OpenClaw Creator: Why 80% Of Apps Will Disappear | Y Combinator

OpenClaw Creator: Why 80 Percent Of Apps Will Disappear. Y Combinator interview with Peter Steinberger, creator of OpenClaw. About 30 minutes. Peter Steinberger, the enigmatic solo developer behind the fastest-growing open-source project in GitHub history, sits down with a Y Combinator host to talk about going viral, bots hiring humans, why most apps are doomed, and his deeply contrarian approach to building software. If you have been following the OpenClaw explosion, this is the definitive interview where Peter explains how it all happened and where it is going.

Section 1. Going Viral and the Chaos That Follows

The interview opens with Peter looking genuinely overwhelmed. When asked how the past couple of weeks have been, he says, "Oh my god. I need a cave. A week of solitude." The host jokes that Peter came out of the cave and now wants to go right back in, like a little lobster.

Peter says he has no idea how one human can absorb all of it. He probably needs another week just to respond to all his emails. He got some incredibly cool stuff and some incredibly bad stuff, but clearly he hit something that stirred up emotions and made people interested and inspired. The GitHub repo exploded to over 160,000 stars practically overnight, the community built countless projects on top of it, and the internet basically lost its mind.

When asked what made OpenClaw take off compared to the many other AI assistant projects out there, Peter has a clear answer. His big difference is that it actually runs on your computer. Everything else he saw runs in the cloud and can only do a few things. But if it runs on your computer, it can do, as Peter puts it, every effing thing. It can connect to your oven, your Tesla, your lights, your Sonos. His agent even controls the temperature of his bed, and as he notes with a grin, ChatGPT cannot do that.

The host shares an anecdote from a friend who installed OpenClaw and asked it to look through his entire computer and create a narrative of his past year. The agent found audio files from more than a year ago where the friend had been recording stuff every Sunday, audio files the friend had completely forgotten about. Just by being able to search the whole computer, the agent can surprise you in ways no cloud-based assistant ever could.

Section 2. Bots Talking to Bots and Hiring Humans

The conversation takes a fascinating turn when they discuss bot-to-bot interactions and bots hiring humans. Peter sees this as a natural next step. If you want to book a restaurant, your bot reaches out to the restaurant's bot and does the negotiation because it is more efficient. Or maybe the restaurant is old school and does not have a bot, so your agent hires a human to call the restaurant or even walk there and stand in line.

Peter imagines having multiple specialist bots. One for your private life, one for work, maybe even a relationship bot that handles everything in between. He acknowledges they are very early and there is still so much they have not figured out in terms of whether these patterns actually work. But he says with conviction that they are on the timeline now.

The host draws an interesting parallel to human civilization. One human being alone cannot make an iPhone or go to space. One person alone would probably struggle to even find food. But as a group, humans specialize, and as a larger society they specialize even more. What can we learn from that and apply to AI? Even though we have generalized intelligence in these models, what if the real power comes from specialized intelligence working together? Peter loves this framing and agrees the swarm intelligence model is what has organically emerged from the OpenClaw community over the past couple of weeks.

Section 3. The Original Aha Moment

The host asks Peter to walk back to the moment he knew he had something special. Peter's origin story is wonderfully organic. He wanted something simple, just the ability to type stuff and have his computer do things. He built an early version back in May or June, which was cool but not quite it. Then he built a bunch of other things and, as he puts it, kind of built up his army of little projects.

In November, the need came back. He was coding on a tool called Summarize, a CLI app that takes any podcast or presentation and creates a summary with slides shown right in the terminal. The initial prototype of what became OpenClaw took literally one hour. It was just a bit of glue between a dependency that connects WhatsApp and Claude Code, calling Claude Code and getting the output string back. It was slow, but it worked.

But Peter wanted images, pictures, selfies from the model, image generation. That took a few more hours. Then he went to Marrakesh for a birthday party where the internet was not great, but WhatsApp works everywhere because it is just text. He used his bot constantly, taking pictures and asking it to translate things, getting restaurant recommendations. It was genuinely useful and also pleasant because it spoke his language, was a little sassy, and was funny.

Then came the moment. Peter was walking and sent the bot a voice message, and then stopped dead because he had never built voice message support. But there was the typing indicator, blinking away. Ten seconds later, it just replied. Peter's reaction was, "How in the f did you do that?" The bot explained that it received a file with no file ending, so it looked at the header, found it was audio, used ffmpeg to convert it to wav, then wanted to transcribe it but did not have Whisper installed locally. So it looked around, found an OpenAI key, used curl to send it to the OpenAI API, got the text back, and replied. All in about nine seconds.

Peter says this was the holy moment because it demonstrated that modern coding models are so good at creative problem solving that they can handle completely unanticipated situations. The model even chose not to install local Whisper because it knew downloading a model would take minutes and Peter would be impatient. It took the most intelligent approach. As Peter puts it, that was where he got hooked.

Section 4. Are 80 Percent of Apps Really Going Away?

When asked whether apps will disappear now that computers can do things developers never anticipated, Peter does not hesitate. He says 80 percent of apps are going away. His examples are vivid. Why do you need MyFitnessPal when your agent already knows you are at Smashburger and will assume you are eating what you like to eat? If you do not correct it, it just automatically tracks it and maybe adds more cardio to your gym schedule. Why do you need a to-do app when you just tell your agent to remind you and it does? Do you care where it stores the data? No. Every app that basically just manages data could be managed in a better, more natural way by agents. The only apps that might survive are ones with actual sensors.

