In natural language, you can say, okay. Do this task over Slack, Notion, and Gmail, and it will get done.

Hi, everyone. I'm joined today by Karen from Composio, an integration layer for AI agents. Karen shares what people are actually using MCP for, the future of function calling, and what an Elon Musk retweet actually means for sign ups.

When something launches, people are eager to use it. So we are very quick on kind of creating examples, launching something that adds value to people's lives. Everybody in the community really liked their examples, reposted it. That got us really good momentum. If one agent using Composio does a mistake, we kind of learn from it, and the next agent using Composio shouldn't make the same mistake.

So Karen, you have had this really explosive growth. I think you've gone from, you know, some in your own words, not many users to a 100,000 users in the space of about six months. You've ridden the wave of MCP, of agents, of function calling. What what has it been like the last six months, and how did you do this?

Yeah. I mean, it has been a crazy ride. So people who don't know, at Composio, we are building the integration layer or skills for AI agents. And as part of it, there has been really kind of increase in demand. We provide, kind of, like, tools.

You can connect, like, Slack, Gmail, etcetera, via MCP, via directly function calling. So, like, everybody, I'm guessing who is listening, if they are not under the rocks, they would have heard about MCP and probably used, it with either Cursor, Cloud, or ChatGPT. So one of the biggest lever that kind of led to our, growth was MCP. We launched our MCP product in March and, like, for example, in the beginning, people were using cursor with Superbase, with Gmail, Slack, etcetera, just to kind of automate a big part of their day to day workflow. So for example, if they want to, like, just get their DB schema and query something while writing code, they can do it in cursor.

Or if they want to use their code to write an engineering doc in Notion, they can use compose MCP. So I think that was really, like, kind of an unlock that led to that virality. People were waiting to do that inside their ID. That post that I made around, like, MCP launch itself got, like, a million plus views. And yeah.

I mean, almost everybody in, like, cursor, cloud community kind of looked at it. And, like, from what I remember, that was one of the instigating moment of, like, that whole MCP hysteria as well that started. Could you could

you tell us about this post? What the post was?

So it was just our MCP launch. We had a video. We had, like, the whole kind of, like, m c p dot composer or dev launch where we were explaining how you can use it, a demo, a small demo of, like, basically connecting your cursor with all these tools and doing, like, almost all the things inside your IDE itself. Like, so when a software engineer who is writing code doesn't have to go outside to, like, chat with his teammate or write a engineering doc, read a PRD, etcetera. All of these things can be literally done inside the ID, which went pretty good.

In addition, we have been really, I would like to say, lucky of sorts, but kind of opportunistic as well where whenever something new launches, we are right then and there. So to give you an idea, when Grok three launched, we made a really cool example with Grok three, which Alan retweeted, and we got, like I I don't remember, but approximately 2,000,000 views. Then when, like, recently Grok four launched, we did a Grok CLI, which was, like, you can use like, ClockCode, you can use Grok CLI to do almost everything on your terminal with Grok. So that also ranged in, like, 500 k views. It was shared internally in x AI, etcetera, etcetera.

So, yeah, that was also interesting. So we what we have done is kind of we have latched on to these right opportune moments. When something launches, people are eager to use it, try it with various form factors. So we are very quick on kind of creating examples, launching something that adds value to people's lives. And I think I still get queries around Grok CLI.

Like, people are using it, like, day to day, like, to use Grok. So I think that, example, I think has around like, in, like, two days, it kind of got around 700, 800 stars. So that was also pretty good. So yeah.

How much I I guess a lot of people will be curious how much an Elon Musk retweet of your your demo. How much that actually translates to kind of, like, users and, like, you know, good users. Like, how what what was the impact if you can remember?

Yeah. I mean, I do remember there was a spike in sign ups. It's very difficult for me to kind of figure out if it was a good user or not, but definitely, like to a lot of eyeballs for Composio. Like, I mean, I could totally see there was, like, I think, at least, like, if not more, 100% jump in sign ups that day when Alan retweeted.

Yeah. And I wonder if it's actually, like, kind of a strategy of just yeah. If you if someone is working on something and they're, like, very famous and then you kind of they launch it and then you yeah. If you it is a good strategy, it seems, to to do that.

