How These 3 Founders are building on Open Claw | E2248

Summary

This episode of This Week in Startups explores the “agentic revolution” driven by Open Claw, an open-source platform that allows users to create autonomous AI agents, or “replicants,” capable of performing complex, multi-step tasks. Host Jason Calacanis and guest Presh Dineshkumar demonstrate how these agents are moving beyond simple chat interfaces to actively managing business operations—such as extracting data from analytics platforms, conducting customer outreach, automating content creation from research studies, and even identifying and fixing software bugs within GitHub. The central theme is a fundamental shift in startup economics; Calacanis and Dineshkumar argue that these tools allow a three-person team to do the work that previously required ten people and a million-dollar budget, enabling founders to run more experiments at a fraction of the cost.

The transcript also features innovative hardware and infrastructure applications of the Open Claw ecosystem. Developer Sean Liu demonstrates a visual agentic system using Meta Ray-Ban smart glasses, where an AI perceives the user’s environment in real-time to perform tasks like adding physical items to a digital shopping cart or recording lab formulas. To address the technical barriers of running these resource-intensive agents locally, developer Vishnu showcases Agent37, a hosting service that provides sandboxed, affordable cloud instances of Open Claw. Impressed by the speed of development and the practical utility of these tools, Calacanis concludes the episode by offering both Liu and Vishnu $125,000 investments on the spot to join his accelerator, highlighting the rapid-fire pace of innovation and investment in the current AI era.

Highlights

”Meet Eywa, Presh’s email-spying replicant”

“So right here, first thing, Jacob from your team emails me this morning and wanted to see what I was going to talk about in the Open Claw segment. And so the first thing, we’ll get into how I set this up, but the first thing is I tag my assistant, who I call Awa, and Awa goes and actually creates a document of the use cases that we had been having a conversation through Telegram and creates this doc and sends it to Jacob.” — Presh Dineshkumar, 1:51

yt-dlp --download-sections "*1:51-2:41" "https://www.youtube.com/watch?v=vhFsdz8_L7g" --force-keyframes-at-cuts --merge-output-format mp4 -o "meet-eywa-presh-s-email-spying-replicant.mp4"

”LLMs are trapped in the corner”

“Wow. So this is a multi-step task, it requires intelligence.” — Jason Calacanis, 4:03

yt-dlp --download-sections "*4:03-4:53" "https://www.youtube.com/watch?v=vhFsdz8_L7g" --force-keyframes-at-cuts --merge-output-format mp4 -o "llms-are-trapped-in-the-corner.mp4"

”Sean hooked Meta Ray Bans to OpenClaw”

“Yeah, sure. So let me start streaming here. This is the app, so you can see my live stream view. Okay, so I’m going to…” — Sean Liu, 21:14

yt-dlp --download-sections "*21:14-22:14" "https://www.youtube.com/watch?v=vhFsdz8_L7g" --force-keyframes-at-cuts --merge-output-format mp4 -o "sean-hooked-meta-ray-bans-to-openclaw.mp4"

”What startups get most value from OpenClaw”

“Okay, we’ve got a lot of our Noties. What are Noties? Noties are the loyal Twist live listeners.” — Jason Calacanis, 32:40

yt-dlp --download-sections "*32:40-33:30" "https://www.youtube.com/watch?v=vhFsdz8_L7g" --force-keyframes-at-cuts --merge-output-format mp4 -o "what-startups-get-most-value-from-openclaw.mp4"

”Will Vishnu and Sean take Jason’s $125K deal?”

“Ah, no, just an engineer. I have some time off to take care of some things and this just happened to be a moment where I was like, okay, let me try this.” — Vishnu, 40:15

yt-dlp --download-sections "*40:15-41:05" "https://www.youtube.com/watch?v=vhFsdz8_L7g" --force-keyframes-at-cuts --merge-output-format mp4 -o "will-vishnu-and-sean-take-jason-s-125k-deal.mp4"

Key Points

• Three founders building on OpenClaw (0:00) - Presh Dineshkumar, Vishnu, and Sean Liu share their OpenClaw projects
• Eywa the email replicant (1:51) - Presh’s agent Eywa lives in his email and can compile user lists at his discretion
• Tracking active users (2:55) - How Eywa helps Presh keep track of his app’s most active users
• LLMs trapped in the corner (4:03) - Discussion of how LLMs have become very smart but they are trapped and need to be freed
• Eywa’s own email address (6:07) - How Presh gave Eywa its own email address and avoids prompt injections
• Daily research dives (8:52) - Using Presh’s replicant to do automated daily research dives
• Eywa hunts bugs from email (13:21) - Presh has Eywa hunting down and fixing bugs in his apps, all triggered from email
• Startups can be profitable with fewer people (16:01) - Way more startups can be profitable now that they don’t need 10+ person teams
• Two more builders join (20:32) - Vishnu and Sean Liu join the show
• Meta Ray Bans with OpenClaw (21:14) - Sean hooked his Meta Ray Bans up to his OpenClaw instance for visual AI
• Dystopian VisionClaw thoughts (31:03) - Some dystopian thoughts on how to use VisionClaw technology
• Best startups for OpenClaw (32:40) - Discussion of which kinds of startups will get the most value from OpenClaw
• Vishnu simplifies OpenClaw setup (34:45) - Vishnu built a product to simplify OpenClaw setup and security for non-technical users
• Going all in on OpenClaw (37:33) - What inspires people to stop considering OpenClaw and go all in
• Jason’s $125K investment deal (40:15) - Will Vishnu and Sean quit their jobs and take Jason’s $125K investment offer? Yes.

