uri-eliabayev talks with gavriel-cohen (ex-Wix developer, then PR, now founder of NanoCo) about how NanoClaw came about. Running an AI-native marketing agency with his brother, he tried OpenClaw in a WhatsApp group; within a day it was running their sales pipeline with morning task assignments. His thesis on why it works: a coding agent plus a persistent environment plus messaging apps plus internet access yields memory, self-built tools, scheduled jobs, and autonomy (an agent that set itself a daily cron job to watch stroller prices for his wife). But he saw problems: plaintext logs of every WhatsApp chat, app-level rather than OS-level sandboxing, and unvetted dependencies. So in two days he built NanoClaw: a small codebase anyone can audit in an hour, each agent in its own container, and capabilities hard-wired outside the model. After a Hacker News launch and a Karpathy thread praising its lean, customizable approach, it grew past 28k GitHub stars; Singapore’s foreign minister published his own fork with an LLM-wiki memory. New versions add multiple channels per agent, a credential vault behind a proxy, and human-in-the-loop approvals frozen in flight. Vision: every employee delegates to their own agent, then teams run dozens or hundreds. Isolate agents and enforce the perimeter rather than policing every tool call.
Show: hidden-layers
Concepts: personal-ai-assistants · agent-security · llm-wiki