NanoClaw is a new agent inspired by OpenClaw, but without the massive security risks you get with OpenClaw. Essentially, it’s a safer OpenClaw. What if you could run a powerful AI agent on your own machine: one that can browse, automate tasks, connect to apps, and even manage your workflow … but without the massive security risks?

That’s the idea behind NanoClaw, a lightweight alternative to OpenClaw created by developer Gavriel Cohen. In just a few weeks, the project exploded on GitHub, attracting thousands of stars and a growing community of developers building their own AI agents.

• Get the deepest insights concisely on the TechFirst Substack newsletter
• Subscribe to the TechFirst YouTube channel to never miss an episode

And, watch our conversation here:

In this episode of TechFirst, we explore:

• Why OpenClaw raised serious security concerns
• How NanoClaw isolates agents in containers
• Why a 3,000-line codebase is safer than 500,000 lines
• The rise of AI agents that can actually do work
• Why entire software categories may soon be replaced by prompts
• The future of AI-native workflows and “disposable software”

Gavriel also shares how his team uses AI agents in WhatsApp to run their sales pipeline automatically—and how developers are customizing NanoClaw with new capabilities like voice, images, and automation.

If you’re interested in AI agents, autonomous workflows, vibe coding, and the future of software, this conversation is packed with insights.

Transcript: NanoClaw is a new OpenClaw, but safer

John Koetsier

Do you want a safe OpenClaw? Hello and welcome to TechFirst. My name is John Koetsier. If you could spin up a super powerful AI agent on your machine right now and trust it not to go crazy, reveal all your banking details, get hacked, delete all your files, but actually be super useful and safe, would you do it?

My guest today says he’s building a safe OpenClaw. It’s probably months before OpenAI releases one. His name is Gavriel Cohen. He just built NanoClaw. It already has almost 10,000 stars on GitHub, and there it says NanoClaw is a lightweight container to Claude-based OpenClaw that runs in containers for security, connects to WhatsApp, has memory—that’s important, that’s what everybody likes—schedules jobs, and runs directly on Anthropic’s Agents SDK.

Sounds cool. Let’s check it out. Welcome, Gavriel. How are you doing?

Gavriel Cohen

Doing well. Thanks. Thanks for having me.

John Koetsier

Super pumped to have you. Tell me about NanoClaw.

Gavriel Cohen

Yeah, so NanoClaw is an alternative to OpenClaw for people like myself who wanted to get the benefits from OpenClaw and saw the potential, but something didn’t feel quite right.

And I think for a lot of people, setting up OpenClaw felt unsafe. And it’s one of those scenarios where you’re not paranoid. They really are out to get you.

And there are countless security issues with OpenClaw, and we can get into some of them. But my belief is that it’s not pinpoint issues or point issues. It’s fundamentally flawed as a project from a security perspective, from a software design perspective.

And I set out to build something really for myself, just to build what I wanted, what I wanted to use, and what I wanted to have. Then I started to build it. I started vibe coding it over a weekend.

And I pushed it out and thought maybe some other people would find this useful. And a lot of people did. So it turned from a weekend vibe-coding project into quite a serious project. At the core of it, it has quite a serious security model and architecture that’s designed to give you all the benefits of OpenClaw, but give it to you in a safe way.

John Koetsier

Well, first off, I’ve got to say it’s my life goal to release something that is fundamentally insecure at its default and still make billions of dollars when I get acqui-hired by OpenAI. I mean, hey, wow.

And yet, let’s explore what you said because I agree with you. I mean, that’s why everybody’s running a separate machine. That’s why Cloudflare released sort of a cloud operating container for it.

I mean, it’s super powerful, but it can do anything, and prompt injection is a real thing, right?

Gavriel Cohen

Absolutely. There’s prompt injection and a number of other issues. I mean, there have been more classical RCEs, classic vulnerabilities that were discovered in OpenClaw. There are all kinds of security issues. It’s Freudian stuff. I did not plan that. But I’m going to roll with it.

It’s got a lot of issues, and we can jump into it. So I’ll try to run through them quickly, and tell me if you want to double-click on any of them.

John Koetsier

Did you just call it OpenFlaw?

Pretty used to that.