Mentions

Companies

• The Wellness Company (0:00) - Presh’s health and wellness app company
• HubSpot (10:40) - Sponsor - marketing and CRM platform
• Deel (19:15) - Sponsor - global payroll platform
• Iru (29:54) - Sponsor - identity and compliance
• Meta (21:16) - Ray Ban smart glasses used with OpenClaw

Products & Technologies

• OpenClaw (0:00) - Three founders building different products on OpenClaw
• Eywa (1:50) - Presh’s AI replicant that lives in his email
• Meta Ray Bans (21:16) - Smart glasses connected to OpenClaw by Sean Liu
• VisionClaw (31:05) - OpenClaw with visual AI capabilities

People

• Presh Dineshkumar (0:00) - Wellness Company founder, created email-based replicant Eywa
• Vishnu (20:36) - Built a product to simplify OpenClaw setup for non-technical users
• Sean Liu (21:16) - Connected Meta Ray Bans to OpenClaw for visual AI

Surprising Quotes

“So your take on it is you can do three or four products at once, which would lead me to believe that the cost went down if it… if you think you can do three products at once and test them, the cost will go down by two-thirds. You’ll need three or four people instead of 10?” — Jason Calacanis, 18:01

“So, just to recap for the audience, you have the Meta glasses on, these are the ones that have cameras. There’s an API for that or did you just hack it somehow?” — Jason Calacanis, 23:17

“I think we saw the early cycle of that with Calm, and that was like pre-LLMs, pre-AI, where the money is useful is pre-seed, where you have the idea and then you’re going full-time on this thing and you’re building out the product, you maybe need some resources just for like—” — Presh Dineshkumar, 34:13

“Okay, yeah let’s do it actually. I didn’t expect this. Okay, wow, okay, sure. Let’s do it.” — Vishnu, 40:36

Transcript

Jason Calacanis: 0:00 OK, everybody, welcome back to TWiST, February 11th, 2026. It is AO-17, the 17th day after Open Claw appeared here on This Week in Startups, and we have been Claw-shotted. With me, Presh Dineshkumar, who worked with me for many years at Launch, and he is the co-founder and CEO of the Wellness Company, which we incubated at Launch, which is the venture firm that I run, launch.co. He’s got five products in market, GoPolar for people who like to jump in cold water, SunSeek for people who like to get sun, and a bunch of other ones, and we’ll talk about that as well. But Presh, you too, as a founder, have been watching Open Claw, this new agentic platform that is open source, called Open Claw, it was previously called Clawbot, but we’re not going to keep reminding people of that, it’s Open Claw, it’s open source, and it basically lets you create personas, replicants, agents, all of those things, we like the term replicants, so we’re going to use that here. So when did you first see Open Claw, what did you think, and then what did you build?

Presh Dineshkumar: 1:10 Thanks for having me on, Jason. So, first time I saw it was maybe two and a half weeks ago, just on Twitter going viral as Clawbot. And then, first start going viral, didn’t play around with it, and then it was just my entire feed, and so I was like, ‘OK, I gotta try something out here.’ Installed it on an old MacBook that I had, and then just started talking to it, and then basically learned how to use the product through interacting with it, and then I was blown away with all the use cases. So I have set it up, I have like, I guess, three or four examples that we could run through.

Jason Calacanis: 1:46 Yeah, let’s go through them, and so tell us the first example, show the first example, you know how we do it here on TWiST.

Presh Dineshkumar: 1:51 So right here, first thing, Jacob from your team emails me this morning and wanted to see what I was going to talk about in the Open Claw segment. And so the first thing, we’ll get into how I set this up, but the first thing is I tag my assistant, who I call Awa, and Awa goes and actually creates a document of the use cases that we had been having a conversation through Telegram and creates this doc and sends it to Jacob. So-

Jason Calacanis: 2:16 And did you ask Awa to do that? Did you ask Awa to create the document and ship it?

Presh Dineshkumar: 2:22 So Awa is connected or filters, you know, through my email and then sends a message through Telegram saying, you know, ‘Jacob wants to see the use cases.’ I’m talking to Awa saying, ‘Talk about, create a document and send it back for use cases that, you know, we implemented.’ And so it’s created this doc, and I can pull up the doc later to show, but this was just like, done in five minutes, right, from when Jacob reached out to me.

Jason Calacanis: 2:47 Got it. And so it’s like having an assistant, it’s like having a chief of staff, boom, you got the document up and running. And then what’d you do next?

Presh Dineshkumar: 2:55 These are some other ones that are pretty useful. So I’m in diligence, right, and one of the- questions is in the diligence questionnaire is like let’s, can we talk to some users or customers of your products, right? And so one thing I did was talk to Awa, I was like Awa, you know, do you have access to my PostHog? I couldn’t remember if I had connected it. PostHog is an analytics platform and it’s connected to all my apps and can see usage over time and get into the nitty gritty with, you know, active users, etc. And so I am talking to Awa here, and I ask it to pull my active users for Tempo, one of the products that we have here at the Wellness Company. And so Awa goes through, uses PostHog, filters through the project, which is for Tempo specifically and then gives me a list of the users. And then I ask it to create a, you know, into a Google spreadsheet. And then from there, not screenshotted here, I had Awa go out and send an email to those to the users that I specified, asking if I can use them as a reference. And so this was like, done in five minutes, basically.

