Full transcript
0:01[Music]
0:15right? Hello everyone. Who's excited to
0:17chat about MCP today?
0:20Okay, we can we can work on that. We can
0:21get it a little bit better by the end of
0:23this talk. Uh but I'm Theo. I am a
0:26product manager at Anthropic work on
0:28MCP. Uh prior to this was also a startup
0:32founder uh working in the AI space. Um
0:35couple fun facts about me because
0:36everyone says make yourself a little bit
0:38more personable. Uh is that I like
0:41playing poker mostly losing money at
0:43poker, not uh making money at poker. Uh
0:46and I also really like coffee. So, uh,
0:49if you're, you know, a huge coffee fan,
0:51um, and want to talk about the best
0:54coffee in San Francisco, hit me up after
0:55the talk. But you didn't come here to
0:58talk about me. You came here to learn
1:00about MCP. So, let's talk about MCP.
1:05I was told not to say MCP is the best
1:08thing since sliced bread. Uh, which I
1:10won't say, but mostly because I don't
1:13actually think it's the best thing since
1:14sliced bread. Uh my goal here today is
1:17to really walk you through the origin
1:19story of MCP, why we launched it, uh
1:22give you a better sense of, you know,
1:25where it can actually help you in your
1:27workflow. Uh and then ultimately give
1:30you a sense of the types of questions
1:32that I'm frequently hearing, where I
1:35think there's a lot of value to build in
1:36the ecosystem, and let you decide for
1:39yourself whether or not it is actually
1:40the best thing since sliced bread.
1:44So, scrolling all the way back to uh
1:48midl last year, the co-creators of MCP,
1:51David and Justin, had this idea. Uh they
1:55were seeing that, you know, classic two
1:57engineers in a garage style. They were
1:59seeing that they were constantly copying
2:02and pasting context from outside of the
2:05context window into the context window.
2:07So, you're doing your workflow and
2:09suddenly you're remembering that there
2:11was a Slack message. which was really
2:12important that had a lot of context that
2:14you could just copy in. Um, so you're
2:16constantly kind of copying things back
2:17and forth from Slack. Maybe you're
2:19copying things in from Sentry, your
2:22error logs. Uh, but they were kind of
2:24realizing, hey, it would be so great if
2:27Claude or any LLM could just kind of
2:31climb out of its box, reach out into the
2:33real world and bring that context and
2:35those actions uh to the model. And so
2:38the genesis of MCP was really around
2:42this big question of uh not just context
2:45but model agency. How do you actually
2:48give the model the ability to interact
2:51with the outside world?
2:53And so as they started thinking about
2:56this uh they came to the conclusion that
2:59it had to be an open-source standardized
3:02protocol in order for this to make sense
3:05uh at scale. And the reason is of course
3:07as you all know if you want to build an
3:09integration uh and the you know the the
3:13actor uh or the client in this case that
3:16has to uh leverage that integration is a
3:19is using a closed source ecosystem then
3:21you need maybe a BD or partnerships uh
3:25angle with that client to actually get
3:27access to the team to integrate with
3:29them. You then have to align on the
3:30right interface and then you get to
3:32actually build the thing itself. Um and
3:35so the idea here was that model agency
3:38was the biggest thing that was stopping
3:41uh LLMs from actually reaching the next
3:43stage of usefulness and intelligence. As
3:46we saw that reasoning models were
3:48becoming uh more and more the future
3:52that tool calling was getting better. We
3:54really wanted to make sure that we were
3:56making it possible for everyone to get
3:58involved in that ecosystem and actually
3:59allow uh the models to again have
4:03agency.
4:04Uh so they form a small tiger team
4:07internally uh work on this protocol and
4:12launch it at our company hack week in uh
4:14November of last year. And this was
4:16really the first turning point of MCP.
4:18It went viral as you can imagine.