Yeah. I mean, like, the it's like whenever something new is there, which is kind of, for example, DeepSeek moment or Grok or Claude four, people are hungry for content around it. People, like, want to know what are the new capabilities that these models unlock. And, like, there are so many opportune moments that are coming just because all these model companies are launching models every other day. So I particularly think on social media, there's always even though if we feel it's all flooded with those type of contents, but there's always dirt of content because people are looking for more.

They want to explore what is possible. Is it even worth their time? So that's the idea where, like, if you post something around that time and, like, it's really good, it really picks up on the social media. And, like, it's not only true for actually social media. It's true for, like, blogs as well.

So some of our blogs when kind of, like, for example, any of these models release, any new frameworks release, really pick up well on like SEO, on Reddit, etcetera.

How do you like actually do that organization? Because I imagine it's like, you know, you just kind of I know there were startups, you don't necessarily have like a long term plan, but you have stuff that you're working on. And then it's like you see you're on Twitter, you see like Grokfri just got released or whatever. Then you just drop everything and like how do you how do you do that? Like is that pretty much what you do?

Or

It's not actually like drop everything, but it's like you carve out those extra us, like probably from your sleep or from your me time and like Yeah. Like, it's important enough that you can like sleep later, I guess, to grab that attention. Because like, those moments are like rare, but you have to capture them.

How important of a strategy do you think it is to do that? Because like, it's kind of an argument. If it is very important, it's like, maybe you should just drop everything. But it's like, you're like, you'll you'll do the the core stuff but you will, you know, still work on it. I don't know.

How how do you think of that around that?

Like, it's a strategy where you don't have to spend a lot of time. You do spend time, but it's fun also. Right? So, like, I'll like, I kind of like to explore when something new comes up.

Mhmm.

So it's more of, like, the strategies around my kind of liking or passion of exploring. So it doesn't feel extra work. I'm kind of anyway willing to do that extra mile of going and exploring that stuff. Like, whenever Grok four releases or plot for Opus releases, Deepsig releases right now, Kimik two. So I'm like like, I like to do that.

I think everybody in my team, my cofounder, my like, the team that we have built, thankfully, everybody likes to do that. So it doesn't feel of an extra effort. It's more around like, okay. We anyway, like, are gonna do that. Why not just kind of like distribute it or make everybody aware of what are some of these findings?

Yeah. That that makes a lot of sense, honestly. Like, it's the stuff that you don't even think about. You're just yeah. It's just like fun.

Yeah. Yeah. It's just natural. Like, it's very hard to kind of like, if you don't like to do that, it's very hard to create strategy around that, very honestly. But if you just like to do that and you like to explore new things, you like to try out, build on top of that, and see how it is working, it's like an extra mile.

But if you already do that, it's like just a hack. Like, just post it post around it and you'll get some momentum.

Scaling DevTools is sponsored by WorkOS. If things start going well, some of your customers are gonna start asking for enterprise features. Things like audit trails, SSO, SCIM provisioning, role based access control. These things are hard to build, and you could get stuck spending all your time doing that instead of actually making a great dev tool. That's why WorkOS exists.

They help you with all of those enterprise features, and they're trusted by OpenAI, Vercel, and Perplexity. And if you use them for user management, you get your first million, yes, million, monthly active users for free. I honestly don't know any dev tools that have a million monthly active users, apart from GitHub maybe. So that'll get you pretty far. Here's what Kyle from Depot has to say about WorkOS.

We use WorkOS to effectively add all of the SSO and SCIM to Depot. It's single handedly, like, one of the best developer experiences I've ever seen for what is, like, a super painful problem if you were to go and try to roll that yourself. So for us, we can effectively offer SSO and SCIM, and it's, like, two clicks of a button, and we don't ever have to think about it. It's like one of the best features that we can add to Depot. It's super affordable, which effectively allows us to like break the SSO tax joke.

And essentially say like you can have SSO and SCIM as like an add on onto your monthly plan. Like, it's no problem. So it really allows smaller startups to essentially offer, like, that enterprise feature without a huge engineering investment behind it. Like, it's literally we can just use a tool behind the scenes, and our life is exponentially easier.