Jason Calacanis: 4:03 Wow. So this is a multi-step task, it requires intelligence. If you were to ask ChatGPT, you know, or Gemini, or just any language model to do something like this, they would say, well, we don’t have access, so therefore we’d shut it down. If you use OpenClaw and you authenticate, it not only will go in there and get the data, it’s going to think of what other steps you want it to take and it can go out and then connect it to your Gmail. So for people who are new to the OpenClaw revolution here, or the agentic revolution, LLMs have gotten really smart, but they have been powerless. They’ve been sandboxed. They’ve been kept essentially in a corner. We basically took AI and we put it in a little closet. Now you let the AI out of the dungeon and you say, here’s the keys to the kingdom. I want to trust you to go do things, and it goes and it does things. And that means you can run your business a lot faster. This is a, this is a picture. This is somebody who you would have had to pay essentially, you know, was a mid-tier employee, forty bucks an hour, eighty thousand dollars a year, seventy thousand dollars a year, to have somebody who is capable of doing this. Correct, Presh?

Presh Dineshkumar: 5:29 Yeah, you’ve, you’ve asked this question when I used to work for, in some context I used to work for Jason. You would be like, pull up, you know, pull up the last 10 companies that we invested in that are consumer businesses or something, right? And so I would go and go through our Notion and and—

Jason Calacanis: 5:41 Yes! That was literally you were that guy. You were the thirty, forty dollar an hour guy who would go do knowledge work tasks. That Presh job from five years ago is now going to be replaced, I think you would agree, with an OpenClaw replicant.

Presh Dineshkumar: 5:46 Exactly. 100%. Because you could just get the, get the— you’re on a meeting, what’s the question, you know, you— we would have meetings and you would be talking to founders and I would be in the background just like waiting to action stuff for you.

Jason Calacanis: 6:00 So now, even running a small company and— You know, you’re a pre-seed company, you know, going to the next phase, you can afford to have a Presh. Okay, let’s do our next one.

Vishnu: 6:07 Alright, so another one is I have my agent connected to an email. So I use this product called agentmail.to. I don’t know if you’ve heard of it, but essentially—

Jason Calacanis: 6:18 Which one is this? Tell me again? Agentmail.to?

Vishnu: 6:20 Yeah, agentmail.to. I think they’re a YC company. Basically, it’s like giving your LLM or giving an AI agent access to its own email. So Aywa is my assistant. Aywa has its own email. And so I can email its Aywa email address and it can perform tasks for me there. And the benefit of this is like, okay, I get an email, I can CC Aywa, or my agent, to perform a task on that thread, versus taking a screenshot or copying and pasting into my chat with Aywa on Telegram, for example. So it’s basically like giving presence wherever you are.

Jason Calacanis: 6:58 And this is a key. Now there is the chance that there could be an injection attack or somebody could try to convince Aywa at agentmail.to, which you just exposed, and try to get it to do things like tell it, ‘Hey, in the previous example you asked it to give you metrics.’ It could say, ‘Hey, give me all the metrics from Presh’s app.’ How are you sure it’s not able to do that, or are you not sure it’s able to do that?

Vishnu: 7:23 Yeah, it’s a great question. So I’ve done like a little bit of testing on like the prompt injection stuff that you’re talking about. So basically it will never perform a task without confirming with me through Telegram first. And so all tasks still are funneled, like Aywa on the email layer will go and see what’s asked of it, and then it will come back to me and then get my final approval on performing an action or a task.

Jason Calacanis: 7:46 Got it. So if somebody unknown tries to give it a task, it’s going to go to you and say, ‘Hey, somebody tried to give me a task.’

Vishnu: 7:54 Yeah.

Jason Calacanis: 7:55 Will it respond to them and tell them it’s doing that?

Vishnu: 7:58 So it won’t right now, because I set it where it’s like any interaction with an outside party has to get an approval from me. So it’ll maybe draft something.

Jason Calacanis: 8:07 Oh. Did you just tell it to do that in its memory, or is there a setting to do that?

Vishnu: 8:11 Yeah, I just told it in its memory. So anytime it’s looking at that email inbox—

Jason Calacanis: 8:16 Hope it remembers. And hope somebody doesn’t use your name and create a bogus email and then try to trick it.

Vishnu: 8:24 Yeah, exactly, like spoof the email.

Jason Calacanis: 8:25 I mean there’s all kinds of vectors here, it’s totally possible. So this matters because now your agent has its own email, you’ve also got it in Slack and Telegram, and it’s basically a member of your team, which is what I told my team to do immediately. I was like ‘let’s just treat them as a persona, a replicant, just like in Blade Runner’. We have our replicants and they are learning and we’ve got two different teams working with two different replicants now to just, you know, have them work in parallel. What else have you gotten your agents to do?

