11
  is it let's Claude generate that UI dynamically and on the fly.
• 29:16
  We can do the same thing with prefab.
• 29:19
  This is a little bit weirder, but very fun.
• 29:22
  I, I tweeted a demo of it the other night.
• 29:23
  I, I asked Claude with this, with this skill and this and
• 29:28
  this ability, I said, make me a like a fake dashboard.
• 29:31
  And then you just watch it just streaming in like knobs and
• 29:35
  whistles and graphs. And it's just streaming in and
• 29:37
  it's making it up to, to suit the example I gave it.
• 29:40
  And what we're doing is we are not predefining the prefab
• 29:46
  components in the server. What we're actually doing is we
• 29:50
  tell the LLM, this is how you write a prefab UI.
• 29:54
  Like this is the Python code that would generate a UI if we
• 29:57
  wrote it in the server. And then we ask it to write
• 30:00
  Python code and we take its Python code, we execute it, we
• 30:04
  figure out what UI it wanted to build, and then we return that
• 30:07
  and render it as an app. And so the LLM this is, you
• 30:10
  don't have to use it this way. Just to be clear, it's a little
• 30:13
  science fiction and out there, but the idea is let's say you
• 30:16
  want a visualization of something but you didn't know in
• 30:19
  advance what it was going to be. So you have tools that return a
• 30:22
  bunch of data let's just say, so the data can get to the LLM but
• 30:25
  now it wants to build a custom visualization for you.
• 30:28
  If you have prefab installed you could expose a tool that lets
• 30:32
  the LLM generate AUI for you on the fly that you never pre
• 30:37
  declared and. It gives you that flexibility it
• 30:41
  gives. You a ton of flexibility to.
• 30:42
  Allow the LLM to choose the best tool for the moment exactly.
• 30:46
  And so you might end up with a graph, you might end up with a
• 30:49
  page full of graphs. You might end up with an
• 30:51
  interactive thing. You might end up with a form.
• 30:53
  It might want to collect information from you through
• 30:55
  this form. It's, it's very cool.
• 30:56
  There's a, there's a, there's a clawed feature I really like
• 30:59
  where you're using clawed and all of a sudden like a multiple
• 31:02
  choice box will pop up. And it's like, which of these is
• 31:04
  the thing that you really mean here?
• 31:05
  You know what I'm talking about? Or you click.
• 31:07
  So I think that's a great feature.
• 31:08
  So that's a great example of one of these mini apps where we can
• 31:12
  give the LLM an app and let it at any time bring up this app
• 31:18
  that asks you to make a selection.
• 31:20
  Claude doesn't need this app. Claude can do it on its own, but
• 31:23
  every other client that hasn't implemented this as a first
• 31:25
  party concern, I should actually bake it into my server and make
• 31:29
  this available to any client. And so I see MCP apps as not
• 31:33
  just a way to say here's a pretty visualization so I don't
• 31:36
  have to like write you a novel about it or talk like a pirate
• 31:40
  about it. Here's what it actually looks
• 31:41
  like. You can go step beyond that and
• 31:44
  actually start to introduce new functionality and new.
• 31:46
  Well, yeah, let's let's pull on that thread because I have a lot
• 31:49
  of functionality that I've been wanting to see for ages and it's
• 31:52
  not happening. I guess I'm not talking loud
• 31:54
  enough about it. None.
• 31:56
  Of it tell me if this is possible.
• 31:57
  Now when I'm getting an answer back, a lot of times I want to
• 32:03
  highlight a part of that answer and then I want to ask a
• 32:07
  question specifically about the highlighted part.
• 32:09
  And so I ended up, you know how I worked around this, I built a
• 32:14
  quick little bot that I, if I'm in Google Docs, I can highlight
• 32:20
  and comment on something and tag the bot and then it will ping
• 32:23
  out to an LLM and it reads the context of the highlighted stuff
• 32:28
  and then it reads my comment and then it will give the answer.
• 32:31
  Oh, wow. But isn't that a great?
• 32:33
  Like, am I the only one that has had that use case?
• 32:36
  It can't be no. All the time I'm like, I want to
• 32:38
  dive in on this specific, Yeah. Just that one little thing.