Do you have any tips for people on like how to, you know because there's playing around with it. I'm sure a lot of people are playing around with new stuff when it comes out. But have you found any like patterns or like tips on posting about this sort of thing?

I would say I generally start reading about it first, like what other people are posting. Because people start like like there are people much faster than me who just get access before me and, like, post things before me. So I make some, like, mental model of, like, okay. This thing is better in this direction. Maybe it's coding.

Maybe it's reasoning. Maybe it's dot dot dot. Like, v o threes, obviously, everybody knows it's better than kind of, like, video generation a lot. And so I think getting some level of mental model okay. This is the direction probably I should push and create something around.

And maybe it will be better than all the past things and maybe it will work out.

And so I kind of interrupted you in the journey though because I we got I got weighed down on like, you're talking about getting retweeted by Elon and then the impact. But you were saying that, you know, you've kind of been able to benefit from a lot of the trends. Could you kinda carry on that that description of the journey of the last six months?

Oh, yeah. I mean, I think I probably will start a bit before, if that's okay. Like, which is, like, we began Composio, the current iteration of product last year around February 2024. So it's been what? Around, like, one and a half year since the beginning.

Like, in the starting phase, as it is, we were just building the product for first few months. And then I think we, like, slowly started launching it in a few places because we were also afraid about the product quality in the beginning. Like, we are providing authentication layer for agents. We are providing the tooling layer for agents. So it's very critical piece of the infrastructure.

So I think anything going wrong can, like, literally muddle our reputation and, like like, once you like, in an infrared product, if, like, people try it once and it doesn't work, then, like, the coming back is somewhat hard to them. They form, like, a perception around you, which is why it's really difficult. Like, the GTM is also difficult because you have to kind of somewhat balance the quality of product because you can never be perfect. So you have to balance that. So kind of we were somewhat always slow, but, like, in the beginning, we realized that there are, like, a lot of people building agents.

If you remember early on, like, whole kind of, like, agentic frameworks, like, langchin, lama index, crew AI, like, still now. But but, like, I think at that time, like, last year, because, like, the inference in time was not there. A lot of things were not there. They were, like, really getting popular. People were building on top of it, and we were part of all these Discord communities, all the different type of communities that these people had, and realized that kind of one of the biggest bottlenecks of them building agents is good quality tools.

To build any agent, you need to connect your LLMs with tools. So that requires authentication, changing the, like, whatever API schema, etcetera, to function calls, which needs to be really LLM friendly so that the LLM actually understands it well. And yeah. I mean, that's where we posted a lot on these Discord communities, got some early adopters, to test the product. And we realized, okay.

This is working. And that's where, like, I would say, like, last late last year, that's when our, like, proper GTM journey started, I would say. And we like, very similar to the trend that I mentioned, we were doing a lot of cool examples with almost all the agentic frameworks. So we made native packages for all the agentic frameworks. So using Composer with, let's say, langchin is very difficult.

You just need to install Composer langchin, and then, like, we do the whole tool formatting, etcetera, directly in the langchin format. You don't have to think about anything. Same goes with Lama Index, Crew AI, any framework that you can think of. And there they these frameworks already had a big community, and they were hungry for the tools for the integrations. And that's where I think when we posted around, let's say, with Clangjin, the community around Clangjin got really interested, and they kind of reacted really well to the content.

They started using Composure with ClangChain, same with Lama Index. And even, like, people building Lama Chain, Lama Index, etcetera, they also realized that with Composure, it's kind of like they can like, their users can build much better agents because now their agents can connect to, let's say, Gmail, like, end to end and kind of manage the emails or calendar to schedule emails. And those examples got some really good attention. I think from Harrison to Jerry to Joao, everybody in the community really liked their examples, reposted it, etcetera, and kind of that got us really good momentum.

And for those that Harrison being the founder of LangChain, Jerry, Lama Index, and then Joel is which one?

Crewe. Crewe. Yeah. Okay. Yeah.

They all are really good friends, I think. Oh, nice. All of all of us are doing really well, so that's great. And yeah. I think other than that, as I just mentioned, our blogs went also really viral on, like, different things.