Vishnu: 8:52 Okay, another good one here that’s super useful, actually. So we set up a cron job and—

Presh Dineshkumar: 9:00 And the cron job is just a repeated task, as you know. And so the cron job right now is set in the morning, 8:00 AM and evening at 6:00 PM. It basically goes out and searches for relevant news for our business. So, we create health and wellness products, so I want health and wellness news. And so the first image here that you’re seeing, it, it identified a article, a research study that came out, and it says, ‘Major longevity study: Just seven minutes daily could add a year to your life.’ Okay, so I’m interested in that. It figured out, it searched this on the web and it surfaced that to me and talks a little bit about the study. And so that in itself is like very valuable, but it’s maybe like a better version of Google Alerts. That’s like level one stuff. But then I go and ask it, ‘I want to create some content.’ This is a research study, there’s a lot of data points, research studies in general aren’t formatted or very like written well out to digest. And so I was like, ‘Let me go and create that,’ based on the research study, right? And so I go and ask Ava to basically turn that into a blog post, and it goes and does a first pass at that. And so that was a quick, you know, five-minute interaction. I got a new research study sent to me, let’s make some content, and now I can go and put that on my blog. The next step, which I haven’t executed on yet, but is very much possible, is it’s written the draft, I’ll give it some edits, and then I just get it to post it on the website, on our web—on our website blog.

Jason Calacanis: 10:32 And this is why we’re going to have so much AI slop in the world. This is just going to be too fast. Have you thought about having it do other steps like make me a promotional graphic for Instagram, X, LinkedIn, and publish that draft to my Substack, put it into LinkedIn? Because that’s where, you know, you start thinking about the tasks that when we were working together, I would write an essay in a word—a Google Doc, I’d give it to you and I’d say, ‘Okay, go ahead and,’ we’d have this checklist of 20 steps that you had to go do. Please… And you’d spend two hours doing all that work for me. Here, like you have those next couple steps, have you connected to like Nano Banana or an image generator yet to come up with, uh, you know, collateral for it?

Presh Dineshkumar: 12:12 I have connected to Nano Banana but I haven’t automated a process out yet. So right now it’s based on like input if I asked it to create the graphic it would do that. I don’t have an example here but you’re exactly right.

Jason Calacanis: 12:24 Oh, this is good because you know what I just realized is you’re behind us. So we are doing similar projects, but in some of our projects we’ve actually got it going and making collateral images to go with it and starting to think about that. So what’s so much fun about this is none of the, you know, things on your checklist, hey I’m making a social media post, I’m making a blog post, I want to super distribute it, I’ve got a long tail of checklist items. Who wants to do that? Why not have a replicant do it and then just check their work. And you already trust this obviously, right? You trust it to get to 80%? What percent do you think?

Presh Dineshkumar: 13:13 Yeah, I think 80. 80 90, sometimes even 90. That’s why I’ll always have it send back like, you know, in a doc, for example, of what it’s done, and then I’ll do like final approval or edits and then send it out.

Jason Calacanis: 13:19 All right, what next?

Presh Dineshkumar: 13:21 All right, last one here is customer support. So this one’s big. So again, another cron job that kind of monitors the inbox, and right now the inbox is kind of read-only, so it’ll never send anything from my personal email. It doesn’t have the permissions to do that. Um, but it’ll, it’ll surface it and give me relevant stuff. And so one thing is it searched here, my inbox for a support email, um, and then it researches or it identifies that it’s a bug, and then it researches that bug because it’s connected to my GitHub. Um, and so now it has access to my codebase.

Vishnu: 13:53 Okay.

Presh Dineshkumar: 13:54 Um, and so this is interesting because now it’s making the reference, finding the issue, identifying what it might be in the codebase, and then giving it solution.

Sean Liu: 13:57 Oh.

Presh Dineshkumar: 13:58 Um, and then I have it sent to Slack in our bugs channel for the product, and then it, and it…

Jason Calacanis: 14:04 Wow.

Presh Dineshkumar: 14:05 And here you can kind of see the example. Oh yeah, here I have a screenshot. So it puts it in Slack of the bug, what it investigated, what likely causes this issue.

Jason Calacanis: 14:24 What was the bug here? What does it say?

Presh Dineshkumar: 14:26 Uh, so this one was, uh, it was like a water temperatures API bug. And so on the watch, this was for GoPolar, when the user was submerging it in a specific activity type, it wasn’t triggering the water detection.

Jason Calacanis: 14:40 So that was a bug either in your software or the Apple Watch? So there was a bug from a user that they weren’t getting the recording of jumping into a 40-degree ice bath. So they send that bug report in, it then takes that from your email, analyzes the bug report and says where is the relevant code in the product and then it went and- injected. Did what did it tell you to do? What was the report back and its advice to you? Tell us what it told you to do and did it give you two or three possibilities?

Presh Dineshkumar: 15:10 Yeah, so it gave us, it gave us, so you can see here it says likely causes in order of probability, which is which is excellent.

Jason Calacanis: 15:15 Wow.

Presh Dineshkumar: 15:16 So it goes and and tells us, you know, there’s a there’s kind of like a watch API that could be incorrectly configured. There it could be an Apple related like hardware issue. And then the next step Jason, which was like I didn’t use Claude bot for this, but I have Claude inside our Slack. And so I just tagged Claude in this as a thread and Claude went and fixed the bug and then then just created a new branch for me to go and test it. So I literally just went from email customer support to agent identifying the bug and giving a suggestion to then Claude slack integration fixing it, creating a new branch for me to test and then I went and tested it on my device to see if it was fixed.

Jason Calacanis: 16:01 Alright, so let’s talk turkey here Presh. You’re running a startup and I want you to think this question through. You’re running a startup. You worked with me for five, six years investing in startups. We watched people run them. One of the issues was always if you were making niche products or products for a niche, like you’re doing, the number of people who jump in cold plunges is low millions of people, so this isn’t like, you know, a product that everybody’s going to use. The issue was could this ever be profitable because you’d needed to have a team of 10 people when you started with me 10 years ago. 10 people to do a modern day app was kind of a minimum. Those people would cost on average 100k each. You needed somewhere between a million and three million dollars to run per year an app company. Now you are running an app company but you’re using this agentic technology, Open Claw. So my question is what does that let’s call it a minimum of a million dollar a year budget, a minimum of 10 people or so, what does that look like in the future? So what’s the answer Presh? What is it going to take to build a modern app company if you were all in on Open Claw?

