NanoClaw - Secure AI Agent for WhatsApp, Telegram & More

Source
ai-agentsclaude-codewhatsapp-bottelegram-botcontainer-isolationself-hostedopen-source

Summary

NanoClaw is a minimalist, self-hosted AI agent that connects to WhatsApp and Telegram, running Claude Code inside isolated containers (Apple Container or Docker). It positions itself as a radically simpler alternative to “OpenClaw” with only 15 source files and ~3,900 lines of code versus 434k+, making the entire codebase auditable in under 10 minutes.

Key Insight

  • The core value proposition is security through simplicity: agents run in OS-level container isolation rather than application-level checks, meaning each chat group gets its own sandboxed filesystem, IPC namespace, and Claude session
  • Architecture is a single Node.js process that polls for messages, manages per-group FIFO queues (default 3 concurrent containers), and uses filesystem-based IPC (JSON files) rather than sockets or network calls
  • Supports “agent swarms” where multiple specialized agents collaborate on complex tasks, which they claim is a first for personal AI agents
  • Per-group memory via isolated CLAUDE.md files means each conversation context stays separate, which is useful for multi-purpose setups (work group vs. personal group)
  • The setup flow is “AI-native” - you clone the repo, run claude, and the /setup skill handles dependencies, auth, container config, and service configuration. No traditional install wizard or dashboard
  • The skills system (Gmail, Telegram, custom capabilities) follows a pattern of Claude Code skills that transform the installation rather than adding code bloat
  • Concurrency control uses exponential backoff on failure, which is a solid pattern for rate-limited AI API calls