gavriel-cohen built NanoClaw after OpenClaw quickly ran his AI-native agency’s sales funnel but proved too insecure to trust with customer data. His threat model assumes any agent can be prompt-injected through email, PRs, or websites, so the architecture must not rely on the agent behaving. Each agent runs in its own container; an orchestrator routes messages; credentials live in an external vault-proxy (OneCLI) that injects them per host; and sensitive actions (sending email, rewiring agents) need human approval executed outside the container. He chose the Claude Agent SDK as the harness: the subscription covers it, the ecosystem is shared, and Anthropic optimizes it. His view is that models aren’t portable. “AI-native open source”: contributions come as skills that teach your agent to customize your fork, keeping the core lean and enabling bespoke software. Co-host Gal says he barely reads diffs now, relying on proof of work.
Show: langtalks
Concepts: agent-security · personal-ai-assistants · skill-engineering · model-selection · ai-verification