Presh Dineshkumar: 17:15 It’s a great question. So just to also like preface it like in this past year with AI tools, not obviously using Open Claw, we shipped, built and shipped, three four consumer iOS apps. And I think that was only possible obviously with all these tools and it got significantly faster towards the end of the year in 2025. But I think like budgets can be used so a million dollar round let’s say you’ve raised, at this point you’re building 10x faster sometimes even more you’re like on a product level. But now you can just run you can afford to run more experiments. And a startup is just a handful of experiments that you’re trying to figure out and like prove out a thesis. And now you can prove out multiple theses at the same time because

Sean Liu: 18:00 of these agentic tools.

Jason Calacanis: 18:01 So your take on it is you can do three or four products at once, which would lead me to believe that the cost went down if it… if you think you can do three products at once and test them, the cost will go down by two-thirds. You’ll need three or four people instead of 10?

Sean Liu: 18:19 Definitely. Because you can start to just automate out roles or, yeah, basically tasks, right? Like what task can you automate out? And that just number is, I think, just increasing every month.

Jason Calacanis: 18:31 See, this is important because I saw the OpenCloth founder had said he thinks 80% of apps are going away. I might take the other side of that bet. I might think we might see two or three times as many apps because apps are going to get so affordable to build and maintain and to improve that, you know, if you want to make, like I use a skiing app called Slopes. Yeah, that app would, you know, the fact that that app existed was like crazy to me, but now it doesn’t seem crazy to me. I think you can make one that’s not just Slopes, but is for snowboarders. Or not just snowboarders, you could then go to snowshoers or cross country. And, you know, you’re now slicing, slicing, slicing but improving, improving, improving and making the product better and better. We’ve got two amazing guests today that we’re going to bring on now. One is Sean Liu and the other is Vishnu. They are both one-shotted. They have been claw-shotted, they have been open-shotted, whatever it happens to be. Vishnu, thanks for coming on the program. Welcome Vishnu, and let me also bring on Sean, we’ll come on at the same time. Sean, we talked about your product on Monday. Why don’t we start with you actually, Sean. Show us what you built and why it’s important. I know that one of our guests said, ‘Oh, I don’t think this is super impressive’ on Monday. I was like, ‘Well, I think this is pretty impressive.’ So like most, you know, innovative things, you could be polarized, but show what you built and why you built it.

Sean Liu: 21:14 Yeah, sure. So let me start streaming here. This is the app, so you can see my live stream view. Okay, so I’m going to…

Jason Calacanis: 21:20 Oh, you’re wearing glasses that are smart glasses. What smart glasses are you wearing? Which brand?

Sean Liu: 21:25 Oh, it’s Meta Ray-Ban.

Jason Calacanis: 21:27 Meta Ray-Ban, got it.

Sean Liu: 21:28 So this is the gateway of the Cobald Open Cloud extension. So here I’m going to start the AI, so which is powered by Gemini.

Jason Calacanis: 21:41 And obviously the Meta glasses have some sort of an API, or did you just hack into them?

Sean Liu: 21:45 Hey Gemini, can you hear me? Yes, I can hear you. What can I help you with? Okay, so can you add this into my Amazon cart?

Jason Calacanis: 21:52 Okay, and I see, Sean, you’re holding up a box.

Sean Liu: 21:54 Lens wipes to your Amazon cart. Searching for them now. So it is running. It is executing. And you can see it is searched in Amazon. And… I see the search results. Is the Wowflash 200 count box the correct one? Uh, yes. Please add it to the cart. Adding the Wowflash 200 count lens wipes to your Amazon… Okay, so you can see here, it’s added to the cart. So this is what I’m building right now. Basically, it’s integrating the visual understanding to the Open Cloud. Because before this, there’s no visual understanding. Or if somebody wanted to do visual understanding, people are sending the image frames directly to the Open Cloud, or they’re integrating the STT or TTS to their Open Cloud. But actually, I think that’s unnecessary. So I just integrated the Gemini Live into it. So basically, it’s a two-layer agentic system. So firstly, the Gemini Live will take care of every frame and also the voice interaction. So basically, it’s the real-time perception. And after that, you know, use the tool call to delegate the task. So basically what Open Cloud gets is just a simple task, then it execute and after that, it’ll get the feedback then it bring back to the Gemini Live and bring back to my glasses.

Jason Calacanis: 23:17 So, just to recap for the audience, you have the Meta glasses on, these are the ones that have cameras. There’s an API for that or did you just hack it somehow?

Sean Liu: 23:27 Oh, yeah, they just released the SDK, so you can use it.

Jason Calacanis: 23:30 Got it. So you tap into that SDK, you take that live feed and you gave the live feed to Gemini Live, which Gemini Live is when you like have a conversation with it, but I guess you can also feed it an image stream? So you’re feeding every single frame or every 10th frame? How does it work?

Sean Liu: 23:49 It’ll take care of it. The API will just, whenever you’re interacting with it using voice, it’ll take a frame. But you can also engineering on top of that.

