WhoBuilt

OpenClaw Creator: Why 80% Of Apps Will Disappear

Feb 7, 2026· 23 min· 1.0M views·Watch on YouTube

Speaker

Peter Steinberger, creator of OpenClaw, talks with Y Combinator about building a computer-based personal AI agent and what agents could change for software and builders.

Why OpenClaw took off

Steinberger says OpenClaw stood out because it runs on a user's computer, where it can work with the files and tools available there. That access let it surprise users, such as by finding old audio recordings and turning them into a narrative of someone's year.

The prototype's unexpected leap

Steinberger built an early version by connecting WhatsApp to Claude Code, then added image support. While traveling in Marrakech, he sent a voice message and was surprised when the agent identified its audio format, converted it, found an available OpenAI key, and arranged transcription. He saw the episode as evidence that coding models can apply creative problem-solving to tasks beyond coding.

Agents could replace many apps

Steinberger predicts that 80% of apps could disappear, especially those that mainly manage data. He imagines an agent tracking meals, adjusting a gym plan, and setting reminders without requiring the user to choose an app or care where information is stored. He suggests apps tied to sensors may be more likely to remain.

Memory, identity, and ownership

Steinberger says chat services keep users' memories in separate data silos, while OpenClaw stores them as Markdown files on users' own machines. He considers those files highly sensitive because people use agents for personal problems as well as practical tasks. He also describes shaping the agent's behavior through identity and soul files, while keeping his own soul.md private.

A low-friction way to build

Steinberger describes running several copies of a repository on main and using parallel terminal sessions to work with coding agents. He says this keeps his workflow simple and avoids branch and worktree overhead. For integrations, he favors turning MCP tools into command-line interfaces, which agents can use without restarting.

Takeaways

  • Give agents access to the files and tools needed to complete useful tasks on a user's computer.
  • Test whether a model can solve an unanticipated task, not only the tasks explicitly built into a prototype.
  • Look for app opportunities where an agent can manage data and workflows through a natural conversation.
  • Treat agent memory as sensitive user data and make its ownership and location clear.
  • Choose a development workflow that reduces coordination overhead when working with multiple coding agents.

Mentioned in this talk

Search inside this talk

What was claimed 9

Show all 9 facts

Related talks