We were doing blogs around, like, generally, whenever something released kind of notes around that, comparing a lot of things, people are hungry for those type of contents. And that goes viral on the SEO side as well as on Reddit. It got, like, million plus views. So at one point, I remember early on, like, January, February, like, every month, we were getting a million plus views across, like, Reddit, like, on Twitter individually. Wow.

So yeah.

When I hear, like, people talk about Reddit communities, I think my mind is, like, stuck in, like, ten years ago or like five years ago where I'm thinking like, oh, like JavaScript, like Python, like, you know, even I don't know if there's there must be a React subreddit online app, but React Native. But the yeah. It's like it hadn't really occurred to me that, yeah, these frameworks are kind of like, you know, like the Lama Index and the Lang Lang chain and stuff. It feels like those are kind of like the new places, I guess, that people hang out and like, I've been doing a lot of stuff with Mastro recently and like really enjoying it. And like, it it does almost feel like that's the way of working and it seems like a really good strategy there.

I think there's a Reddit for almost all these frameworks as well as well as, like, I think one of the most popular Reddits is obviously Singularity where people post a lot of things. ChatGPT. A lot of these Reddits like Llama. I think one of our post on, like, Llama Reddit subreddit went really viral.

What was the post?

I I don't remember particularly well, but it was around, like, probably one of the examples built with Composure where kind of like it was a user, like, kind of one of the Composure users built it and we posted around it. I think on Reddit, what works is like a detailed story of what you did and how it affected your work. Like, okay. I improved my day to day by doing using Composeo and building something around it.

Okay. So some someone, not you, the company, but someone saying, I used Composeo and it saved me a ton of time or I I was able to build this thing, I wouldn't have been able to. Yes. Yeah. Do you normally post sort of like a graphic, like a GIF or something or like a on Reddit?

It's more around like text content, but the blog itself has graphics. Yeah.

You you guys are very interesting copy. I saw like on your website, have like pricing is like it's like ridiculously cheap. And like what what is it exactly? I had to put it off. It's like

Yeah. I think totally free, ridiculously cheap, some serious business.

Yeah. So that's a very interesting bold way to pay your prices. Totally free, ridiculously cheap, serious business. And serious business is $229 a month, which is like which is I guess it's some serious business at that point. Relatively serious business at least.

Yeah. I mean, like, very honestly, right now, we are focused on getting the right usage. So I think pricing is something that we have not even put a lot of our mind about. But at the same time, obviously, everything around, like like so one thing is everything that you have, like, the white space that you have is content. Right?

I think someone said it. I don't remember who. Like, you you should use everything very wisely. And that's where kind of, like, our audience is developers and, like, they all like something interesting, funky, and that's where I think, like like, regular pricing, nobody will, like, look at it and get, like, oh, I want to use this company. Why you, like, kind of waste that space also to kind of like not attract developers, I think.

That's the idea.

Yeah. It's it's really cool. I don't I was wondering, like, one of my friends, Joe, launched this product and it was I don't think he ever monetized it, but it was like a database. And I think he it was before the pricing and he on the on the pricing page, I think he just had a picture of like Karl Marx and was like, it's free or something.

Oh, that's interesting. That's a good idea.

People like loved it. I think at one point, I was on like hack like big on hack news and like, that was the thing that people didn't know about the most. It was like, I don't know if it's like like, I I guess it feels like we have to be super serious or maybe maybe we don't on the pricing pages.

Yeah. Yeah. I mean, like, we haven't kind of like an enterprise, like, pricing, but you have to anyway get on a call for that. So, like Yeah. Why be serious on that difference?

So no one's gonna have to justify the line item that just says serious business, $229 necessarily.

Yeah. Most of the startups use that. So yeah.

Yeah. Okay. That's brilliant. Would you be able to talk about MCP? And also, I think you're in a good position to kind of separate what's the noise and what is the real where is this going?

What are the real use cases that people are are using in, you know, serious business?