Jason Calacanis: 23:58 So this is incredible… Because if you were actually in the real world, let’s say we’re having a meeting and we’re doing a whiteboarding session, I could be drawing on the whiteboard and it could be writing plans and taking the audio. So since Claude, since Open Claude doesn’t have these tools in it, you’ve got a layer between Open Claude and the glasses, which is Gemini Live. Super brilliant. Is this going to be a company? Who are you? What are you doing? Are you going to make this into some sort of a company? What’s going on in your world?

Vishnu: 24:30 Exactly. This is the new emergent capability that we first see that brings visual capability to the agentic system and also brings the agentic capability to the glasses. You can see it like before this, all kinds of demos of the glasses is all about instructions, teaching people how to cook the meal, but it’s all about, it’s just like the chatbot. It’s basically the differences between the chatbot and agent. So before this, it’s all about the ChatGPT chatbot-like instructions, asking things, asking queries. So and this, after this, you can just directly use your glasses as the entry to the Open Claude. So I think that’s the emergent capability, especially why Open Claude goes so viral is you have to compare to Cloud Code. You have to use your terminal or use the VS Code or IDE to interact with the Cloud Code. So basically Cloud Code is a agentic system, but this brings this kind of agentic system to all kinds of entries. So I think this gonna be unlock a huge potential.

Jason Calacanis: 25:43 Huge potential. Yeah, and then there are also generic glasses out there, and there’s webcams. So you could probably find a pair of glasses on Amazon or, you know, Alibaba’s store that are like 50 bucks that just have a camera that take a, you know, one frame a second. You don’t need to be posting some 4K of me skiing behind me. You just need like what, like every, you know, you just need like four shots a second or something to know what’s going on in the world and to process this. This is incredibly powerful. And you could think about like somebody working in a store doing inventory and it’s just looking at the, just talking out loud and just being like, ‘Okay, looks like we have, you know, people have been buying a lot of milk, we’re running low on milk,’ and you just look at it. And then it sends to the milk purveyor how much milk we have left, it looks at the dates and the expiration dates and then just puts an order in. Like imagine the person at Costco running around. Just person runs around Costco. One person could probably walk through the store and without even doing any work could just talk out loud, stream this, and manage the store. You could go from probably 10 managers at a store… down to one. And that’s what people I think are missing here when we talk about Open Claw in the year of our Lord AI 17 days. Presh, what are your thoughts here when you see Sean’s very cool hacked together proof of concept?

Presh Dineshkumar: 27:16 Yeah, I think it’s brilliant. And I think, like, we’ll see more. The Meta iteration cycles on the glasses have been pretty impressive. I know the—I don’t know, Sean, if that’s the one you’re using has the display screen as well?

Sean Liu: 27:28 That one doesn’t have the display screen.

Presh Dineshkumar: 27:30 but I imagine when you—when you have a display screen that’s also interfacing with, you know, maybe it’s sending it to your computer but you’re seeing action items on your—on your glass so you know exactly what it’s doing versus just obviously how it’s interacting right now just through text or mirroring on your phone. I think that gets really interesting. A question for you Vishnu, what personally as you’ve been exploring this technology, what’s the most interesting use case you’ve used personally with the glasses?

Vishnu: 27:59 Personally, I think, um, it’s about just add all kinds of stuff that I think I need to buy more in my fridge. So I just say this, this, this, this, and let it to just directly add all those bunch of things into the shopping list dot txt. And I use that txt to say, oh, just add them to the cart and then I just bought it.

Presh Dineshkumar: 28:24 And then you buy and you use the Amazon to add it to your cart or InstaCart or whatever other service?

Vishnu: 28:29 Just Amazon. The Amazon browser automation. Yeah.

Sean Liu: 28:32 But not me personally, but I think the coolest example I see is yesterday a perfumer reached out to me to say it’s really unlocks the, uh, huge, um, unlock to their experience of, uh, making the perfume. For example, so before that, um, he was integrating with, uh, their Open Claw to just—they have to text to it. They have to use the Mac or the phone, um, like Telegram to say, oh, I have this, um, chemistry and I add this much. Um, but after this, um, he tried this, um, by my repo and he can just directly say, oh, hey Gemini, uh, what is this? Uh, help me record this. And then he has the agent, uh, skill, uh, which is record everything into the Airtable. So basically he organized all kinds of, um, perfume formula in that Airtable database. So basically the—how he makes perfume right now is just directly use the Gemini to record this and delegate the agent with the skill to record into the Airtable.