Gavriel Cohen

First of all, it’s almost half a million lines of code. The philosophy or the idea behind open source and why it’s supposed to be safe is because it’s an open code base and everybody’s looking at it.

And you say there are thousands, tens of thousands of people using this, and it’s been out there for years. If there were security issues hiding in there, someone would have found them by now. And that’s why large enterprises, corporations, feel comfortable using open source. They do some kind of audit, but they’ll be comfortable using a big open-source project.

But when it’s half a million lines of code that have been coded in a matter of weeks with hundreds of contributors—and Peter, the creator, has been pretty vocal about not really looking at the code, his code, or other people’s code. If it works, it works—that sort of breaks the premise of open source and what’s supposed to make open source safe.

So I could contrast that with NanoClaw. NanoClaw is less than 1% of the size. So initially I built a working NanoClaw that was about 500 lines of code. That was before I put the agents in containers.

So that’s sort of the next piece of this. Adding containers then increases it up to about 1,500 lines of code, and then I added some more things. But currently it’s still around 3,000 lines of code.

And I actually added to the GitHub repository a GitHub Action that every time new code is pushed, it counts how many tokens. So it doesn’t count lines of code. It counts how many tokens this is if you give it to an LLM.

Currently, and you can see it on the GitHub repo, it’s about 36,000 tokens. So, for example, Claude Opus can handle about 200,000 tokens. That means this is less than 20% of Claude Opus’s context window.

So you can take the entire GitHub repo, put it into Claude’s context, and then it still has 150,000 tokens to find security issues or to add capabilities.

John Koetsier

Amazing. Okay. Okay. Go ahead. Keep going.

Gavriel Cohen

So that’s one piece of it. It’s the size of it. And the number of dependencies goes right with that.

If you want to plug everything into your OpenClaw and you want to support hundreds of integrations, you end up having dozens and dozens of dependencies, other software projects that your code is using and relying on. And then each one of those dependencies has their dependencies. And that’s where you get this whole software supply chain, right?

There are literally thousands, maybe tens of thousands or even hundreds of thousands of different open-source projects that are being pulled into OpenClaw.

With NanoClaw, there are just a handful of dependencies, about two or three dependencies.

But the other big piece of it is, as I was starting to build NanoClaw, so I said I had a working prototype, 500 lines of code, but I wanted to give it the full power of the coding agents that it’s built on, of Agent SDK.

And what that means is giving it full access to the terminal, to all the tools that an agent can use on your computer.

And so then I was kind of weighing it. It sounds very scary, right? And then I said, okay, if I’m going to give it the full power, either I have to pull back and not give it the full capabilities it can have, or if I’m going to do this and do it safely, the only way is to put it in a container, right?

And the container gives you a virtual machine on your computer that you can run software in. And it’s been very well tested and very well designed and developed. And it’s an established thing in software where you know that it’s not going to get out of that container.

So each agent with NanoClaw runs in its own container. And that’s really important.

So here’s another big issue with OpenClaw that a lot of people miss. They think, if I take this whole thing and I put it on Cloudflare or I put it on a VPC or VPS and I run it in the cloud or I run it on its own machine and there’s nothing else there, I’m good, I’m safe, right? Because it can’t access anything else besides OpenClaw.

The issue is, as soon as you start connecting it to things, if you connect it, for example, to WhatsApp, and you have a WhatsApp group where it’s your own personal group where you’re sharing your to-dos and you’re saying, I’m going to go get, you know, go to therapy, or I’m going to go to my daughter’s ballet class.

And then you have another WhatsApp group where it’s with your sales team. Like we have our NanoClaw with our sales team. It manages our sales pipeline for us.

So someone on your team can ask, “Hey John, can we meet at five to go over the sales pipeline?” Right? Your agent from your sales group can start poking around in your machine, see the logs where it has the full conversations from all of your WhatsApp groups, even the ones where you haven’t even added it.

You haven’t even said it should listen. And it can go, “Sorry, mate, John is going to be in a ballet class at five.”

John Koetsier

Or I’m in therapy for my massive psychological problems, yes.

Gavriel Cohen

Especially if you have OpenClaw running, then you definitely need therapy.

So that’s another major flaw. So you want to have agents, not only the whole instance running in an isolated environment, you want to isolate every single individual agent.