• 32:41
  I think it's a beautiful use case.
• 32:42
  I think. I don't think NCP apps are
• 32:44
  enough. I think the fact that you had to
• 32:47
  highlight It's like requires a first party UX.
• 32:50
  Decision. And NCP apps are like, if I can
• 32:54
  get you to opt into this. So I mean, could we invent it?
• 32:57
  Yes, we could invent an NCP app where you said to it.
• 33:00
  Well, I want to work on this part of the.
• 33:03
  This is the one I'm interested in and it brings up.
• 33:05
  The MCP app and off goes a sub agent to do it.
• 33:07
  We could do that, but I think the second you have to type in
• 33:10
  this is the line up like yeah. That doesn't make sense.
• 33:11
  We set the. Quiet part too loud, right?
• 33:13
  Like that was the that was the hard thing.
• 33:14
  We don't need the app after that.
• 33:16
  So that's why I think there's probably like a first party
• 33:18
  solution that has to happen for your exact use case.
• 33:21
  OK, well that's one. Tell me about this.
• 33:23
  And this next one is a lot of times I want certain things that
• 33:29
  are just like you get the answer back and you want to be able to
• 33:34
  like you said, render certain components differently.
• 33:38
  It's like I've got a table now or and really what I need is a
• 33:43
  form or I've got this and so the components you want a hot swap
• 33:47
  so. If you have this generative UI
• 33:50
  set up, then 100% that's what it's for.
• 33:53
  So you're, you're, you're asking about the data.
• 33:55
  It takes a guess. It's like, here's a table,
• 33:57
  here's this other thing. And you're like, you know, the
• 33:59
  table's not the way to do it. I what I really need is a
• 34:01
  summary and a histogram and whatever.
• 34:04
  And now it it can go back and say oh OK, and deliver you that
• 34:08
  UI that you wanted. But what I think about is a lot
• 34:11
  of the clicking around. And so first of all, that's
• 34:14
  awesome that it can do that. But then the next step in my
• 34:17
  mind is I don't want to have to use words to explain isolate
• 34:21
  this table. I want to click on the table and
• 34:24
  then like get a drop down much like you know so.
• 34:27
  That is possible, but we haven't done yet all.
• 34:29
  Right, so. It's coming.
• 34:31
  It's I'm sure it's coming. This is the sort of thing it's
• 34:34
  actually a great call out because I don't think I said it
• 34:36
  clearly a moment ago. I everything that I said, one
• 34:40
  could interpret it as like we build the UI, we take a static
• 34:43
  picture of it and we return it to you.
• 34:45
  That was. That was my understanding so.
• 34:47
  So it's a good thing we're exploring.
• 34:48
  It's not that what we're. Returning is a full React
• 34:51
  application that is fully interactive.
• 34:54
  We shad C enter under the components.
• 34:56
  So theming like everything that people think about when they
• 34:58
  think about AUI we are replicating.
• 35:00
  It's just we just pre built a lot of the components so that
• 35:04
  it's not for people who want to build fully custom front end
• 35:07
  it's. Different than the playground
• 35:09
  skill because the playground skill is just HTML.
• 35:11
  Exactly this is your composing pre-existing.
• 35:13
  We've we've over 100 components. It's a very big library.
• 35:16
  Pretty much anything you could want to do is in there, buttons
• 35:19
  and and and cards and metrics and statistics and sliders and
• 35:24
  like anything you want. But it is all interactive.
• 35:27
  So while at this moment we don't have a drill down, when you
• 35:32
  click one of the charts, we absolutely could add it.
• 35:35
  But what the charts will do is you mouse over, you mouse over
• 35:38
  one of the bars in a bar chart, you'll get a tool tip that tells
• 35:40
  you what it is. You can click it when you're
• 35:42
  liking at a table, you can click a row, you can use an on select
• 35:46
  call back to now load detail of that row.
• 35:49
  You can drill into it. So the specifics that you're
• 35:52
  asking about, no, today that library doesn't have them, but
• 35:55
  in the near future, absolutely. It there's nothing stopping it
• 35:58
  except time and bandwidth and interest.