4:20Engineers from various teams were
4:24working on building MCPs to automate
4:26their own workflows. They were working
4:27on MCPs to uh automate other teams
4:30workflows. Uh this was really kind of a
4:32cool moment to see how it went from
4:35again like two engineers in a garage all
4:37the way to uh this is a major moment in
4:40turning point where we think we actually
4:42unlocked some uh true value for for
4:44other people. And so we ultimately ended
4:47up open sourcing uh MCP in November of
4:50last year and that's when uh we
4:52introduced it to the rest of the world.
4:55But as most builders know, uh when you
4:58build something 0ero to one, you think
5:00the launch moment is going to be really
5:02impactful. But it actually usually is
5:04not. Uh at launch, most people were
5:08saying things like, "What's MCP?" or
5:11even worse, or maybe, you know,
5:13rightfully so, what's MPC? Uh and more
5:17often than not, we got this question of
5:19I don't really understand why you need a
5:21new protocol. I don't really understand
5:23why it has to be open source. Camp
5:26models call tools already. Uh this was
5:29the slew of questions that kind of came
5:32uh again and again for probably from the
5:34era of November all the way even to uh
5:38early uh early this year. And it really
5:42took uh making it possible for builders
5:47to kind of get their hands dirty uh with
5:49building MCPS to automate their own
5:51workflow for for uh for this to take
5:54off. And so the next turning point uh as
5:57Henry alluded to was when Cursor kind of
5:59adopted MCP and after that a lot of
6:02other coding tools also adopted MCP um
6:05VS Code uh source graph uh etc. But we
6:10had a lot of coding ideides um start
6:12adopting MCP and that's really where
6:14that uh next stage of momentum came in
6:18where agent uh agency was given to
6:20builders to actually build uh MCPS for
6:22themselves
6:24and more recently we've seen uh kind of
6:27another turning point where Google,
6:29Microsoft, OpenAI uh and many others
6:31have uh also adopted MCP. So really
6:34excited to see this kind of become more
6:36and more uh the standard. But ultimately
6:39uh standards uh become standards because
6:42they are actually useful to builders.
6:44And so uh I uh kind of want to ask all
6:48of you to to keep us honest. Um
6:50contribute when you see you know issues
6:53with uh the way that the the protocol is
6:55built today. uh or uh if you uh even
6:59want to take that one step further and
7:01submit a PR directly to the GitHub repo
7:03and uh fix the issue that'd be even
7:05better. Um but our goal here is really
7:07to make it maximally useful for uh for
7:10you all and for uh model providers. So
7:13uh thank you for for your help in even
7:15getting us to the point where I can be
7:17speaking on stage uh about this uh less
7:20than one year later.
7:23So just to get a little bit deeper into
7:26uh what we were solving for at the start
7:28of building MCP is again this kind of
7:32idea of of model agency. Um and part of
7:36that means uh agents is kind of the
7:38direction that that we think is is going
7:40to be the future. That's no surprise to
7:42anyone in this room. You are probably
7:44going to hear the word agents said in
7:47every talk if not almost every talk. Uh
7:50but the way that we think about agents
7:54is that you are giving the model or
7:57you're rather depending on the model's
7:59intelligence to choose actions and
8:01decide uh what to do. Uh in the same way
8:04that you know maybe when you talk to a
8:07human and you ask them uh for a
8:09response, you don't know exactly what
8:10the response is, but based on your
8:12understanding of maybe the task that
8:14you've given them, your hope is that
8:16they are going to give you the right
8:18response. And uh we want to kind of
8:20enable that world where you're uh uh
8:23depending on the model's intelligence
8:25scaling over time. So uh that leads to
8:30principles in how we actually build the
8:32protocol itself. Uh recently we uh
8:36launched the support for streamable HTTP
8:38which uh changes the the transport from
8:41SSE. uh and as you all might know
8:45streamable HTTP is is more the uh
8:47enables more birectionality and so that
8:50was uh a very controversial decision
8:52actually but uh if you're keeping agents
8:54in mind as the future makes a lot of
8:56sense because you want to make sure that
8:57agents can kind of communicate with each
8:59other. The other thing that we believe
9:02uh is that there will be a lot more
9:05servers than there are clients. Uh this
9:08we could be totally wrong on this. Uh I
9:11would love to see where the future plays
9:13out. But because we think that there
9:15will be a lot more servers than there
9:17are clients, uh we optimized for server
9:20simplicity and for the server uh server
9:23builders to have better tooling. And
9:25that does mean when we have to make a
9:27trade-off between client complexity or
9:29server complexity, we tend to optimize
9:31for pushing the complexity down to the
9:34client. So apologize in advance to
9:36client builders. Uh but it was an
9:38intentional decision. again uh would
9:41would uh be curious to see if if this
9:43plays out uh the way that that we
9:44thought it would.