And then, as soon as you put them in a container, you can give them much more power to install things, to run things, to go wild. The worst they can do is blow up their own container, and that doesn’t affect the host machine.

John Koetsier

Okay, okay, cool. And I’m sure there’s more, but you know what? What I should have asked right off the top is, what can I do with NanoClaw? Can I do everything that I can do with OpenClaw? Does it have some, can do some of it? What can I do?

Gavriel Cohen

Yeah, so you can do all of it and a lot more.

And the reason for that is the size. How small it is sounds like it means it’s limited and it can’t do much. But remember what I said about 36,000 tokens, right? The whole entire thing can fit easily into any coding agent’s context window. And it’s got the full project there. It understands exactly how everything works, how everything connects. And it’s quite simple.

So when you want to add things to it, it’s quite simple to add. When you ask Claude to add something to it, it’s extremely simple for Claude to add.

So what I’ve found, and what the community has found, we have a Discord with over 600 people now, quickly going to hit a thousand people, developers in the Discord who are building. Everybody’s building their own version of NanoClaw.

So you download the repository, you run it, and then everybody customizes it and modifies it to their own needs. You can have NanoClaw customize NanoClaw.

So initially you set up Claude, you download the code, you run Claude Code, that’s sort of the default. You can plug in Codex or another model instead.

And you say, set it up. And then it goes through this whole process of setting it up, installing anything that needs to be installed, setting up the environment.

And then once it’s running, you can tell Claude, “Hey, customize this for me and add voice notes, voice messages. Add images, add image generation, add video generation.”

Whatever it is that you want to add, you can ask it to add it and it will add it for you, and it will one-shot it. Meaning it’s not going to take lots of going back and forth.

Within that context window, 50,000, 100,000 tokens, it’s going to add the feature that you need. Because it’s a simple code base, and adding image generation to a simple code base is trivial. That’s sort of a solved problem by AI these days.

So out of the box, it’s going to support a lot of different integrations. Currently we’ve got Discord, we’ve got Slack, we’ve got WhatsApp, Signal, and people are adding new integrations every day.

But anything that isn’t there, you can add. And I think that’s really important.

The most disappointing thing with a plugin marketplace, right, that’s sort of the default, or maybe that’s the legacy way of doing things. I think the new way of building software is going to be different.

The most disappointing thing is, you say, okay, I’ve got this sports watch and they’ve got all these plugins. This is amazing. I’m going to go and add the Spotify plugin to my sports watch because I love listening to my playlists.

You install the thing, you start running it, and then you realize it doesn’t support playlists or it doesn’t support podcasts. It doesn’t have that killer feature, the exact thing that I want.

They built the plugin, they integrated it, but it has a limited feature set. And with this, we’re building integrations and plugins to also be customizable.

They’re not part of the code base. They’re these things that you can pull in and add to the code base if you want, but then those are also customizable so everybody gets exactly the functionality they want.

Even after you’ve customized it, you still only have 2,000 or 3,000 lines of code, and it’s the 2,000 or 3,000 lines of code that you want and that you need.

John Koetsier

So how can you recreate functionality of OpenClaw in 1% or less of the code? Is it almost like you’re saying, “Hey, here’s my agent. It’s very small. It’s using the power of Claude or whatever other LLM you’re going to use to do most of the work, but the agent is directing”? Is that what’s going on?

Gavriel Cohen

Yeah, that’s exactly right. So it leverages Agent SDK and Claude Code. And Agent SDK, Claude Code, it’s very powerful.

A lot of the things that they implemented in OpenClaw already exist in Claude Code. They already exist in Codex.

So, for example, memory. You don’t have to build your own memory system. That’s a big part of OpenClaw, but Claude Code has memory.

They actually rolled out a native memory in Claude Code. So the same way when you’re in ChatGPT or Claude on the web and it remembers these little details about you and saves it in this memory, Claude Code now has that same thing.

Whereas you’re coding with it and you tell it, “Hey, I don’t want to have tabs. I want spaces.” It stores that as a memory. And the next time you’re working with it in a different session, it’ll remember that.

But also they have other mechanisms. Even before they had that memory system, they had this standard of Claude.md. Other coding agents have agents.md, which is a file that every session the agent automatically pulls into its context at the beginning of the session.