Yeah. So there are two parts of it. Like, MCP is being used by direct consumers presumers, I would say, with their client of choice being Cursor, Claude, or Chachiopi. Chachiopi is somewhat like I won't say it's there because, like, that like, MCP integration is still not complete. They just support it in deep research, not in anything else.

And then there is a second audience which is, like, people building on top of MCP where, like, I'm building an agent and I use MCP for my tools to connect my like, to get my users' data or take action on my users' account. So there are two audience. The first one is, like, I want to in my day to day, while I'm building or while I'm doing my work, I already use ChatGPT, I already use Claude, but I have to like, for example, if I'm drafting an email, I'll draft it on Claude, but then I'll go to my Gmail and then, like, copy paste it, send it across. So that's a very small use case, but if you think about it while kind of, like, do writing a PRD or a design doc, I'll do it probably inside cursor because that's where all my code sits. A lot of design docs, like readmes, etcetera, are present there.

So it has, like, good amount of context. But then I'll make it there, and then I'll go to Notion to copy paste it. The idea that kind of, like, MCP brings and this is not all, by the way, is to connect where I'm kind of actually chatting with the LLMs to my actual work where I can directly create all the docs inside Notion. One of the biggest use cases of Composure MCP right now is the Figma MCP. So a lot of our users are literally like, we have a tool where you can get the Figma screens in adjacent format and essentially, Claude Opus does a very good job of converting that directly to code.

So a lot of our users get their first version of Figma to code directly via the Figma MCP. And, like, that's very interesting because, like, I think that just means as these models are improving, that's the way like, you'll be, like, mostly closer to these models. Right? So the idea is kind of, like, right now, Cloud four Roper does a 90% job of getting it right. But as kind of like in the future, it will be probably really soon, like 9900%.

Mhmm. And the then, like, why would you want to go away from, like, this place, like, where you're interacting with someone who can just do your job? Yeah. So the then the idea is, like, you'll connect your Figma, you'll connect your database, Supabase. I think Supabase was one of the most used MCP of Composeo as well, where people were doing a lot of data analysis on top of their data in Supabase, even changing schemas, or kind of just changing the code around DB, like, using super base, MCP.

So that was very interesting. So these are, like, different use cases and, like, as in when, like, model keeps on improving, that's happening at, like, really rapid pace as we know. It will be just kind of, like, people will be glued to where the models are and, like, all the other things will be coming in via MCPs or to their cursor, Claude or ChetGPT. So that's on the prosumer side. And on the people building their agents there, I've seen a mixed reaction till now.

So the benefit of MCP firstly is that you get the whole corpus of, like, MCP servers. If you build a client once, you can connect to all the MCP. You can connect to like, we have, like, around five fifty toolkits. Essentially, one tool kit is one app. So Gmail, Slack, etcetera, that you can directly connect to.

There are some, kind of, like, native, MCP's, like, Notion has released an MCP, Linear has an MCP, etcetera. You can direct to them directly and community based MCP. So that's one of the advantages of, like, if you want to build, let's say, a client and want, like, a tail end plethora of application, you can just create, like, one MCP client and connect to all the MCP servers outside. Now there are some limitations there. A lot of people want to customize different things like function schemas that the LLM sees or control the request that LLM makes, etcetera.

Control the response that LLM gets. In some cases, like, the response can be huge and, like, the LLM context can kind of overblow. MCP as a protocol and I was just listening to one of these talks where one of the main MCP creator in Claude mentions that the whole protocol, they're designing it in a way where ease of server creation is really high because they they believe, like, everybody will create MCP servers. But at the same time, the trade off is they're not trying to optimize the client DX and ability to do things is lower because they think there will be only a few clients. Their idea is there will be only ChatGPT, Cursor, and, like, probably Cloud, like, very few singular clients that will access plethora of servers.

So that's where their Yeah. Focus is not on, like, client side developer experience. And that's where a lot of our users are still using us via our DirectDX and, like, directly using function calling because then they get a much better developer experience

Mhmm.

And they can customize the hell out of, like, whatever they want to do. Mhmm. So because, like yeah. That's somewhat difficult because you are interacting with the server on MCP side. So that's where I think the whole DX is not that optimized for like, to build on.