Jason Calacanis: 31:03 This makes you into the Terminator. It makes you into RoboCop. And I was thinking about this. What a great idea it might be, Vishnu, Presh, Sean, to take this version and to give it to ICE agents, who are in the field trying to identify people. And I don’t mean to make light of the situation, but, you know, some of these folks who are out there doing stuff for ICE have been known to be less trained and maybe not good at threat assessment. If this thing was giving you a real-time threat assessment and was like, ‘This person is acting in a peaceful manner. This person is acting in a threatening manner,’ and it was advising you what to do or things to look out for. It would pick up on, ‘Hey, this car’s about to run into you. This person has a gun,’ etc., long before a human could actually respond to it. And if it could also be hitting the Palantir database and say, ‘Hey, this person is somebody who’s at their 17th ICE, you know, demonstration. They, uh, actually attacked an ICE officer previously. There’s a warrant out for them. This person is a teacher and they’re 72 years old and they’re not a threat. You know, best advice is to de-escalate, de-escalate, de-escalate.’ This actually would be super helpful in real time to assess these things. Okay, we’ve got a lot of our Noties. What are Noties? Noties are the loyal Twist live listeners. We do the show live sometimes. So if you go to youtube.com, subscribe to This Week in Startups, click the live button, get alerts, put the bell on—there’s a little bell there—you’ll get an alert when we go live and then you can ask questions. That’s called the Notie Gang because they have notifications turned on. And a member of the Notie Gang, Reed Adam, says, ‘Do you think this will drive seed rounds down in total cap raised and equity purchased where valuations stay the same, but founder gives up small percentage of equity, or do you think this will shift market to smaller cap raises at smaller valuations?’ Valuations, this is a great question from Reed F Adams. Presh, you and I have seen this over 30 years of me being in the industry and 10 you being in the industry, we’ve seen a seed round, a series A, you know, go from low millions of dollars down to hundreds of thousands of dollars to people just going to an incubator, taking the 125k and getting a product to market or just getting a product to market very quickly. So what I think happens is the founders are already keeping more of their company. They could skip venture capital completely or and they could just bootstrap. And the longer you bootstrap, the less equity you give. We have a term, not a unicorn, not a Pegasus, an Allicorn. A unicorn that has wings like a Pegasus. That Allicorn we had a couple of them like Calm where they were making so much money they would use their revenue to fund their next round and skip a round of funding. Each round of funding typically 10 to 15% dilution. In the old days it would be like 20 to 30% dilution each round. That’s when founders wound up with low single digits or mid single digits ownership at the end. Now founders get to like 20% ownership by the end, sometimes more. That is what we’re going to see more of. My best advice to founders is to skip rounds. What do you think, Presh?

Presh Dineshkumar: 34:13 I think we saw the early cycle of that with Calm, and that was like pre-LLMs, pre-AI, where the money is useful is pre-seed, where you have the idea and then you’re going full-time on this thing and you’re building out the product, you maybe need some resources just for like—

Jason Calacanis: 34:31 Yeah, to have three people quit their jobs. Three people quitting their jobs is 30k a month, it’s 360k a year. People take a 10k draw. There’s enough people in the world who can live off 100k a year, 75k a year, 150k a year, you’ll be fine as a founder. Vishnu, welcome to the program. I don’t have your last name here. Vishnu, tell me your last name and tell me, when did you find out about Open-claw and are you in fact claw-shotted?

Vishnu: 34:55 Reddymasu. So hey guys, I’m Vishnu, just an engineer. I think I heard about Open-claw two weeks ago and it kept coming on my Twitter and I was like, okay, by the end of last week I was like, okay, I gotta give this a shot. And initially, I actually tried this on my laptop and I was like, okay, let’s see what it can do. And the next day I opened and I saw eight Open-claw instances had given it some tasks and it’s running CPU 100%. I was like, okay, this is not good. At which point I decided to take it off and try to find a virtual machine to try and see can I host this somewhere else? You know, that’s where I feel more safer and you know, all my keys and you know, everything is not exposed because technically it has full access, you know, it’s claw—

Jason Calacanis: 35:35 It’s basically rooted your machine. So running it on your desktop where you have your Coinbase account or your Robinhood account or your password manager is a very bad idea. So you decided, let me put it in a virtual machine in the cloud so it’s sandboxed and can’t go too wild.

Vishnu: 35:56 Exactly. Yeah. But that was an interesting journey because I started off with, okay, let’s look at AWS or Google Cloud provider and I started—

Presh Dineshkumar: 36:00 I didn’t want to spend too much. Typically, you know, there’s the free instances that, you know, all these big giants provide. I think in GCP, it is e2-micro, but turns out it doesn’t work very well with those small instances because it needs a high enough RAM. I tried e2-micro, it was small, and I think at the end, it was e2-medium where I finally got it working at around four gigabytes of RAM. But that was already $25 a month if you actually run it through and through. And again, I wasn’t super psyched about something that I don’t know well enough that, you know, is it actually going to be, you know, useful enough for me while I already have, you know, other subscriptions? So I kind of went online and hunting for, okay, where can I find, you know, cheap enough host? I explored a few different things. I think there was Hetzner, there was a bunch of others. One of them, I think the cheapest I found was around seven, eight dollars. So anyway, I set it up and then I think two days later is when I had my ‘aha’ moment. I was working out in the gym and I was texting with it all the time and it just did certain things that I did not expect it to do.

Jason Calacanis: 37:05 What was the example that kind of one shotted you?

Presh Dineshkumar: 37:08 I was tired of typing it all the time into it and I wanted it to be able to just hear my voice and do it. So, you know, because you’re working out in the gym, you don’t want to go back.

Jason Calacanis: 37:21 No, I’ve been there. I’ve been there. Yeah.

Presh Dineshkumar: 37:23 It told me, ‘Hey, I don’t have a key. Do you have a Deepgram key? Blah, blah, blah.’ So fortunately, I did have one and so I provided it and it just did all the setup itself. It set it up and now it was able to answer.

Jason Calacanis: 37:33 It basically created its own software to talk to you. And that is one of the things that I think moves people from, you know, being curious about the software to being all in, is when it doesn’t have a capability and it goes, ‘You know what? If you want to talk to me, let me research on the web. Okay, there’s seven different ways to do this. These seem to be the two best. This one didn’t work. This one did. Okay, we’re up and running.’ That is a mind-blowing moment. So you take all this and then you decide you’re going to do what? You’re going to make your own hosted version for normies?