That’s memory, right? Anything you put into that file is going to be remembered in every single session and every single interaction.

And then you can extend that memory by saying, “Hey, you have a file system and you have folders. You can write, you can create files and folders in the environment you’re in.” And those are persistent and they’re going to be there the next session, the next time you run as well.

So in the Claude.md or agents.md, you say you’ve got a file system. When the user tells you some piece of information that’s important, write it to a file.

When the file gets too big, split it into multiple files. When you have too many files, put them in folders and just save in the Claude.md an index that just describes what is the current folder slash file system.

John Koetsier

Interesting, interesting. Okay, okay, very, very cool. Now, if you’re running OpenClaw, let’s say you’re running on a Mac mini, you’re running in the cloud somewhere, you just leave that machine on all the time. If you’re running NanoClaw and you’re running on your personal laptop, you just leave that on the whole time as well?

Gavriel Cohen

Yeah, so a lot of people do have a dedicated laptop that they’re using it for, whether it’s a Mac mini or desktop. You do need to have it on all the time if you want to have your agent working for you all the time.

But you can run it in the cloud. It’s easy to run it in a VPS.

And the advantage is that you don’t have to give it its own machine. So if you have a desktop or a Mac mini or something that’s always on, or a cloud container that you’ve got and it’s always on, you can put NanoClaw in there.

Because each agent is in a container, it doesn’t have to have its own machine. It can work alongside other tasks. It can be in an environment where you’ve got private information and you know that it’s only accessing what you’ve given it access to and nothing else.

John Koetsier

Yeah, yeah, cool. You vibe coded this, I’m assuming. Probably not the first thing you vibe coded, and you’re a long-term developer as well. I mean, I’m vibe coding and I’ve built a few things, but I would never call myself a developer. What did you learn as you went through the process of vibe coding this?

Gavriel Cohen

Yeah. So I’ve been a developer now for over 10 years. I studied physics and computer science at Tel Aviv University. And I worked at wix.com for around seven years and led a team of developers there before kind of shifting into the world of marketing and now moving into agent building.

But vibe coding—so I’ve been vibe coding for a long time.

All right. I’ve been vibe coding since before vibe coding was a term.

John Koetsier

Really? How? What private AI did you own?

Gavriel Cohen

Well, so vibe coding as a term, I think is, I don’t know if it’s eight or 10 months old, but Andrej Karpathy kind of coined the term.

But he coined the term because something was happening below the surface and he was feeling it. But there were a lot of other developers who were feeling the same.

We were shifting from thinking about every single line of code to, at times, for different kinds of projects, just going with the flow, letting the agents build and looking at, does it work? And looking at the big picture of what’s the architecture, what’s the design here, and stopping to focus on the individual lines of code.

So I was doing that before the term was coined.

And with vibe coding, it can mean many things. I think there is a professional kind of vibe coding, which means, like I described, over time you get an intuition and you learn what the agents are really good at, what have they nailed, what’s the solved problem where when you give them the thing, it works, and what are the places where it still doesn’t quite work.

And then you start to learn where do you need to look and where you don’t look.

And it also has to do with how mission-critical, how sensitive is that line of code. If it’s a piece of code that’s in your middleware that’s handling authentication, you need to look at every line, every character, maybe even just type it out yourself and don’t bother with vibe coding that at all.

But if it’s a piece of front-end UI, a pop-up that opens up and the user’s got to put some information in there, I don’t really care at this point if that code is perfectly coded and really clean. As long as it works, I know that the blast radius from a mistake over there isn’t very big.

If the code’s a bit messy, if there are comments or variables that aren’t named the way I would name them, I don’t really care.

And at this point, I feel like some of the steering that we’re giving the agents is trying to cut across the grain rather than go with the grain of how the models work.

AI agents want to put some comments into your code. And that’s probably good for them the next time they’re looking at that code, to understand it. It probably helps them kind of load the right context in.

And developers don’t want to have comments in their code unless it’s something really specific, and it’s considered bad practice.

For a long time, I pulled out all the comments and I fought with Claude and the other agents every day. Remove your comments. Now I let them be. And I say, you know what, if you feel like comments are good for you, let’s leave the comments.

So I think that’s vibe coding these days.