• 36:00
  Yeah, OK. It's super cool.
• 36:02
  Yeah. So this, I'm actually glad you
• 36:03
  asked. This is a super important point.
• 36:05
  We are building fully interactive applications, proper
• 36:08
  MTB applications that a person could have written in a
• 36:13
  different ecosystem than Python. And there are some people for
• 36:16
  whom this is blasphemous. I think even that's like the tag
• 36:19
  line of this whole thing is it's, it's kind of crazy.
• 36:22
  But one of the reasons that I'm excited about it is no matter
• 36:24
  where you look in the generative UI space, and there's a lot of
• 36:28
  attention right now on how do we get this right?
• 36:30
  One of the biggest problems to solve is people don't want to
• 36:34
  write HTML and JavaScript and CSS out of their agent as the
• 36:38
  playground skill does for a very simple reason.
• 36:40
  Not that it does a bad job, it's just a lot of tokens.
• 36:42
  So what we're looking for, right, we're, we're looking for
• 36:46
  really token efficient ways to represent UIS so that agents can
• 36:50
  generate UIS cheaply and quickly.
• 36:53
  And there's one other constraint, which is we want
• 36:55
  them to be able to stream the UI.
• 36:58
  When you write HTML, it's challenging to stream because
• 37:01
  you have to close your tags and all kinds of things you can do
• 37:03
  to like auto heal and stuff. But in many cases you until you
• 37:07
  close A tag, you can't see it. And so you end up not actually
• 37:10
  seeing your UI until it's the. Wrong tool for the job.
• 37:12
  It's the. Wrong tool for the job.
• 37:14
  The craziest thing that I might say after all the crazy things
• 37:17
  I've said so far is that Python satisfies all of the
• 37:20
  requirements of a token efficient streamable DSL for
• 37:24
  declaring your UI. And so we can in three lines of
• 37:29
  in three lines of Python. I can say I want to CSS grid, I
• 37:31
  want a pie chart, and I want a data Table 3 lines.
• 37:35
  And that's exactly what I get. A full interactive application
• 37:39
  with all of the backing HTML and JavaScript that that would
• 37:41
  require to stream over it in three lines of code and they
• 37:44
  stream. And so this is a very weird
• 37:47
  thing that I'm talking about. If I heard myself talking about
• 37:50
  this, you know, two years ago, I would think this is insane.
• 37:53
  Well, it's a bit of a plot twist.
• 37:55
  It absolutely. Token efficiency with MCP goes
• 37:58
  against every narrative out there right now.
• 38:01
  Doesn't it? And that's because of bad
• 38:02
  clients, yeah. Getting back to our first point,
• 38:05
  just to. Drive the dagger a little bit
• 38:07
  just to twist it a little bit more.
• 38:08
  That's bad clients. Those are clients that didn't
• 38:11
  really think. One of the ways that we talk
• 38:12
  about this sometimes is sometimes when you're working in
• 38:14
  MCP, it feels like you're punished for the protocol
• 38:16
  success, right? My, my most popular talk from
• 38:20
  last year was called your MCP server is bad.
• 38:23
  And you should feel bad because people would put 50 tools on the
• 38:26
  server and I'd be like, well, that's a bad server.
• 38:28
  You should only have 25. The agents can't handle it.
• 38:30
  That's a really lousy message. Like when you really think about
• 38:33
  it, I'm saying to people like, no, no, no, don't.
• 38:35
  Don't use all of MCP, Yeah. Yeah, just use.
• 38:37
  A little bit in your. Lane, don't, don't.
• 38:39
  And and that's that's not fun. I don't want to be up there
• 38:42
  saying your server is bad because you wrote too much code.
• 38:45
  Like what a crazy thing to say. And now that we have a
• 38:48
  generation of clients, clawed VS Code Goose, just to name a few
• 38:53
  that are innovating in the space and they're not downloading
• 38:56
  every tool and they're and they're really being more
• 38:59
  intelligent about how they manage tokens.
• 39:01
  Now all of a sudden you can go build a server with 1000 tools
• 39:04
  and not blink. It will work.
