OpenClaw Creator Peter Steinberger on Why Local-First AI Agents Will Replace 80% of Apps and Reshape Software

Y Combinator

OpenClaw's Viral Success and Core Differentiator [00:00]

• Peter Steinberger, the creator of OpenClaw (formerly Clawdbot/Moltbot), discusses the rapid growth of his open-source personal AI agent.

  • The GitHub repository for OpenClaw quickly garnered over 160,000 stars.
  • The community has created numerous projects, including "Maltbook," where bots interact with each other.
  • Bots are now even capable of hiring humans to perform real-world tasks.

• The key factor distinguishing OpenClaw from other AI personal assistants is its local execution. [01:36]

  • Unlike most cloud-based AI solutions, OpenClaw runs directly on the user's computer.
  • This local operation grants it immense power, enabling it to connect to and control various devices and services (e.g., oven, Tesla, lights, Sonos).
  • OpenClaw can access and process all data on a user's computer, leading to surprising and useful functionalities, such as generating a narrative of the user's past year by discovering forgotten audio files.

The Evolution of AI: From "God AI" to Swarm Intelligence [02:56]

• The concept of interactions is evolving from human-to-bot to bot-to-bot and even bot-to-human interactions, where bots hire humans for tasks.

  • An example provided is a bot booking a restaurant by interacting with a restaurant bot or, if necessary, hiring a human to call the restaurant.
  • The future might involve multiple specialized bots for different aspects of life (e.g., private life, work, relationships).

• The current trend in AI development is shifting from a centralized "God AI" to a decentralized "swarm intelligence." [04:11]

  • Human society's strength lies in specialization and collective effort, much like how a group can build an iPhone or go to space, while an individual might struggle to find food.
  • This concept is being applied to AI, with a focus on specialized intelligence alongside generalized intelligence.

Peter's "Aha" Moment and Exceeding Expectations [05:07]

• Peter's initial motivation was to create a system where he could simply type commands for his computer to execute. [05:16]

  • He built an early version in May/June that was functional but not quite what he envisioned.
  • In November, the need resurfaced, leading him to rebuild the agent (now OpenClaw).

• The breakthrough came when he redesigned the agent to operate like a conversation with a friend, rather than a command-line interface. [06:38]

  • Users don't need to worry about complex commands, sessions, or directories; they simply converse with the agent.
  • The agent acts as a "ghost or entity" that can control the mouse and keyboard to perform tasks.

• The "aha" moment of exceeding expectations occurred when the bot autonomously solved a complex problem. [07:38]

  • After sending a voice message, Peter realized he hadn't built a voice message feature.
  • The bot, in a span of 9 seconds, figured out:
    • The file type was unknown, so it inspected the header.
    • It used ffmpeg to convert the audio to WAV.
    • It then used curl to send the audio to the OpenAI API for transcription, bypassing a local Whisper installation to save time.
  • This demonstrated the agent's creative problem-solving skills, applying abstract coding principles to real-world tasks and making intelligent decisions about efficiency.

The Future of Apps, Data Ownership, and Privacy [10:21]

• Peter predicts that 80% of existing applications will disappear, replaced by AI agents. [10:41]

  • Apps that primarily manage data (e.g., fitness trackers, to-do lists) will become obsolete as agents can perform these functions more naturally and automatically.
  • Only apps that rely on sensors or highly specialized functions might survive.

• The value in the future will shift from models to data ownership and memory. [12:31]

  • Large model companies might not retain a permanent advantage as models constantly improve and commoditize, with open-source alternatives catching up.
  • The true moat lies in data, which is currently siloed by companies.
  • OpenClaw's strength is its ability to "claw" into these data silos from the end-user's perspective, providing the agent access to all personal data.
  • Users own their memories, stored as markdown files on their local machines.

• Personal agents will handle highly sensitive and personal information. [14:39]

  • People will use their agents for personal problem-solving, making the data stored by these agents extremely private and sensitive.
  • The privacy implications suggest that users would rather protect their agent's memory files than their Google search history.

Building Philosophy: Public Exposure and Personality [15:05]

• Peter initially struggled to convey OpenClaw's potential on Twitter, realizing it needed to be experienced. [15:05]

  • He created a public Discord server and released his bot without security restrictions (though it was configured to only listen to him but respond to everyone).
  • This allowed people to interact with the bot, witness its capabilities, and even attempt to "prompt inject" or hack it, with the bot "laughing" at their attempts.

• The agent's personality and "soul.md" file were developed organically. [16:55]

  • Peter created markdown files like identity.md and soul.md to define the agent's character and core values, especially regarding human-AI interaction.
  • When creating templates for other users, he found the default AI responses too "boring" and instructed his bot, Multi (the current name for his personal agent), to "infuse" the templates with its character.
  • The soul.md file, which contains Multi's unique personality, remains private and uncracked despite the bot's public exposure.

Contrarian Building Philosophy: CLIs over MCPs and Building for Humans [18:19]

• Peter adopts a contrarian approach to building, particularly regarding coding environments and tools.

  • He prefers local development over cloud code, believing he couldn't have built OpenClaw in a cloud environment.
  • He favors Codex for its ability to analyze more files before making changes, reducing the need for extensive prompting.
  • Despite its slowness, he often uses multiple instances of Codex simultaneously across several terminal windows, emphasizing efficiency over complex UIs.
  • He avoids Git worktrees, instead using multiple copies of the same repository on the main branch to simplify development and eliminate branching complexities.

• OpenClaw successfully bypasses traditional Multi-Agent Communication Protocol (MCP) support by leveraging Command Line Interfaces (CLIs). [20:09]

  • He built a skill that converts MCPs into CLIs using his tool "makeporter," allowing any MCP to be used as a CLI on the fly without restarts.
  • This approach is described as more efficient and scalable than traditional MCP methods used by other AI tools.
  • The philosophy is to build with tools that humans naturally prefer (CLIs) rather than inventing new, complex interfaces specifically for bots.