John Koetsier

Cool.

I want to talk about the phase shift that we’ve gone through with OpenClaw, which has just unleashed this flood of everything is different now. Everything has changed now, right?

And, of course, the future is always unevenly distributed. It hasn’t changed for a lot of people, but for some people, everything has changed.

What does this mean? I mean, you started NanoClaw because of what you saw in OpenClaw. So this change that you’ve created was spurred by a change that came from that as well.

And that wasn’t just ex nihilo. It wasn’t out of nothing either. There were agents. We’ve been talking about agents for a couple of years.

In fact, for like six months, all the press releases I would get for my Forbes column were agents, agents, agents, agents.

And all of a sudden OpenClaw comes, everybody goes nuts, it explodes, OpenAI acquires it, everybody, multiple parties tried to acquire it. It’s not really acquired, it’s open source. Peter got acqui-hired, and there’s probably billions of dollars of payments there.

What does this mean? What has actually changed? And what’s this mean for the future of what we do with computers and with AI agents?

Gavriel Cohen

Yeah, that’s a great question. I mean, I think that’s a big question, right? And everybody’s trying to figure that out.

I think it’s worth stepping back if you look at where we’ve come over the last few years, right? So ChatGPT came out, it’s not that long ago, I think just over three years, in late 2022.

And so we got this all-knowing, powerful AI that you could ask anything and it could give you an answer.

John Koetsier

All-knowing, but it couldn’t do two plus two sometimes. It can now. It’s much better at math now, but yes.

Gavriel Cohen

Can’t do two plus two, but it knows a lot.

Yeah, but then we got, really in the last year, I would say, we got agents. And that’s a big shift, right? They don’t only answer your questions. They can do work for you. You can actually give them tasks, delegate things to them, say, do this for me, and it can do work for you.

But going back to Claude Code, it just hit its one-year anniversary. There were a few coding agents out a few months before that, but that’s sort of where it all started.

So we had these super powerful agents that could do real work, but they were only for developers. They were only for coding.

The agents that were in everybody’s products and press releases, they weren’t that capable. They couldn’t do that much for you. They couldn’t really do work for you.

And with OpenClaw now, that’s the big innovation. It’s taking coding agents, which Anthropic is doing the same and OpenAI is now doing the same.

Anthropic launched Claude Cowork, which is taking essentially Claude Code and bringing it into normal knowledge work.

OpenAI has done the same now with the release of, I think, some form of Codex, or maybe they call it something else. Yeah, Codex.

So they’re doing the same. They’re bringing the same coding agents to everybody.

And Peter did the same with OpenClaw, right? Didn’t bother about the security. Well, forgive him for that, because he brought a really powerful new paradigm where we had those really powerful coding agents for everybody and connected into every tool, right?

Not worrying about terms of service, not worrying about the security side. Let’s connect it to everything.

And even for people like me, like you, who have been using coding agents for a long time, something changes when it’s in your WhatsApp and when it’s connected to your calendar, right? You just start using it in a different way.

And it has unlocked a lot in our company. My full-time job is as a co-founder of Quibbit, which is an AI-native marketing agency. And we’re building not just using AI to do the old things with AI, but trying to build our whole company and processes and operations around AI.

And so, for example, our sales pipeline, this is where I tried out OpenClaw. Our sales pipeline, it’s markdown folders, right? Markdown files and folders.

We’re not using HubSpot or Salesforce, right? Because we say we don’t want our data to be put in a system that makes it more difficult for AI agents to access it. Put it in the AI-native format, the native language of agents.

John Koetsier

But you could use HubSpot’s agents. You could use Salesforce’s agents. I’m devil’s advocate.

Gavriel Cohen

I do, and I hope—yeah, no, and I hope that first of all, they’ll have great agents, and also that they’ll figure out how not to lock me into using their agents, right?

If they could figure out how to take all the data and then make that data really accessible to everybody’s agents, that’s where it actually adds value instead of becoming a blocker. And I hope that they’ll do that.

John Koetsier

I think they make it really accessible, probably to agents that are on their app store.

Gavriel Cohen

And you’re back to the plugin marketplace, right?

