NanoClaw is a new open source AI agent platform built by Gavriel Cohen as a security-focused alternative to OpenClaw. Its key differentiators are container isolation per agent (limiting blast radius if an agent misbehaves) and a small ~4,000-line codebase that is realistically auditable. OpenClaw's 400,000-line codebase and bare-metal execution model have led to real security incidents, including an agent deleting a user's inbox. Cohen built NanoClaw in a weekend using Claude Code, layering on Anthropic's Agent SDK. Andrej Karpathy highlighted the project for its manageable size and skills-based configurability. Cohen argues that container-per-agent isolation, combined with minimal integrations, gives enterprises a practical path to deploying agents safely. The project is open source and Cohen plans to keep it that way while building a commercial platform on top.
Sort: