Conversation with OpenClaw founder: AI is a leverage, not a replacement; 80% of apps will be replaced

Author: Baoyu

This is another 40-minute interview with ClawdBot/OpenClaw author Peter Steinberger, hosted by Peter Yang.

Peter is the founder of PSPDFKit, with nearly 20 years of iOS development experience. After the company received a strategic investment of 100 million euros from Insight Partners in 2021, he chose to “retire.” Now, his developed Clawdbot (now renamed OpenClaw) is exploding in popularity. Clawbot is an AI assistant that can chat with you via WhatsApp, Telegram, iMessage, connected to various applications on your computer.

Peter describes Clawbot like this:

“It’s like a friend living in your computer, a bit weird, but incredibly smart.

In this interview, he shares many interesting insights: why complex agent orchestration systems are “slop generators,” why “letting AI run 24 hours” is a vanity metric, and why programming languages are no longer important.

One-hour prototype, 300,000 lines of code

Peter Yang asks him what exactly is Clawbot and why the logo is a lobster.

Peter Steinberger didn’t directly answer the lobster question but told a story. “After coming back from retirement, I threw myself into vibe coding—using AI agents to help write code. The problem was, these agents might run for half an hour or just two minutes before stopping to ask questions. When you come back from eating, you find it’s already stuck, which is frustrating.

He wanted something that could check his computer status anytime on his phone. But he didn’t do it himself because he thought it was obvious—big companies would do it.

“By November last year, no one had done it, so I thought, fine, I’ll do it myself.

The initial version was extremely simple: connect WhatsApp to Claude Code. Send a message, it calls AI, and sends back the result. It took an hour to build.

Then it “came alive.” Now Clawbot has about 300,000 lines of code, supporting almost all mainstream messaging platforms.

“I think this is the future. Everyone will have a super-powerful AI that follows them through their entire life.

He says, “Once you give AI access to your computer, it can basically do anything you can do.”

That morning in Morocco

Peter Yang says, now you don’t need to sit in front of your computer and stare at it; just give it commands.

Peter Steinberger nods but wants to talk about something else.

Once, he was in Morocco celebrating a friend’s birthday and found himself constantly using Clawbot. Asking for directions, finding restaurant recommendations—these are small things. What really surprised him was that morning: someone tweeted that a certain open-source library of his had a bug.

“I just took a screenshot of the tweet and sent it to WhatsApp.

The AI understood the tweet content, recognized it as a bug report. It checked out the relevant Git repository, fixed the issue, committed the code, and then replied on Twitter that it was fixed.

“I thought, this actually works?

Another even more amazing story. He was walking on the street, too lazy to type, so he sent a voice message. The problem was, he hadn’t even enabled voice message support for Clawbot.

“I saw it show ‘typing,’ and thought, oh no. But it replied normally.

He later asked the AI how it did that. The AI said: I received a file without an extension, so I looked at the header and found it was Ogg Opus format. You have ffmpeg on your computer, so I used it to convert to WAV. Then I looked for whisper.cpp, but you didn’t have it installed. However, I found your OpenAI API key and used curl to send the audio for transcription.

Peter Yang listened and said: These things are really clever, though a bit scary.

“Much better than web-based ChatGPT, it’s like a liberated ChatGPT. Many people don’t realize that tools like Claude Code are not just good at programming—they’re capable of solving any problem.

The command-line tool (CLI) army

Peter Yang asks how those automation tools are built—whether he writes them himself or lets AI do it.

Peter Steinberger laughed.

He’s been expanding his “CLI army” over the past few months. What are agents best at? Calling command-line tools, because that’s what the training data is full of.

He built a CLI that accesses all of Google’s services, including Places API. Also one specifically for searching emojis and GIFs, so AI can send memes in replies. He even made a tool to visualize sound, aiming to let AI “experience” music.

“I also hacked into the local food delivery platform’s API. Now AI can tell me how long until my food arrives. And I reversed the API of Eight Sleep, so I can control my bed’s temperature.

【Note: Eight Sleep is a smart mattress that can adjust surface temperature; its API is not officially open.】

Peter Yang asked further: Did you have AI help you build these?

“The most interesting thing is, I used to work at PSPDFKit for 20 years, developing in Swift and Objective-C, very specialized. But after returning, I decided to switch tracks because I was fed up with Apple’s control and the limited audience for Mac apps.

The problem was, switching from a highly skilled tech stack to another was painful. You understand all the concepts but don’t know the syntax. What is ‘prop’? How to split an array? Every small question requires looking up info, and you feel like an idiot.

“Then AI came along, and all that disappeared. Your system-level thinking, architecture skills, taste, judgment on dependencies—those are what really matter, and now you can easily transfer them to any field.