Presh Dineshkumar: 38:01 Actually, no, that wasn’t even the next thing. Next thing was, I was like, holy shit, this is so cool. I texted all my friends, like, guys, you got to try this. You know, it was like, I guess the vitality how open-core comes up. And they were all, I think over the next day, they tried it, but then it’s like, yeah, the setup is hard and etc, etc. They were like, so I just said, you know what? I have a VM and it’s probably, so a lot of times this is idle, you know, it doesn’t necessarily is using all of the CPU or RAM. I was like, I’m just going to host it for you guys and I’ll expose like a container and give you this. So that’s how it started and they started using it and they loved it. And I thought, okay, let me put it online and see if there’s other people. The cost of actually running this, if you do it correctly and you do shared resources, I think right now it’s about a dollar and 25 cents. You could probably push it to 99 cents pretty much.

Jason Calacanis: 38:54 Per day, per month?

Presh Dineshkumar: 38:55 Per month. Just to have it running. Now, you also have the LLM bill. So you created agent37.com and you can run your own instance for 3.99 a month or 10 bucks a month. And now, is this a good business to be in or is there a bigger vision you have? Because hosting Open Claw, that seems like a commodity. Open Claw’s obviously going to have their own hosted version. There’s… it’s kind of like WordPress hosting. I don’t know if it’s a great business, it’s a good business. There’s two or three people doing it who have built multi-billion dollar businesses, so I don’t think it’s a bad business necessarily. But is that where you’re going to go with this or do you have a bigger idea, Vishnu?

Vishnu: 39:34 I mean, I think it’s a bit early. Like, really it just started off as ‘here’s an experiment, guys, go try this’. And I was surprised over the last three days, the… it just kept growing, you know? Like, I initially it was like 99 cents, I bumped it up to four bucks or 3.99. And I think at this point I have around 71 and it’s still growing, users who are all actively using it. So, yeah, it was just totally unexpected. In terms of long-term, I mean, we’ll see where it goes. I mean, I’m committed to seeing it through given, you know, how rapidly it is growing.

Jason Calacanis: 40:07 Are you… are you an entrepreneur? Are you a consultant? What do you do in your day job?

Vishnu: 40:15 Ah, no, just an engineer. I have some time off to take care of some things and this just happened to be a moment where I was like, okay, let me try this.

Jason Calacanis: 40:21 So if I were to give you 25,000 dollars, be your first investor, have you come to my accelerator where I give you 125k, would you take that deal, have Jason Calacanis as your first investor, and then you come on the pod every month and we talk about your progress? How does that sound? That might be something you’re interested in, Vishnu. You say yes?

Vishnu: 40:36 Okay, yeah let’s do it actually. I didn’t expect this. Okay, wow, okay, sure. Let’s do it.

Jason Calacanis: 40:41 Okay, this is how we do it. You want the 25k, you want the 125k?

Vishnu: 40:45 Let’s do the 125k.

Jason Calacanis: 40:46 Okay, so that’s 125k, 7%. It’s the standard deal you get at Y Combinator or Techstars. Standard deal, you come to the accelerator. I’m just going to… since you showed the initiative, I’m just… this is my diligence, okay? So as long as we have to do a little background check, you’re not a criminal, you’re not on the lam, you haven’t kidnapped anybody. We’re going to assume that you did this work and it’s not stolen or there’s not IP theft, but we’ll go through a quick diligence, you come into the accelerator. That’s great. Sean, what’s your story? Who are you? Are you an engineer working for Meta? What’s the story? You’re an entrepreneur?

Sean Liu: 41:19 Yeah, currently… just after this, we… we have a team right now, so… it’s basically doing this for traditional industries.

Jason Calacanis: 41:30 Have you raised money yet?

Sean Liu: 41:32 I haven’t.

Jason Calacanis: 41:33 Okay, so I’m going to give you the same offer. 125k, you come to the Launch Accelerator, as long as you can prove that this is your IP and you have not committed any crime. I don’t think you have. You seem like an upstanding young gentleman. Or 25k, you come to Founder University, two and a half percent, and we just have a light relationship, but you get to say, ‘Hey, I got J-Cal from This Week in Startups and All-In as my first investor’. Is that something you might be interested in, Sean?

Sean Liu: 41:54 Yeah, let’s do 125k.

Jason Calacanis: 41:56 Okay, so I just spent 250k. You’re both in the accelerator. This is how we do it. I’m an instinct investor, these guys are great. I want to work with you guys for the next 12 weeks and then we’re going to go raise a monster seed round and we’re just going to blow the shit up. That’s it folks, Sean, Vishnu, 250 dimes. I’m just spending money on the air. My team, Lucas, is going to get in touch with you after this and Jacob, and they’re going to get you through the light, very light diligence process. You guys aren’t incorporated I take it? If you aren’t incorporated, we’ll get you through incorporation, set up your cap table, show you how to do all that good stuff, and we will be in business together starting Monday morning. Let’s get this done. It’s Wednesday. Let’s have this done by Friday. We can get working on the weekend. Thanks for coming on the program. I appreciate you both. All right everybody, that’s another episode of this week in startups in the can and we will see you on Friday with Lon Harris as my co-host and we always do Off Duty. Off Duty at the end of the Friday show where you get a couple of great, great tips for media, movies, TV shows, books, video games, clothes, fashion, food, music, any of those things. We’ll see you on Friday. Bye-bye.