Yeah. So, yeah, I think, like, it's it's extremely useful for, like, universal clients of, like, sorts, like ChatGPT, Cursor, because you get, like, the tail end applications be capable to build their own servers very easily and, like, use whatever they want to use, their internal, their self made, etcetera, with these clients.

Yeah. Do you think you're gonna get products which are basically just I don't know if I can explain this in my head. It might be like dumb question, but like kind of the product is you send an you send a request and then it goes to like some virtual, you know, like some server that has cursor in like kind of a sandbox and that has all these MCPs and then kind of the product is just like the agent running cursor with all these like tools and sends it back and it's like a standalone kind of thing. I don't know. Like it because if if that's the pro perfect prosumer way to build things, maybe it's like almost like the product as well as like kind of a wraps up developer in a box kind of thing.

I don't know.

Yeah. I mean, like you're talking about like a background agent of sorts. Right?

Yeah. But I guess I was just because I always think of like cursor is like where you yeah. I suppose I suppose. Yeah. I suppose it is.

Yeah. Yeah. It's a background agent. Yeah.

Yeah. I see. I Yeah. Yeah. I mean, like so by the way, we are launching this amazing new way of using Composio.

So which might be very, like, similar to what you just said, where you don't have like like, right now, you have to kind of think a lot about, okay. I have to use Gmail. I have to use Slack. And, like like, a user has to do multiple things, get the URL for Gmail, get the URL of Slack. It's kind of a very big friction.

We are launching a product with a single URL. You don't need to decide what capabilities you want, etcetera. In natural language, you can say, okay. Do this task over Slack, Notion, and Gmail, and it will get done. So, like, the whole kind of agent group, etcetera, will be on our side.

We'll manage everything on that, and you just need to give the instruction. Okay. You want a summary of last twenty four hours of messages on Slack and create that create a Notion docker on that and send me an email and send me an email at 11AM when I start my day, it will be done.

That's really cool. With agentic loops, because I I started to, like, dabble and I'm using like Maestro and with Maestro you don't I think they handle the agentic loop in a way like, because you just declare your tools and then like, you say what you want to happen which I think is like kind of the vibe like that you're describing. But it's you probably just run that like incursor. Right? Like what goes on under the hood there?

Is it just like how how are you like what does the code kinda look like? How how does the logic work well for this kind of agentic loop? Is it just like a full loop or like how are you

I think like like there there's one good post of like building agents by Devin. There's one, obviously, by the clot, like, clot team as well, like, the clot code team. I think most of the agentic loops are literally kind of, like because the models have improved so much, are converging to, like, a for loop of, like like, basically, react type of a thing. And the idea is how well do you summarize the information and the, like, the tool call responses to the agent so that it can react well. It can kind of understand and react well.

So that's the idea, giving it the right information. That's what we're trying to do and, like, providing the right tool calls so that, like, you can like, if you think about it, like, we have, like, around 550 tools right now, which are increasing at a very rapid pace. We added, like, for example, 140 tools like, 140 apps in last two to three weeks. So it's kind of increasing at a really rapid pace. If you give all of them to the LLM, it will just blow up.

It can't process that much. Right?

Yeah. So it's kind of the narrowing it down is important.

Exactly. Exactly. So we have, like, around 20,000 actions if you think about it all. So that's not even possible to kind of give in context. Even if you give, like, full context, it's like attention is sparse.

You can't give everything and, like, expect it to perform really well. So that's where I think it's about, like, designing the context that you pass to the LNB, designing the tools such that, like, the capabilities can be extended, but in the right direction when needed. So that's the idea. Making sure that you have the right what we call internally a dependency graph of sorts of okay. For doing this thing, this this needs to happen and this this needs to happen because then only you can get the like like, very in a very abstract manner, then only you can get the direction right.

Right? Where you have, like, limited context, what needs to come first, and then second, you need to decide that while passing it to the LLL.

Wait. Can you explain that again? I just I feel like I didn't quite

yeah. Yeah. So example, like, it's a kind of like, the example that I gave, like, summarize my Slack messages Yeah. And then pass it to Notion. If I pass all the plethora of Slack tools, Notion tools, etcetera, directly to LLM, it will just blow up.