He paused:

“Suddenly, I feel like I can build anything. Language doesn’t matter anymore; what matters is my engineering mindset.

Controlling the real world

Peter Steinberger started demonstrating his setup. The list of permissions he grants AI is astonishing:

Email, calendar, all files, Philips Hue lights, Sonos speakers. He can wake himself up in the morning, gradually increasing volume. AI can also access his security cameras.

“One time I had it watch for strangers. The next morning, it told me, ‘Peter, someone’s there.’ I checked the footage, and it was screenshotting my sofa all night because the camera quality is poor, and the sofa looked like someone sitting there.

In his Vienna apartment, AI can also control the KNX smart home system.

“It can really lock me out.

Peter Yang asked: How do you connect all these?

“Just tell it directly. These things are very manageable; it will find APIs, Google, and look for keys in your system.

Users’ ways of using it are even crazier:

• Someone had it shop on Tesco’s website
• Someone had it order from Amazon
• Someone had it auto-reply to all messages
• Someone added it to their family group chat as a “family member”

“I had it check in on British Airways’ website. It’s basically a Turing test—operating a browser on an airline site, which is notoriously user-unfriendly.

It took nearly 20 minutes the first time because the whole system was rough. AI had to find the passport in his Dropbox, extract info, fill out forms, pass CAPTCHA.

“Now it only takes a few minutes. It can click the ‘I am human’ verification button because it’s controlling a real browser, with behavior indistinguishable from a human.

80% of apps will disappear

Peter Yang asked: What’s a safe way for new users to get started?

Peter Steinberger said everyone’s path is different. Some start using it to write iOS apps immediately, others manage Cloudflare. One user installed it for himself in the first week, then for family in the second, and started making enterprise versions for his company in the third.

“I installed it for a non-technical friend, and he started sending me pull requests. He’s never sent one before in his life.

But what he really wanted to say was the bigger picture:

“Think about it—this thing might replace 80% of the apps on your phone.

Why still use MyFitnessPal to log food?

“I have an infinitely resourceful assistant that already knows when I make bad decisions at KFC. I send a photo, and it stores it in a database, calculates calories, and reminds me to go to the gym.

Why set the temperature of Eight Sleep with an app? AI has API access and can do it directly. Why use a to-do app? AI keeps track for you. Why check in for flights with an app? AI does it for you. Why shop with an app? AI can recommend, order, and track.

“There will be a whole layer of apps gradually disappearing because if they have APIs, they’re just services your AI can call.

He predicts 2026 will be the year many start exploring personal AI assistants, with big companies entering the space.

“Clawbot isn’t necessarily the ultimate winner, but this direction is correct.

Just Talk to It

The topic shifted to AI programming methodology. Peter Yang said he wrote a very popular article called “Just Talk to It,” and wanted to hear him elaborate.

Peter Steinberger’s core point is: don’t fall into the “agentic trap.”

“I see too many people on Twitter discovering how powerful agents are, then wanting to make them even more powerful, and falling down the rabbit hole. They build all kinds of complex tools to accelerate workflows, but they’re just building tools, not creating truly valuable things.

I’ve fallen into it myself. Early on, I spent two months building a VPN tunnel just to access terminals on my phone. It was so good that once, while dining with friends, I was vibing coding on my phone the whole time instead of engaging in conversation.

“I had to stop, mainly for mental health reasons.

Slop Town

He recently got frustrated with a system called Gastown.

“A super complex orchestrator that runs a dozen or twenty agents simultaneously, communicating and dividing tasks. It has watchers, overseers, mayors, pcats (probably ‘civilians’ or ‘pets’), I don’t even know what else.

Peter Yang: Wait, there’s a mayor?

“Yes, Gastown has a mayor. I call this project ‘Slop Town.’

And RALPH mode (a ‘use-and-discard’ single-task loop mode, where you give AI a small task, then discard all context and start fresh, looping endlessly)…

“This is basically the ultimate token burner. Run it all night, and in the morning, you get pure slop.

The core problem: these agents lack taste. They’re frighteningly smart in some ways, but if you don’t guide them, tell them what you want, they produce garbage.

“I don’t know how others work, but I start a project with only a vague idea. During building, playing, feeling, my vision gradually becomes clearer. I try things, some don’t work, and my ideas evolve into the final form. My next prompt depends on what I see, feel, and think at the moment.

If you try to write everything into a detailed spec upfront, you miss this human-machine feedback loop.

“I don’t see how you can produce good stuff without feeling and taste involved.

Someone on Twitter boasted about a note app generated entirely by RALPH. Peter replied: Yes, it looks like RALPH generated it; no sane person would design it like that.

Peter Yang summarized: Many people run AI 24 hours not to make apps but to prove they can run AI 24 hours.

