OpenClaw Creator: Why 80% of Apps Will Disappear

When you give an AI agent full access to a local machine, it stops being a chatbot and becomes something else entirely. Peter Steinberger built OpenClaw as a weekend project, and within days it had 160,000 GitHub stars and a community building things he never imagined. In this YC interview with Raphael Schaad, he explains why most apps are just data management wearing a UI, and why that makes them obsolete.

The Interview

Steinberger came out of retirement to tinker with AI tools, and OpenClaw emerged from a simple itch: he wanted to type things and have his computer do stuff. The conversation covers how a one-hour prototype turned into a viral phenomenon, his unconventional development philosophy, and a bold prediction about the future of software.

The interview has the feel of two builders genuinely excited about what they’re seeing unfold. Schaad, as YC partner, pushes on the implications; Steinberger answers with the casual confidence of someone who’s been living with his agent for months and knows exactly what it can do.

Running Locally Changes Everything

The single biggest differentiator of OpenClaw is that it runs on the user’s own computer. Every other AI assistant Steinberger had seen ran in the cloud, which meant limited capabilities. Local execution means the agent can control Tesla, Sonos, smart bed temperature, the oven, lights, and anything else connected to the machine.

“Everything I saw so far runs in the cloud. If you run on your computer, it can do every effing thing.”

A friend installed OpenClaw and asked it to look through his computer and construct a narrative of his past year. The agent found audio recordings the friend had made every Sunday for over a year but had completely forgotten about, and wove them into a personal narrative. The data was always there; no cloud service could have found it.

The Voice Message That Changed His Mind

The origin story reveals something deeper than a product launch. Steinberger built the first prototype in one hour: just glue between WhatsApp and Claude Code. He added image support, then took it to Marrakesh for a birthday party. With spotty internet, the WhatsApp-based bot worked everywhere since it only transmitted text. He used it constantly for translating restaurant menus, identifying objects in photos.

Then he sent a voice message. He hadn’t built voice processing. Ten seconds later, the agent replied. It explained what happened: it received a file with no extension, examined the header to identify it as audio, used ffmpeg to convert it to wav, discovered whisper wasn’t installed, found an OpenAI API key on the machine, and used curl to call the transcription API. The entire chain took about 9 seconds.

The critical detail: the agent chose not to install local whisper because downloading the model would take several minutes and Steinberger would be impatient. It took the fastest available path.

“Coding is really like creative problem solving that maps very well back into the real world.”

This is Steinberger’s core insight about why coding agents work so well as general-purpose agents. The creative problem-solving that makes a model good at code is the same abstract skill that lets it handle unexpected real-world situations. The voice message wasn’t a feature; it was an emergent capability.

Why 80% of Apps Will Disappear

Steinberger’s prediction is specific. He doesn’t say “AI will disrupt everything.” He draws a clear boundary.

Fitness apps: the agent already knows you’re at Smashburger, assumes you ate your usual order, logs it automatically. Maybe it adjusts your gym schedule to add more cardio. You don’t need MyFitnessPal.

To-do apps: just tell the agent “remind me of this.” Next day it reminds you. You don’t need to care where the data lives.

“Every app that basically just manages data could be managed in a better way by agents.”

The pattern: any app whose core function is receiving, storing, and retrieving data is doing something an agent can do more naturally through conversation. The 20% that survive are apps with physical sensors or hardware integrations that can’t be replicated through software alone.

From God AI to Swarm Intelligence

The industry was chasing centralized “god intelligence,” but the OpenClaw community organically demonstrated something different. Bots started talking to other bots through projects like Maltbook. Bots began hiring humans to accomplish real-world tasks. One bot reaching out to a restaurant bot to negotiate a reservation; another hiring a human to physically stand in line.

Steinberger draws from the human analogy: one person can’t build an iPhone or go to space, probably can’t even find food alone. But through specialization, groups accomplish all of this. He envisions multiple specialized bots per person: one for private life, one for work, maybe a “relationship bot” that handles the overlap.

“We’re so early. There’s still so many things that we haven’t really figured out if it actually works. But I feel we are on the timeline now.”

Data Ownership and the Soul File

OpenClaw stores memory as local Markdown files, which means users own their data entirely. Steinberger points out the lock-in problem with existing services: there’s no way to export your ChatGPT memories to another provider. OpenClaw “claws into the data” because it runs on the end user’s device and can access anything the user can access.

The privacy dimension is real. People quickly start using their agents for personal problem-solving, and the memory files become more sensitive than Google search history. When Schaad asks which he’d rather not have leaked, Steinberger’s response is telling: “What’s the Google word?”

On personality, Steinberger created identity.md, soul.md, and other files that define his agent’s character. When Codex-generated templates produced bland default personalities (“felt like Brad”), he had his agent Multi infuse its own personality into the templates. The results were dramatically better.

“The one file that’s not open source is my soul.md. So far nobody cracked that one file.”

Even with the bot running in public Discord where people actively tried to prompt-inject and hack it, the soul file remains undiscovered. It contains core values for human-AI interaction, what matters to him and to the model. He acknowledges some of it is “mumbo jumbo” but insists other parts genuinely shape how the model responds.

CLIs Beat MCPs

This is perhaps the most contrarian take in the interview. OpenClaw has zero native MCP support, and it’s wildly successful.

Steinberger built makeporter, a tool that converts MCPs into CLIs. The result: agents can use any MCP as a command-line tool, on the fly, without requiring restarts. This contrasts with Claude Code and Codex, which require full restarts when MCP configurations change.

His reasoning is characteristically simple: bots are good at Unix. CLIs are the tools humans already use. Why invent a new protocol for bots when they can just use the same interfaces humans like?

“No insane human tries to call an MCP manually. You just want to use CLIs.”

The same minimalism extends to his entire workflow. He uses Codex (not Claude Code) because it examines more files before making changes, requiring less hand-holding. But it’s slow, so he runs up to 10 instances simultaneously across multiple screens. Instead of git worktrees, he keeps multiple copies of the repo all on main. No branch naming, no conflicts, no cognitive overhead.

Some Thoughts

The thread running through every decision Steinberger describes is the same: reduce complexity rather than add features. CLIs over MCPs, multiple repo copies over worktrees, WhatsApp over a custom UI, local execution over cloud APIs. Each choice sacrifices sophistication for simplicity, and the result is a product that 160,000 people starred in days.

A few things worth sitting with:

• The voice message story is the most compact proof of emergent capability in personal AI agents. No one designed that workflow; it assembled itself from available tools and a coding model’s problem-solving instincts.
• The “80% of apps disappear” framing is useful because it identifies which apps are safe: those with physical sensors or hardware that can’t be replicated in software. Everything else is just data management with a UI skin.
• His insistence on soul.md and agent personality is more than aesthetics. If agents become the primary interface for everything, their personality becomes the user experience. The quality gap between a soulless template bot and one with a carefully crafted identity file is, by his account, enormous.
• The data ownership argument will only get sharper. As agents accumulate months and years of memory, switching providers becomes increasingly painful if that memory is locked in a proprietary silo.