It won't perform really well. So kind of like, okay. First, I need to do this, then I need like all the related tools around, like Slack of kind of like conversation related. I pass that and then once that done, pass the tool for, like, specifically selecting the right tool for Notion, passing that. And then, like, if I want to send an email, passing that.

It's like a very simplified format, but, like, essentially, there can be, like, really complex dependency graphs internally Okay. That we make and pass the right context.

And you're calling the depend the dependency graph is like, okay, they want to do this thing. Therefore, first we need to like of the dependencies, we need to decide which tools. And then maybe that has a dependency somehow of like, okay. It comes back up to just the final. Yeah.

Yes. Okay. Yeah. Interesting. Yeah.

I guess I guess like a lot of calls.

Yeah. It's it's it's a tough problem. I think we are internally also realizing we need to kind of abstract a lot of things out and, like, we are essentially following the skill layer. If you if you read the copy on the website also, it's like skills. So kind of like learning from these kind of patterns and, like, forming a higher abstraction skill so that the like, it's kind of like if you think about it, there's a saying in developers, developers are as good as their tools.

Mhmm. So and, like, we as human, we evolve. Right? Like, our tools also evolve with us. Yeah.

So, like and we learn things. Kind of, like, we learn something to do. We have, like, a subconscious memory of sorts which we develop of, like, let's say how to play chess and like when this pattern emerges, what to do. Yeah. And that's where we are trying to go with this, where LLMs are great conscious brain, but doing things require a subconscious of remembering the patterns and that's what we want to develop at Composio.

Interesting. So if someone just so people are just thinking, oh, okay. You connected this tool to another. But actually, you're more like, if you just give a tool that you DIY ed yourself to an agent, it's not gonna use it as well as it will use our tools because it has more kind of guidance and learned patterns

and stuff. Exactly. So like like it's like if one agent using Composio does a mistake, we kind of learn from it. And the next agent using Composio shouldn't make the same mistake.

That's very interesting. I hadn't thought about that at all, to be honest.

Yeah. Very cool. Yeah. It's not just a hard problem. So like a lot of research were going around it internally.

They are kind of like looking out like, we have, like, some really cool folks working on top of this problem, but we are always looking out for more folks interested working with this problem.

Okay. Well, people, if that sounds interesting to you, you should reach out to Karen. Help us use tools better. Yeah. It's it's it's very exciting.

I I'm personally I'm I just love this stuff. I think it's like I I recently got extremely excited about all the tools and like just even just like like simple things like you give it you give an agent the ability to do Google searches. And it's like, suddenly you can do so much more stuff. And you think that like, oh, it just the the LLM can like know stuff. But it it it just works so well.

Like it I don't know. It's it's very cool. I'm very excited about this space.

I agree. I think really cool things are about to come.

What what are you most excited about? Because you've you've been able to see the pat or or you've, like, been very fast with the patterns. Are there anything that you're what's the next MCP? What's the next, like, agents? Or or is that the end?

What are you what are you most excited about?

Yeah. Honestly, it's very difficult to predict in the current world because, like like, every other two weeks, there is a new release. But at the same time, I think one thing is for sure, things like, there are two things that will definitely improve. One is, like, latencies. I think the same intelligence that you're we are using right now will definitely kind of, like, fasten up.

Like, we'll get much more, like, tokens per second, etcetera. And, like, the intelligence, which direction it improves, it's very difficult to kind of predict. Like, obviously, coding has improved a lot. Reasoning has improved a lot. We do generation has improved a lot.

But, like, it's very difficult to predict. Okay. Next model, I think maybe, like, Frontier more labs can predict it. Like, what are the next capabilities and log? What are they focusing on?

But from outside, it's somewhat difficult to predict. But I think, like, I'm like, because of this kind of, like, constant direction of latency decrease, cost decrease, I think I'm really excited next about, like, for example, voice related stuff. So a lot of things which are, like, people building in voice and, like, want to use MCPs around it. Like, for example, in a call, I'm sitting, like, we are both are interacting, but there might be a kind of, like, a third bot where we'll just be okay. Go and do go and, like, summarize the meeting notes till now and put it on Notion.