“This is like a size comparison contest without any reference. I also let it run for 26 hours once, feeling proud. But it’s a vanity metric, meaningless. Building everything doesn’t mean you should build everything, nor that it will be good.

Plan Mode is a patch (hack)

Peter Yang asked how he manages context. When conversations get long, does AI get confused? Does he need to manually compress or summarize?

Peter Steinberger said this is an “old mode problem.”

“Claude Code still has this issue, but Codex is much better. On paper, it might have only 30% more context, but it feels like 2-3 times more. I think it’s related to internal thinking mechanisms. Now, most of my feature development happens within a single context window, with discussion and building happening simultaneously.

He doesn’t use worktrees because that’s “unnecessary complexity.” He simply checks out multiple repositories: clawbot-1, clawbot-2, clawbot-3, clawbot-4, clawbot-5. Whatever’s free, he uses, tests, pushes to main, and syncs.

“It’s like a factory—if they’re all busy. But if you only run one, the wait time is too long, and you can’t get into the flow.

Peter Yang said it’s like an RTS game—you have a squad attacking, and you need to manage and monitor them.

Regarding plan mode, Peter Steinberger has a controversial view:

“Plan mode is a patch that Anthropic had to add because the model is too impulsive, jumping straight into coding. If you use the latest models, like GPT 5.2, you just talk to it. ‘I want to build this feature, do it this way, I like this style, give me a few options, let’s discuss.’ Then it proposes, you discuss, reach consensus, and start.

He doesn’t type; he talks.

“Most of the time, I just talk to it.

Discord-driven development

Peter Yang asked about his process for developing new features—exploring problems first? Planning first?

Peter Steinberger said he did “probably the craziest thing I’ve ever done”: he connected his Clawbot to a public Discord server, so everyone can chat with his private AI, with his personal memory, in a public space.

“This project is hard to describe in words. It’s a mix of Jarvis (Iron Man’s AI assistant) and the movie ‘Her.’ Everyone I demoed it to was super excited, but posting pictures and text on Twitter didn’t go viral. So I thought, let people experience it themselves.

Users ask questions, report bugs, request features in Discord. His current process: take a screenshot of the Discord chat, drag it into the terminal, and tell AI, ‘Let’s talk about this.’

“I’m too lazy to type. If someone asks ‘Do you support this or that,’ I let AI read the code and generate a FAQ.

He also built a crawler that scans the Discord help channel at least once a day, summarizes the biggest pain points, and they fix them.

No MCP, no complex orchestration

Peter Yang asked: Do you use those fancy things? Multi-agent systems, complex skills, MCP (Model Context Protocol)?

“My skills are mostly life skills: tracking diet, grocery shopping, that kind of stuff. Programming-wise, very little, because I don’t need it. I don’t use MCP or any of those.

He doesn’t believe in complex orchestration systems.

“I’m in the loop, and I can make better products that feel right. Maybe there are faster ways, but I’m already near the bottleneck—not AI, but my own thinking speed, occasionally limited by waiting for Codex.

His former PSPDFKit co-founder, a former lawyer, now also submits PRs (pull requests) to him.

“AI allows non-technical people to build things, which is amazing. I know some oppose it, saying the code isn’t perfect. But I treat pull requests as prompt requests—they convey intent. Most people don’t have the same systemic understanding to guide the model to the best result. So I prefer to get the intent and do it myself, or rewrite based on their PR.

He tags them as co-authors but rarely merges their code directly.

Find your own way

Peter Yang summarized: The core point is, don’t use slop generators; keep humans in the loop because the human brain and taste are irreplaceable.

Peter Steinberger added:

“Or, find your own way. Many ask me ‘how did you do it,’ and the answer is: you have to explore yourself. Learning these things takes time and making your own mistakes. It’s the same as learning anything, just in a field that changes extremely fast.

Clawdbot can be found on clawd.bot and GitHub. Clawd with W, C-L-A-W-D-B-O-T, like lobster claws.

【Note: ClawdBot has been renamed OpenClaw】

Peter Yang said he also needs to try it. Doesn’t want to sit in front of the computer chatting with AI, but wants to give commands anytime while outside, taking care of his kids.

“I think you’ll like it,” Peter Steinberger said.

Peter Steinberger’s core message can be summarized in two sentences:

1. AI is powerful enough to replace 80% of the apps on your phone
2. But without human taste and judgment in the loop, the output is garbage

These two statements seem contradictory but actually point to the same conclusion: AI is a lever, not a replacement. It amplifies what you already have: system thinking, architecture skills, intuition for good products. Without these, running multiple agents 24 hours a day is just mass production of slop.

His own practice is the best proof: a 20-year veteran iOS developer built a 300,000-line project in a few months, not by learning new language syntax, but by leveraging language-agnostic skills.

“Programming languages don’t matter anymore; what matters is my engineering mindset.”