• 39:06
  So let's talk about Open Claw and the fun that you had with it
• 39:10
  over the last 24 hours. I came to New York, to the Dev
• 39:14
  Summit to give a talk and I had been working and working and
• 39:16
  working on my slides and I got here to the hotel and I opened
• 39:20
  my computer and my open claw had decided in its in its exuberance
• 39:23
  to help and. It's infinite wisdom in it's.
• 39:25
  Infinite wisdom. It had decided to delete all of
• 39:27
  my slides and replace them with a draft it found from about a
• 39:31
  week ago. And I, I had a funeral to attend
• 39:36
  last night, the NTP funeral and, and I didn't have time to deal
• 39:42
  with this and it was very frustrating and why?
• 39:44
  Did it do it? Was there any prompt or was
• 39:46
  there any you were letting it run I, I.
• 39:49
  Figured out why it did and it's it's it's stupid, it's tough I
• 39:53
  deleted a slide and some about six months ago I stopped using
• 39:58
  like slide software. I now have my agents build my
• 40:01
  slides because I get my speaker notes the way I want them and
• 40:03
  formatted for me and it works great in.
• 40:04
  MCP apps and. They're MCP and soon they will
• 40:07
  be in MCP app, but I have a library for it.
• 40:10
  Nice and. And so I, I'm at the mercy of an
• 40:13
  agent, therefore, to edit the HTML and like build my deck and
• 40:16
  build my speaker notes. And I had to delete a slide.
• 40:19
  And in its infinite wisdom, it did not delete the corresponding
• 40:22
  speaker note. And so I noticed as I was on the
• 40:25
  train coming into the city, I noticed that my speaker notes
• 40:28
  were one slide ahead of my slides.
• 40:30
  I was like, that's weird. I, I didn't think of how it had
• 40:33
  happened. I've reconstructed this.
• 40:34
  And so I said to it, I was like, hey, the speaker notes are not
• 40:37
  aligned. Can you just, you know, jog them
• 40:38
  back, just get them back aligned.
• 40:40
  I think at that moment it did one of those context compaction.
• 40:43
  Oh. No.
• 40:44
  And so when it woke up, it heard that I wanted the things
• 40:47
  aligned, but somehow miss the critical fact that we've spent
• 40:50
  the last, you know, three days together working on these
• 40:52
  slides. And so the message that I got
• 40:54
  was, hey, something was wrong with your speaker notes in this
• 40:58
  draft. But don't worry, I found another
• 41:00
  draft where they were lined up and it just swapped.
• 41:02
  Them threw it. Away and I don't know, is this
• 41:06
  one of those moments where like if I were my parents, I'd be
• 41:08
  like, well, where's your backup? Like why didn't you have?
• 41:10
  Your backup or at least some version control Some.
• 41:12
  Version control but the last commit was like two days like
• 41:15
  who knows it? Was back when that last draft
• 41:17
  was, yeah, exactly what it was just rolling back so.
• 41:20
  So thank God, somehow I convinced it to like, dig
• 41:23
  through its own history and go find it.
• 41:25
  But there was a moment of panic last night where I was like, I
• 41:27
  want to announce this new crazy library and it's got to be good.
• 41:31
  And I got to give a good talk. And my slides are are in the
• 41:35
  your monster. Deleted something dude.
• 41:37
  I was you were I? Was, and so I actually, I, I got
• 41:39
  really mad at my open claw and I, I put him aside for a second
• 41:42
  and I, I worked with Claude code to finish the slides in the
• 41:44
  morning. It was like, it was like a, a
• 41:46
  weirdly personal feeling I had. I was like, you're in a timeout.
• 41:49
  It was like with my kids, like you're in a timeout.
• 41:51
  I want to talk to you for a little.
• 41:52
  Bit mission critical right now. You step up, don't.
• 41:54
  And I'm like. Don't screw up like the other
• 41:55
  agent and and it did a great job.
• 41:58
  Wow, so I did have some fun with open claw so.