9:47So I'm going to speedun through uh some
9:50project updates mostly because other
9:52talks are going to go much more in
9:53detail here. Um but last six months we
9:56launched uh ability for uh folks to
9:59build remote MCPs.
10:01We fixed o
10:03which we got wrong initially. Thank you.
10:06Uh I know that was a huge huge thing
10:08that that we got wrong initially, but it
10:09is now fixed uh in the draft spec and so
10:12would love folks to you know continue
10:15helping to push on on these things that
10:18they see don't match their mental model.
10:20Uh this was actually fixed via a series
10:24of of people from the community jumping
10:26in to work on saying hey this is how you
10:29know uh O works with identity providers
10:32and here's how we can update the
10:33protocol. So very much a community uh
10:37community effort. Um again uh launched
10:40removable HTTP as the primary transport.
10:43Uh and lastly made a couple of updates
10:45uh to developer experience um by
10:48updating our SDKs and also uh making
10:51updates to inspector which if you aren't
10:53familiar with is a really good uh
10:55debugging tool for for your server. I
10:57think it is probably our most
10:58underutilized uh tool.
11:02Looking forward, we're going to be
11:03focusing a lot more on uh that agent
11:06experience. So, we just added
11:09elicitation uh to the draft spec. This
11:13uh allows servers to ask for more
11:17information from end users. So, you can
11:18imagine you're building a uh maybe
11:22you're building a flight booking tool
11:24and uh the end user says, "Hey, book me
11:28the best flight to Atlanta." And so as
11:30the server you have a question which is
11:32what does best mean to you? Is it
11:34cheapest or is it fastest? So you ask
11:36the end user uh and now you can pass
11:39through that elicitation. The end user
11:40can respond and have that response
11:42ultimately sent back to the server. Uh
11:45we are also making progress on the
11:47registry API which would make it a lot
11:50easier for models to actually find MCPS
11:54that weren't already given to them up
11:56front. So this is again kind of on that
11:58theme of model agency. Uh we're really
12:00betting on the intelligence of models
12:02going up over time.
12:04Again working on uh developer
12:07experience. We've heard often from you
12:09all that there are uh that you know
12:12you'd love to understand what kind of
12:14the best patterns are in the ecosystem
12:16or what the standards are. And so we
12:19want to make sure that there are open
12:20source examples that uh that both we've
12:23contributed to and also the community
12:24can contribute to to kind of help build
12:26those standards and patterns together.
12:28And lastly uh we're making sure that MCP
12:33stays open uh forever and we are
12:36investing heavily in thinking about the
12:38next phase of governance. Uh so there
12:41will be more updates on that soon.
12:44And just to do a quick call out to uh
12:46the graphic in in the bottom. So a lot
12:50of people have asked uh us what it looks
12:53like to actually build an agent with
12:54MCP. Our take is that an agent really
12:58is, you know, just a server acting as a
13:00client and vice versa. Uh where you can
13:02then kind of chat back and forth with
13:05other agents, uh other servers, other
13:07clients. Um so I won't go into too much
13:09detail there. I know a lot of other
13:11people are going to be uh talking about
13:13agents in more detail, but just wanted
13:16to make sure that uh I call that out
13:18here.