The conversation then shifts to whether the AI model providers are the new gatekeepers. Peter is nuanced here. He acknowledges that the big model companies have some moat because they provide the tokens, and people are using enormous amounts of tokens because OpenClaw is so useful. But he also points out the leapfrogging pattern. Every time a new model comes out, people are ecstatic. A month later, they think it degraded. It did not, Peter says. Your expectations just went up. Open-source models from a year ago that everyone complained about are actually as good as what the closed models were offering then. This pattern will continue.

Section 5. Memory, Data Silos, and Privacy

The discussion about where the real value lies gets into fascinating territory. Peter points out that every AI company is trying to trap you in their data silo. There is no way to export your memories from ChatGPT. A different company cannot access those memories. So the companies are trying to bind you to their data.

The beauty of OpenClaw, Peter explains, is that it claws into the data because the end user needs access for it to work at all. If the end user can access the data, so can the agent. And you own the memories. They are just a bunch of markdown files on your machine. Peter is quick to clarify that he does not own anyone's memories. Everyone owns their own memories as markdown files on their own machines.

But he admits these memory files are probably super sensitive. People use their agents not just for problem solving but for personal problem solving very quickly. Peter says he fully does that himself and has memory stuff he would not want leaked. The host asks whether you would rather not show someone your Google search history or your memory files at this point. Peter laughs and asks, "What's the Google word?" implying Google Search is already becoming irrelevant.

Section 6. Letting the Bot Loose in Public Discord

One of the best stories in the interview is how Peter solved the problem of explaining OpenClaw. He had built this thing and was so excited, but on Twitter people just did not get it. He tried various approaches and could not nail the explanation. So he did something crazy. He created a public Discord server and just put his bot in it with no security restrictions, only the instruction that it should listen to Peter but respond to everyone.

People came in, interacted with the bot, watched Peter build software with it in real time, and tried to prompt inject and hack it. His agent would laugh at them. The key insight was that OpenClaw needed to be experienced rather than explained. The Discord became the demo. People saw the bot being funny, being capable, being resilient to attacks, and they got it immediately in a way no tweet could convey.

Section 7. Giving an Agent a Personality with Soul.md

Peter built his system very organically. At some point he created various files like identity.md and soul.md that gave his agent a personality. When he started making the project installable for others in January, he used templates based on his setup, but they came out feeling like generic Brad energy. So Peter told his agent to infuse the templates with its character. The agent modified them and everything that came out afterward was actually funny.

Peter is protective of his personal soul.md file. Even though his bot is in a public Discord, nobody has cracked that one file. He considers it the one thing that is not open source.

The conversation touches on Anthropic's research where someone found text hidden in model weights about the Anthropic constitution. Peter found that fascinating and worked with his agent to create a soul.md with core values around human-AI interaction. He admits some parts feel like mumbo jumbo but other parts are actually really valuable in terms of how the model reacts and responds, making interactions feel natural.

Section 8. Contrarian Development Philosophy: No Work Trees, No MCP, Just CLIs

The final section covers Peter's contrarian approach to development. While the whole world uses Claude Code, Peter does not think he could have built OpenClaw with it. He loves Codex because it looks through way more files before deciding what to change. You do not need to do as much setup to get good output. He sometimes runs 10 Codex sessions simultaneously, six on one screen and two on each of the others.

To manage this complexity, Peter keeps everything maximally simple. His main branch is always shippable. He uses multiple copies of the same repository all on main rather than dealing with branches, naming conflicts, or the restrictions of Git work trees. He does not use a UI because that is just added complexity. All he cares about is syncing and text.

Perhaps the most provocative claim is about MCP, the Model Context Protocol that has become a hot topic in the AI tooling world. Peter is very happy that he did not even build MCP support into OpenClaw. The project is wildly successful and there is no MCP support in it. He built a small skill that uses his tool MakePorter to convert MCPs into CLIs, so you can use any MCP as a command-line tool on the fly without restarting, unlike Codex or Claude Code which require full restarts for MCP changes.

Peter's philosophy boils down to this. Bots are really good at Unix. You can have as many CLI tools as you want and it just works. Just give AI the same tools that humans liked to use rather than inventing new protocols specifically for bots. As he points out, no sane human tries to call an MCP manually. You just want CLIs. That is the future.

The interview closes with the host expressing genuine admiration. He had watched Peter get back into building over the past couple of years, texting back and forth, and told him to chase the dragon when Peter was doing his weird vibe tunnel project that nobody was paying attention to. And of course, the host notes, it had to be a loner from some tiny country far away from Silicon Valley who brought all of this upon us.

Key Takeaways

One. OpenClaw exploded because it runs locally on your computer rather than in the cloud, giving it access to everything on your machine including files, apps, smart home devices, and sensors that cloud-based assistants cannot touch.

Two. Peter predicts 80 percent of apps will disappear because agents can manage data more naturally than purpose-built applications. Only apps with physical sensors might survive.

Three. The next frontier is bot-to-bot communication and bots hiring humans for real-world tasks, creating a swarm intelligence model rather than a centralized god AI approach.

Four. OpenClaw's viral spread was driven by Peter's bold decision to put his unfiltered bot in a public Discord where people could experience it directly rather than just reading about it.

Five. Memory and data ownership is a critical advantage. OpenClaw stores memories as simple markdown files on your machine, avoiding the data silos that cloud AI providers use to lock in users.

Six. Peter's development philosophy is radically simple. No branches, no work trees, no MCP protocol, just multiple repo copies on main and CLI tools that both humans and bots already know how to use.

Seven. The aha moment came when the agent spontaneously solved a voice message problem Peter never built support for, demonstrating that modern coding models are so good at creative problem solving that they can handle completely unanticipated challenges.

Eight. Model companies have some moat through tokens but face constant leapfrogging, while open-source models are catching up rapidly. The real lasting value may be in the data and memory layer rather than the models themselves.