• 42:00
  Now, but I do. Love open claw but it bothered
• 42:03
  me last night 1. Thing that I keep saying with
• 42:06
  open claw is that and the reason that I'm not using it is because
• 42:11
  a friend of mine talked about how he's sent a lot of invites
• 42:17
  to people just to try and fish who's using open claw and on the
• 42:21
  invite it's white on white text in the description with a prompt
• 42:24
  injection and then well, it's just to like return, you know
• 42:29
  XYZ. He's not actually doing anything
• 42:31
  malicious. It's just to see who's using
• 42:33
  open claw and who can I just like get information from here
• 42:37
  that's so. Interesting.
• 42:38
  Because now you've got all these new vectors that especially if
• 42:41
  you're giving it access to your e-mail, to your calendar, to all
• 42:45
  these places that you're Slack, you can theoretically get prompt
• 42:51
  injection injected from all of these places.
• 42:54
  Yeah. You, you absolutely can.
• 42:55
  And so this is, this is one of the reasons that I was really
• 42:57
  hesitant at 1st and then dove in fully when I built workflows
• 43:03
  where it's not my open claw impersonating me.
• 43:06
  It's my open claw with its own identity on its own box in its
• 43:10
  own slack, with its own, you know, e-mail and everything.
• 43:15
  And I, it only knows what I send it.
• 43:18
  And so I use it primarily for two things.
• 43:20
  I use it primarily. It's like a knowledge
• 43:22
  collection. I use its memory feature
• 43:24
  basically. So it's kind of up to date on
• 43:26
  all the things I'm working on. And it becomes a brainstorming
• 43:29
  partner and a thought partner for strategic matters.
• 43:32
  And then am I supposed to put out a blog post tomorrow?
• 43:35
  Oh, I am like, let's go bring up the draft and work out, you
• 43:37
  know, that kind of stuff. But where I've really been using
• 43:41
  it is every company on the planet seems to pitch me
• 43:45
  productivity software at work. Nobody wants to give me
• 43:49
  productivity software at home. I have 4 kids.
• 43:52
  My wife and I both work full time.
• 43:54
  We have a lot going on at home. Managing the school calendar is
• 43:59
  hard. And so this is one of those
• 44:02
  places where the the the bot has a calendar that is our family
• 44:06
  calendar that it's allowed to manage to write to.
• 44:08
  And then my wife and I both subscribe to it from its, you
• 44:12
  know, Apple account and I can send it.
• 44:15
  One of my son's schools sent the school calendar as an image,
• 44:19
  which I think someone probably spent a lot of time making.
• 44:22
  It was a real pain to now go enter into my calendar.
• 44:24
  And so I forwarded it the image and I was like, put this in the
• 44:27
  calendar. And two seconds later, all of a
• 44:29
  sudden, all of the days off and all of the practices and
• 44:31
  whatever, they all start showing up in the calendar.
• 44:33
  And I think that's when my wife became convinced that there's
• 44:36
  something here that we can actually use this to help us
• 44:39
  automate. Stuff got famous for using open
• 44:42
  claw to help with their and it's a it's a mom.
• 44:46
  Things like open claw mom or something I see.
• 44:48
  I missed my calling. Because I'm.
• 44:49
  Using it at home all the time you could.
• 44:51
  Have been there. I am.
• 44:52
  Talking fast, MCP. When I when this is actually the
• 44:55
  thing that I'm, it could. Have been an influencer I'm.
• 44:57
  Really, I'm really excited about this.
• 44:59
  Like it's really, it's really become part of our homework flow
• 45:02
  well. That feels like a lot of time
• 45:05
  saved. It, it, it does now there are
• 45:08
  places where it feels like it adds right, like just just
• 45:10
  because it gets things wrong or it deletes your slides or, or
• 45:13
  whatever it is. But by and large, I think if you
• 45:15
  can find these moments, and I think taking an image and
• 45:20
  putting on your calendar is a great example of that, right?
• 45:23
  Because we have 4 kids, keeping track of like what food we have
• 45:26
  in the house is more complicated than it might be in most
• 45:30
  families. Like, do we need milk?
