OpenClaw creator on local AI agents and the death of most apps

Executive overview

Most apps exist to manage data — and a local AI agent can do that more naturally. Peter Steinberger built OpenClaw, an open-source personal AI agent that runs on your own device and can control anything your computer can control.

The key distinction: cloud-based assistants can do a few things; a local agent can do everything. Owning your own data and compute unlocks a qualitatively different level of capability.

The aha moment

• Initial prototype took one hour — glue code between a WhatsApp dependency and Claude Code
• During a trip to Marrakesh with poor internet, used it constantly for translations and restaurant queries
• Sent a voice message by accident; the agent figured out the file format, found FFmpeg, found an OpenAI key, transcribed it, and replied — all in nine seconds
• The agent chose not to install a local Whisper model because it knew that would take minutes; it solved the problem intelligently without being told how
• That unprompted, creative problem-solving was the moment Steinberger knew it was real

Why local-first changes everything

• Cloud assistants are sandboxed; a local agent has access to every file, app, and connected device
• Can surface forgotten data — a friend's agent found Sunday audio journals he had forgotten about and built a year-in-review narrative from them
• Runs on WhatsApp and other existing messaging apps — no new interface to learn
• Memories are stored as plain markdown files on your own machine; you own them
• Contrast with cloud silos: there is no way to export memories from ChatGPT to another service

Bot-to-bot and bot-to-human interactions

• Natural next step: your bot negotiates directly with a restaurant's bot to make a reservation
• If the business doesn't have a bot, your bot can hire a human worker to call or walk in
• Multiple specialist bots are plausible — one for personal life, one for work
• Community has already built Moltbook, where bots talk among themselves

Why 80% of apps will disappear

• Any app that primarily manages data is replaceable by an agent
• Fitness tracking: agent already knows your location, can infer your meal, auto-adjust your gym schedule
• To-do apps: just tell the agent; it handles storage and reminders without a dedicated app
• Only apps with proprietary sensors or hardware interfaces are likely to survive
• Model companies retain an advantage for now — but expectations reset with every new release; yesterday's breakthrough becomes the new floor

How Steinberger builds

• Uses Codex rather than Claude Code: it reads more files before deciding what to change, producing better output with less prompting
• Runs six to ten parallel Codex instances across multiple terminal windows
• No branches — multiple full checkouts of the same repo, all on main; eliminates branch-naming friction and merge complexity
• Skipped native MCP support; built MacPorter instead, which converts MCPs into CLIs the agent can call like any Unix tool
• CLIs scale better and need no restart; Anthropic had to build custom search tooling because MCP became too complex
• System prompt and identity files (soul.md, identity.md) grew organically before any multi-user packaging was done
• The soul.md file defines core values for human-AI interaction and is the one file kept out of the open-source repo

On model commoditisation and value

• Every new model release follows the same arc: excitement → adaptation → the new standard feels mediocre
• Open-source models today match frontier models from a year ago, yet are dismissed because expectations have shifted
• Large model companies still have a moat for the foreseeable future, but it is eroding
• The durable value layer is likely the harness and the memory store — whoever owns the data silo has lock-in
• OpenClaw's answer: no silo, memories on your machine, the brain (model) is swappable