OpenClaw Creator: Why 80% Of Apps Will Disappear
These notes were generated by AI and may contain inaccuracies.
Peter Steinberger created OpenClaw, an open-source personal AI agent running on GitHub with over 160,000 stars overnight. Community projects include Maltbook where bots interact. Bots now rent humans for real-world tasks. Discussion covers his aha moment, contrarian development philosophies, and implications for builders in 2026.
Past one to two weeks have been overwhelming for Peter; needs solitude to process emails and attention. Received cool and bad feedback, but hit something that stirred emotions, inspired people.
OpenClaw runs locally on your computer, unlike cloud-based alternatives. Local access enables controlling anything on the machine: oven, Tesla, lights, Sonos, bed temperature. JPD can't do that.
Friend installed OpenClaw; it scanned computer, created narrative of his last year by finding forgotten audio files from over a year ago recorded every Sunday.
Moving from human-to-bot to bot-to-bot interactions. Bots negotiate restaurant bookings efficiently. For old restaurants without bots, bot hires human to call or stand in line.
Multiple specialist bots possible: private life, work, relationship bot. Still early, much to figure out, but on the timeline now.
Shift from centralized god intelligence to swarm and community intelligence. One human can't make an iPhone, go to space, or find food alone; society specializes. Apply to AI: specialized intelligence alongside generalized.
Wanted AI to type commands to automate computer tasks. Built version in May/June, then more tools. In November, needed it again while coding a CLI app (summarize) that handles podcasts/hot seats, summarizes, shows slides in terminal.
Came out of retirement after Wipe Tunnel (two months work), addictive coding. Rebuilt as OpenClaw (initially cloudbot): talk naturally like to a friend who controls mouse/keyboard, no thinking about sessions, folders, models.
Initial shitty prototype in one hour: glue between WhatsApp dependency and Claude code. Added image generation.
In Marrakesh for birthday party with poor internet, WhatsApp worked: translated menus via pictures, sassy/funny responses. Walking, sent voice message; saw typing indicator, got reply explaining process: detected text message without file ending via header, used ffmpeg to convert to WAV, no local Whisper so curled OpenAI key for transcription—all in 9 seconds.
"Holy fuck."
Coding models excel at creative problem solving, applicable to real-world tasks. Model intelligently avoided installing local Whisper to save time.
80% of apps will disappear. No need for MyFitnessPal: agent tracks eating at Smashburger, assumes preferences, adjusts gym schedule with more cardio. No need for to-do apps: just remind verbally, it handles storage.
Apps managing data can be handled naturally by agents. Apps with sensors may survive.
Models commoditize; swappable. Model companies face issues with high token usage from popularity. Models leapfrog, but degrade in perception as expectations rise.
New models hyped, then become standard. Open-source matches last year's models, but complained about until it becomes norm.
Big companies have moat in harness (data silos like ChatGPT memories inaccessible). OpenClaw accesses user data as owner; memories are markdown files on your machine, highly personal/sensitive (personal problem solving).
Failed to explain awesomeness on Twitter; created public Discord with unrestricted bot (locked to his user ID via system prompt). People interacted, tried prompt injection, bot laughed.
Built organically: identity.md, soul.md. In January, made installable; used Codex for templates, then infused with character via Multi (personal bot name, due to naming challenges). Kept soul.md secret (not open-source), core values on human-AI interaction, inspired by Entropic research on hidden text in model weights (nicolity constitution). Mix of mambo jumbo and valuable response guidelines.
Prefers Codex over Claude for coding (looks at more files, less charade). Slow, so runs 6-10 in parallel terminals.
No git worktrees: multiple main checkouts (always shippable), avoids branch naming/conflicts/restrictions. No UI (added complexity); focuses on syncing/text.
Discuss design with agent. No native MCP support; skill uses makeporter (his tool) to convert MCPs to CLIs. On-the-fly, no restarts (unlike Codex/Claude). Scales better than custom MCP tools like Entropic's search feature. Bots good at Unix/CLIs, like humans.
Huge inspiration; chased the dragon from vibe tunnel etc. Loner from tiny country far from Silicon Valley.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