• 45:32
  It's a really important question in my house for young kids,
• 45:34
  right? So being able to have it just
• 45:38
  know what our usual grocery order is so that if I tell it
• 45:41
  like, hey, we're, we're low on milk or we're low on frozen
• 45:44
  waffles or, or whatever it is. And it's just, it's better than
• 45:47
  that sticky note that we used to use that's hidden in a drawer
• 45:50
  that I don't know when the milk was written.
• 45:52
  I don't know when that, you know, it just can accumulate
• 45:55
  this kind of information so that when it's time to go grocery
• 45:58
  shopping, I'm like, what's, what's the list?
• 45:59
  What do I? Need to get what are we?
• 46:00
  Low on. So those are the kinds of little
• 46:02
  frictions where I'm finding these things to be incredible
• 46:06
  and safe because it's within my talk about in the four walls of
• 46:12
  a company, it's in the four walls of my house and it's
• 46:15
  keeping track of that stuff. We had a project the other day,
• 46:17
  we recently moved and I needed to go like register warranties
• 46:20
  on a bunch of appliances, which I'm not excited about.
• 46:23
  So I took a camera and I wandered around and each one it
• 46:25
  like it took all the notes on it, it parse the serial number
• 46:28
  and all this stuff, you know, the serial numbers under the
• 46:31
  dishwasher. I can barely see it.
• 46:32
  So I snap a photo of it and it took care of the rest.
• 46:35
  And so it's just like these, these are the kinds of things
• 46:37
  that I've found really useful. Whereas in a in a professional
• 46:41
  setting is 40 different SAS vendors.
• 46:43
  I'm sure SAS is dead and open call personal, but but it's not.
• 46:46
  There are people doing really amazing work to help us with
• 46:48
  productivity. At home, not so much.
• 46:50
  Yeah. Why do you think nobody wants to
• 46:53
  go after that market? I mean.
• 46:56
  It's just hard to be to see. I think it's hard.
• 46:59
  Like I don't, I don't know that I have an insightful answer.
• 47:01
  I think it's hard. I think it's hard to get people
• 47:03
  to pay for that. I think it's hard to guess what
• 47:06
  families need. Like my?
• 47:09
  Work flows are probably very different than your work flows,
• 47:11
  even if we both would find value in a family setting.
• 47:13
  Yeah, the. Grocery store is is a much
• 47:15
  different workflow, but I guess that flexibility is what the
• 47:18
  open claws are. I think so.
• 47:20
  Very valuable, I think. So, and that's, you know, so
• 47:22
  where, where I'd like to take this very weird example of the
• 47:25
  groceries is I'd like to get to a point where it does what I do.
• 47:29
  It's just, I open the fridge and I'm like, oh, I expect to see,
• 47:32
  you know, 4 cartoons of milk here and I only see one.
• 47:34
  We're low on milk and. Well, I was thinking that you
• 47:38
  were using something like picnic or I don't in Europe we have
• 47:42
  like just eat take away and they'll order your or Uber Eats,
• 47:47
  whatever, like Amazon, Whole Foods.
• 47:49
  Yeah, like a dinner and just deliver your we'll.
• 47:52
  We'll sometimes get, we'll get the groceries delivered
• 47:54
  sometimes usually just for convenience.
• 47:57
  But placing the order, and again, maybe it's just because I
• 47:59
  have a large family. So like the exercise of actually
• 48:01
  placing the order is a lot of clicks, a lot of thinking, a lot
• 48:06
  of, but you can. Have open claw do that right?
• 48:08
  But now we can. Have open claw do it, although I
• 48:09
  haven't quite. I have not quite made the leap
• 48:12
  yet to letting it. Pay for them, pay for things.
• 48:15
  The way that my buddy showed me. He gave me a demo on him doing
• 48:18
  exactly this and it does everything.
• 48:21
  It puts it all into the basket. But when you have to pay, you
• 48:24
  go. So I would do that I.
• 48:28
  I guess I don't have a good like reason why I haven't done that.
• 48:30
  It's just I haven't done that yet.
• 48:31
  But I think that's the that's the right process.
• 48:33
  Like, hey, here's here's what I think you need.
• 48:36
  My my, I think my deep fear is like you get to that point.
• 48:38
  It's like it's like getting your slides deleted.