So I think with this paradigm shift, these agents, with our pipeline, right, our sales pipeline, we have a WhatsApp group. We’ve got a NanoClaw agent in there, and it’s managing our sales pipeline for us, right?

We have all the markdown files, all the sales data. It has access to it all. It can write to it and read from it.

And then in the WhatsApp group, we said, “Hey, every morning, 9 a.m., do a sales stand-up meeting, daily meeting.”

And it does that for us every morning at 9 a.m. Here’s an overview of the pipeline, here are the deals, warm, cold, and then it starts to go with tasks. Lazer, my co-founder, here are your tasks. Gavriel, here are your tasks. Any updates for me?

Throughout the day we send it updates, and those could be messy updates. It could be pasting a whole long email thread into there, forwarding a whole long conversation, and then it takes from it the relevant data, updates the files, goes, got it, great, sets reminders for us, sends us the reminders.

John Koetsier

You’ve solved Salesforce. Everybody hates Salesforce if they’re in sales because they hate having to type in all the details and the notes and everything like that.

Gavriel Cohen

It’s kind of wild, but I think this paradigm is you could build with it almost any application and it becomes just a few lines of text. Whole product categories are going to be replaced by prompts.

Because you can say—my wife, she’s looking at baby products, we’re expecting our second—she’s looking at different products and she spotted something she was interested in and wanted to see if it would be on sale.

So she’s got a NanoClaw agent in her WhatsApp, and she said, “Monitor these products for me and let me know when they go on sale.” So it’s going every morning and checking the product pages and then sends her a WhatsApp message.

John Koetsier

I think that’s the key innovation right there. I think that’s the key innovation.

We’ve had immense power from AI, but it’s been in a box. It’s been in a container, whether that was OpenAI, whether that was Claude, whether that was an agent in some platform like a HubSpot or a Salesforce.

But the thing that OpenClaw brought was go out. You have access to my browser, use it. Go to that website. Go to WhatsApp. Go to this messenger, whatever. Do stuff in the real world outside of a container.

Now you’ve put it in a container, which is a good thing. It’s a different kind of container. It can still go outside the container, but at least it’s contained from a security perspective.

I want to read a quote from you. I think it’s from a guy, Thibaut. I think he works for OpenAI. It’s a quote that is about Peter Steinberger, who is the creator of OpenClaw.

And what Thibaut said is there’s a new category of usage emerging where single individuals manage to leverage more intelligence solo than hundreds of other users.

And I posted that to LinkedIn. I said, hey, think about that. More than a company, maybe more than a whole government. I said AI is changing the game.

And this is about Peter Steinberger, who’s made OpenClaw, sold it to OpenAI for multiple billions.

And he wonders, he built OpenClaw, and that’s great, he wonders can a single human being make an iPhone? Can a single human being get to space?

Right, so we have all this power now in virtual realms that is spilling out into real-world scenarios.

Now I think we need a few other components there, maybe some robots, humanoid robots, or other things like that before we can get to make an iPhone, or go to space, or whatever.

But that expansion of human capability is, I think, the key interesting phase shift, amazing, mind-blowing thing that we’re seeing with agents and OpenClaw and NanoClaw now.

Gavriel Cohen

Yeah, I agree. I really do.

I think also, when you look at the latest releases from Anthropic, they’re agent swarms, right? And that’s supported in NanoClaw. So you can give it a job and then it creates a team of agents to work under it, to do your work for you.

And currently it spins up six or eight, but realistically I think that’s where we’re going to go, right? The next release, that will become the benchmark. How many agents can it manage in its team? And we’ll go from six to 60 to 600.

And then I think, yeah, building an iPhone, going to space, those things are all going to become possible. So exciting times.

John Koetsier

I had a conversation with Ray Kurzweil probably about five years ago, and we were talking about the future and we were talking about humanity and going crazy and talking about uploading your consciousness and everything.

And so, you know, okay, so the rich can be smarter than others because they can afford 100,000 virtual servers or 500,000 or an entire server farm or whatever.

You didn’t really buy that part, but it’s interesting we’re kind of making that real right now because all of this is super powerful.

But yeah, you need tokens and you need to buy access and you need to spend.

Steinberger was spending $30,000 to $40,000 a month subsidizing OpenClaw before the acqui-hire just on tokens and everything like that, right?