13:21So, the uh thing that everyone has
13:25probably been waiting for and that I've
13:27been told uh over and over again when
13:29when I talk to founders uh what they're
13:31asking me about is uh what should I
13:34build in this space? you know if uh MCP
13:37becomes a standard what is where are the
13:39interesting opportunities so before
13:42jumping into this the first thing I'll
13:43say is that we are really early right
13:46now and that means that even if the
13:49standard exists we still need the
13:51ecosystem to be filled out and I uh
13:55would urge you to build more and more
13:56and more servers if I had to put a
13:58waiting on these three bullet points I
14:00would put 80% on the first one 10% on
14:02the second one and 10% on the third one
14:04Um so we have a lot of opportunity to
14:07build a lot more servers uh that are
14:10higher quality uh and for different
14:13verticals. Um and just to touch quickly
14:16on what I mean by higher quality. Uh a
14:20lot of people you know maybe hot take
14:22but I think a lot of people are wrapping
14:23their API endpoints one to one and just
14:26exposing that as tools. I don't think
14:28that's the right way to build an MCP
14:30server. That in and of itself could
14:32probably be a 20-minute talk. Uh but
14:35what you really have to remember when
14:37you're building a server is that you
14:38have three users. You have the end user,
14:40the client developer, and the model. So
14:43a lot of people forget that the model is
14:45a user here as well. You want to uh just
14:47as you would for API design, you want to
14:49think about what are the use cases that
14:51your end users are going to have. What
14:52are the prompts that they might actually
14:54be uh putting into the the model? and
14:57ultimately what are the tools that you
14:59then need to expose to the model to
15:01enable the model to respond correctly to
15:03those uh to those prompts. So uh higher
15:07quality servers uh and also servers for
15:10different verticals. A lot of the
15:11servers today um have been for dev
15:14tools. We would love to see uh this
15:17expand to be useful beyond engineers
15:20into verticals like sales, finance,
15:22legal, education, pick your poison, uh
15:25whatever you know best. um that uh we we
15:28would just love to see more servers. The
15:31next piece is on simplifying server
15:33building. So again, as I mentioned, we
15:36believe strongly that uh servers are
15:39going to be the vast majority of the
15:40ecosystem. There will of course be a lot
15:42of clients as well, but we think the uh
15:45order of magnitude of of servers is
15:47going to uh outweigh the order of
15:48magnitude of clients. And so would love
15:51to see a lot more tooling to actually
15:52make it easier and easier to build
15:54servers. um both for enterprises uh that
15:58are deploying MCPs internally uh as
16:01interfaces between teams and for indie
16:03hackers uh and everything in between
16:05that uh are building MCPS for external
16:08users. So anything from hosting tooling,
16:12testing tooling, uh eval deployment,
16:15etc.
16:16And then uh I snuck a bullet in here
16:19that's maybe a little bit more of a
16:21moonshot and a bet on the future, but
16:23the uh there's a bullet for automated
16:26MCP server generation. And uh again, if
16:29you kind of think back to our bet on
16:32model intelligence and model agency for
16:35the future, uh at some point models will
16:38be so good at writing code and
16:41interacting with the external world that
16:42they will actually be able to write
16:44their own MCPS on the fly in real time.
16:47And so, uh, this might be a little early
16:51for where we are today, but I do think
16:52that there will be an opportunity for
16:55automated MCP generation, um, as models
16:57get smarter and smarter.
17:00And, uh, last but not least, uh, wanted
17:03to do a quick call out for any tooling
17:05around AI security, observability, uh,
17:07auditing, etc. I don't think this is
17:10actually specific to MCP. This is true
17:13for any AI application. But I think the
17:16more that you enable those applications
17:18to have access to the outside world to
17:20start playing with uh real data uh of
17:23course the security and privacy etc
17:25implications also go up and so I think
17:28if you're going to build uh a startup in
17:30that space now is is the time.
17:34So with that uh happy MCP. Thank you.
17:39[Music]