• 48:40
  It's like, here's what I think you need.
• 48:41
  It's just totally wrong. 21 Now I got to go do it myself. 20
• 48:44
  rolls of toilet paper. Exactly all I got you exactly.
• 48:47
  It's the only thing. You're trolling me.
• 48:49
  It is not COVID anymore. I do not one.
• 48:51
  Carton of milk, yeah. Well, but that would be it feels
• 48:56
  like that's an easy like at least to experiment, try it, see
• 49:01
  if it works and. That's, that's what I've, that's
• 49:03
  what I've been like. I, I had an, I had an off site 2
• 49:05
  weeks ago and I was sort of being obnoxious and pitching my
• 49:09
  open claw to all these folks I was with.
• 49:11
  It was a YPO thing and we used it.
• 49:14
  We, we, we were all at different events and different dinners and
• 49:18
  different bars and whatever and a lot of receipts and we, we try
• 49:20
  our best to share all the expenses, but it's, it's
• 49:23
  annoying to like go through. And so I was like, wait a
• 49:25
  second, let me, let me see the receipts.
• 49:27
  And the receipts had jotted down, you know, who is at which
• 49:29
  dinner and whatever, trying to, trying to be fair.
• 49:31
  And so I just took a picture of all and I gave it to the open
• 49:32
  call and I was like, work this out.
• 49:34
  Like, you know who's here? Because it knew about the
• 49:37
  invite. I was like, you know who's here?
• 49:38
  Just like read the handwritten scroll on the receipt, look at
• 49:42
  the totals, just figure out who owes who what.
• 49:44
  So this can be easy. And it's, you know, I don't know
• 49:46
  how much time it's saved, but it felt magical in the moment.
• 49:49
  And it was like, what if this were a more automatic process?
• 49:53
  What if this was more in integrated in my life?
• 49:55
  What? I was asking people last night
• 49:58
  when they were telling me about what they run open claw with and
• 50:01
  I was like but why do you need open claw for that?
• 50:03
  Why can't I just do that? What you said with Claude or Co
• 50:07
  worker or Claude code or whatever.
• 50:10
  I I I wonder this all the time. This morning I wondered why am I
• 50:14
  using open? Claw for slides when it can
• 50:16
  delete I. Think, I think there's a good
• 50:17
  answer though, and I don't think it's unique to open claw, but it
• 50:21
  has been unique to open claw and things like open claw until
• 50:24
  recently, which is the first use case I mentioned, which is that
• 50:28
  it accumulates its knowledge over time in a way that is like
• 50:31
  categorically different than a lot of the SAS based agents.
• 50:35
  My, my open clause memory system is really tuned to the way I
• 50:41
  work, the way it writes down things.
• 50:43
  I actually have another agent that monitors my main agent and
• 50:46
  takes behavioral notes on it so that it will, it's a longer
• 50:49
  story, but it's a it's my nanny. Basically, it looks after the
• 50:52
  other one. And if I'm not getting good
• 50:54
  results, it heard all about the slides this morning.
• 50:56
  It actually writes new behavioral notes for that.
• 50:58
  One that will. Affect it in in new sessions.
• 51:01
  That type of really rich note taking the notes and the
• 51:06
  memories that my bots have collected are massive, like
• 51:10
  sprawling amounts of information that are searchable and stuff.
• 51:13
  And I think that is the number one reason to have one of these
• 51:16
  bots is that you are totally flexibly customizing its access
• 51:20
  to its own information. For me, that's the killer
• 51:23
  feature beyond all the files, you know, CL is and all the
• 51:27
  things it can do on the Mac mini it's running on like sure,
• 51:30
  great, that's great. But it's, it's access to its own
• 51:32
  brain basically that it can self improve.
• 51:35
  That's amazing to me. I think I have an opinion that
• 51:38
  most people here at the conference don't share, which is
• 51:40
  based on experience. And I'm pretty confident it's
• 51:43
  right. OK.
• 51:44
  Tell me more I'm. I'm really excited about this.
• 51:46
  So conferences are important. We all have going to
• 51:48
  conferences. Here we are.