And it will do that all, like, in the call while we are kind of talking and, like, come back to us. So that time that future, I'm really excited about for sure.

I I completely agree. I think that's extremely exciting. You should be like, oh, like, you know, we were talking about like, me, it's always like podcast related stuff. So it's like, oh, like, let's say we were you know, you and I were talking about, oh, what should be the thumbnail for this episode? And we were like, okay.

Well, it's probably like, you know, a 100,000 users in six months or something. And it's like, okay. Hey, like, you know, not Alexa, but some other thing. Go make that and like, you know, just like message us when it's done and like add it in. It's just like I feel like it's so cool.

I don't know.

Yeah. Yeah. Yeah. That could be really cool, I think. And yeah.

So much can be unlocked. Right? Like, think I think we always have, like, these Fireflies or kind of note takers that are on. And they they also prove to be so valuable. So I'm just thinking if there's, like, literally a AI person sitting there, how much valuable it can be.

Yes. I completely agree. Yeah. Karen, this has been really fun. I I've learned a lot.

I'd I'd never thought about the fact that you're learning how to use the tools better. That was like kind of blowing my mind. So that's that's very exciting. Is there anything that you would say to DevTools founders? I know you said that you've listened to the show as well.

So by the way, really appreciate that. What what kind of things have you do you feel like you've learned in the last six months that you'd already say to other DevTools founders?

I mean, like, I think one of the biggest learning source has been your show. So thanks for Sorry. I'll Like Thank you. I always look forward to new episodes. That's my gym time.

Listen without any I think whenever a new episode comes in my gym, kind of put it on. So I think two things that I've learned really well, in kind of, like, developer tools and scaling developer tools. One is developers don't it's like a common thing, but developers don't like to be sold. So that's why we have to like, if we develop something cool, like the examples that I mentioned, they just, like, like that. And if they like it, okay.

I like, for example, with me, I'm a developer. Right? So if I like something, I'm like, okay. I want to use that. I want to build that for myself.

And then I'll go and, like, kind of try to, okay, see what's there and, like, experiment with it, etcetera. That works the best in the developer community. Build something cool with whatever product and, like like, sometimes it will work, sometimes it won't, but, like, just kind of write about it, post about it, like, just let it open. And, like, in most cases, if it's really good, like, people will be excited and, like, people will try to use it and kind of that's how, like, most of the developer tools I have seen work really well. And the second is, which is, like, a long kind of, like, lesson and, like, we are also learning on the ways you have to optimize that, like, moment really well.

Because in the early on, if you don't get that moment well, and if something is not like, for whatever reason, if you don't get that moment and, like, you get the reverse moment, which is like, okay. This is not working. That's, like, catastrophic. So I think that, like, onboarding journey, etcetera, which we are also, like, I think correcting as we speak all the time because, it's always kind of you find a new friction and you have to optimize for that. So that's a constant learning and iterating.

But those are the two things that, like, are the most important Yeah. In my mind.

So the the first thing of just, not not trying to sell too hard and kind of exploring and and putting out, actual like good stuff that is just genuinely interesting. And then optimizing for the moments and also removing the like, oh no moments of like like this sucks. That's that's amazing.

Yeah. Those are the two things that I've kind of like I think we are trying to optimize a lot, always.

Amazing. If people wanna learn more about Composeo, where would you like them to go?

Yeah. We have like our website is composer.dev. Our Twitter handle is composio h q and like yeah. If they want to reach out to me, my Twitter handle is karen matthessex. Yeah.

And you've just raised $29,000,000, so people can start placing bets on how long it's gonna be beforeyour.com on the compose here as well.

Yeah. Yeah. I think people should check out that video.

Oh, is it

those is

it yeah.

Yeah. The launch video is like, I think, really good. Like, a lot of like we got, like, some really good views and some really good reactions.

Amazing. Yeah.

It's a La La Land themed video. So people who are, like, La La Land or musical fans, they will really like it.

If If you like La La Land and MCP, check it

out. Yeah. That should be soon. We are doing another one really soon. So

Amazing. Karen, thank you so much, and thanks everyone for listening.