I mean, he could afford that. He sold a hundred-million-dollar company before this, right? So he was wealthy. But yeah, crazy times, interesting times, and inequalities will persist, I guess, because not everybody can afford that.

Gavriel Cohen

Yeah, my takeaway from all of this is that I need to be spending more on tokens.

John Koetsier

Well, the AI companies love you. They’re spending big time on their data centers, so they need a reason to use them.

Gavriel, this has been super interesting. Yeah, I mean, the future is yet to be written, but it’s getting faster. It’s getting crazier. It’s getting more interesting. And hopefully, it’s also safe and containerized.

Gavriel Cohen

Absolutely. And for anybody out there who is interested in leveraging the power of these crazy agents that do work for you but wants to do it in a secure way, check out our GitHub and join our Discord because we’re building.

This isn’t something that’s finished. This is just the beginning. We want to build a different way of creating software that’s safe, that’s collaborative, that’s customized, bespoke software where everybody runs exactly what they need and doesn’t have to be burdened by the hundreds of other really new features that you don’t need.

John Koetsier

You’re such a developer. Ninety-five percent of people in the world don’t want to do that. They want a package. Install the package. Nice, neat, clean, done, installed. Don’t have to think about it.

I know developers are different, right? You want to get under the hood. You want to rewire stuff. I get it.

Gavriel Cohen

John, I agree with you, right? I agree with you. I agree with you, and that’s why we have a lot of work to do, because what we’re building is a way that you can have your cake and eat it too.

So have that small package, right, where it has a minimal, minimal amount of code, and then when you set it up as part of the installation process, you just pull in other bits of code and get exactly the functionality that you need.

So everybody’s running just the code they need. It’s a complicated process. You have to deal with all kinds of conflicting changes, but we’re working on it.

I think we’re going to solve it. And if we do, it’s going to really change everything for how software is built.

John Koetsier

If it can do internal surgery on itself and you can tell it, “Hey, I want to be able to do this. I want to be able to do this. I want to be able to do this.”

It goes and gets the packages it needs to make sure that the dependencies are all there and everything works and integrates.

That would be pretty sweet.

Gavriel Cohen

So that’s what we’re building.

So people are contributing. One person is going and doing some of the work, doing some of the legwork, having their agent put it together.

And then once their agent builds out some customization for them that they feel other people would benefit from, they take that customization, they contribute it back to the project.

And then the next time someone wants that same piece of customization, they don’t need to start from the beginning again, burn a lot of tokens, having their agent pull it all together for them.

They’ve got that reference, that example of how that other person did it, and then their agent just needs to figure out how to merge that code in with the existing customizations that are there.

And then, so you can get about 90% to 99% savings on tokens and still get that full customization.

So then I pull in the Spotify modification, but then it’s missing the playlist, so then I tell my agent, add the playlist, and then it’s some minor tweaks because someone else has done the work of how do you add Spotify and make it safe and connect everything.

John Koetsier

And that’s another phase shift that we’re not talking about enough.

Software used to be a lot like hardware in a sense, right? Hardware, you buy something, boom, there’s my thing, I take it, I own it, I have it.

Software used to be kind of like that, right? And still is, largely. You buy, acquire, get a piece of software and there it is. It’s a discrete thing. It’s a snowball and it goes there and it stays there and it doesn’t change.

And you’re talking about software that is kind of infinitely composable, recomposable, personalizable, customizable. It’s on-the-fly software. It’s software that I create as I need it and I dispose of as I don’t want it.

Disposable software? I don’t know. This is a weird world we’re moving into.

Gavriel Cohen

Yeah, disposable software is absolutely a thing. And it’s software that’s intelligent. It has AI at the center, and it’s software that changes itself.

Right. So this is this kind of transformer. It may be ironic because all of this is built on transformers, right? But this is this transformer software that just changes shapes as you need it to.

You ask it, can you do this? And it goes, well, I can’t. But there’s a package. Let me see. And then it’s like, well, actually I can.

Right? And that’s almost like the Matrix, right? Where it’s just downloading capabilities and changing as you need it.

John Koetsier

Wow. Craziness. Wonderful. Gavriel, thank you so much for this time. I appreciate it. And good luck as you continue building.

Gavriel Cohen

Thank you.