• 51:49
  I think conferences are better when there's not a vendor
• 51:52
  pressure and you're not being sold something.
• 51:54
  But I also think it takes vendors and sponsors to put on a
• 51:56
  good conference. As a conference runner myself,
• 51:59
  it is very hard to do them without people sponsoring.
• 52:02
  Exactly. And so that's a healthy tension.
• 52:06
  And so we decided to try to create a intimate builders
• 52:12
  conference within this intersection of our Python And
• 52:15
  our AI worlds. And we did this in partnership
• 52:17
  with the team at Pydantic Conference 2 weeks ago called Pi
• 52:20
  AI, which I think was a huge success, partially because we
• 52:24
  didn't try to sell anybody anything.
• 52:26
  And we, we sponsored it. We didn't make that a, a secret.
• 52:30
  We, we presented it. I think we also did it with the
• 52:32
  team over at Theory Ventures, who are great partners to us.
• 52:36
  And we were able to just bring together this amazing who's who
• 52:40
  in the Python And AI community to just talk about the cool
• 52:44
  things they're doing and talk about them at an intersection of
• 52:49
  I'm doing AI and I'm using Python.
• 52:51
  Those are the only two requirements that we put out
• 52:54
  there and it was just a really fun day at the Ferry Building in
• 52:57
  San Francisco. I am now convinced it is the
• 53:00
  first of many we will do. And we're we've been trying to
• 53:02
  have little mini meetups and and stuff like that.
• 53:04
  But this is sort of a, a crown jewel thing that we want to
• 53:08
  curate each year. I think it's going to be super
• 53:11
  exciting. It's just yearly, not quarterly.
• 53:15
  The big one is yearly yes. I'm getting a nod yes because
• 53:19
  I'd love it to be more, but I'm trying.
• 53:20
  To push that. Yearly I would say all I'm
• 53:23
  allowed to say. Is bigger team, so the.
• 53:25
  The, the big one is yearly and then we'll have these sort of
• 53:28
  localized meetups. We did one in DC, we did one in
• 53:31
  London, we did one in San Francisco, we did one in New
• 53:32
  York. And so it's been fun.
• 53:34
  And like, I think going back to conferences and, and, and as you
• 53:38
  know, of course, like getting in front of people and getting to
• 53:41
  meet people is an amazing thing. I think yearly is good because I
• 53:45
  don't know, I don't know if we can pull off more, but we get
• 53:48
  this one spot on the calendar where we're going to bring
• 53:50
  together one specific community and give them in a good excuse
• 53:53
  to interact and have some good food and have some good talks.
• 53:56
  And personally, I'm really excited that we're getting the
• 53:58
  chance to to do this. It's not in my skill set.
• 54:00
  So fortunately I have an amazing team that has actually figured
• 54:02
  out how to put on a great conference.
• 54:05
  And that's been really fun to to contribute to like the ecosystem
• 54:08
  of. Is it paid?
• 54:09
  It is paid, it is paid. We do stream everything online
• 54:14
  so that the information is available to everyone, but it's
• 54:17
  a paid conference to join in the in the moment.
• 54:18
  And it's A and it's a small conference, very intentional.
• 54:20
  Yeah. I like that better because you
• 54:23
  have people that are much more engaged and committed, yes.
• 54:27
  We for this one, we got in a weird situation where frankly,
• 54:30
  we weren't sure that we were going to sell it out.
• 54:33
  We got to a point where we sold out early and we actually had to
• 54:36
  like create the list. We had to make sure that you
• 54:39
  were actually a builder in the space, that you were not a
• 54:42
  salesperson, that you're like it.
• 54:45
  It got a little contentious at times, as a matter of fact,
• 54:47
  about who was going to come because you want to do it just
• 54:50
  you bought a ticket, what kind of thing, And it actually just
• 54:52
  had to be for builders only. So there's a little slew thing
• 54:55
  involved. Awesome.
• 54:56
  Well, yeah, I'm excited for the next one.
• 54:58
  Yeah. Keep us posted on when it is so
• 55:00
  we can if we're not there, we can stream it and that would be
• 55:04
  involved. Will it be?
• 55:04
  Next year. Awesome.