What it does

NanoClaw is a personal AI assistant that connects to WhatsApp, Telegram, Discord, Slack, Gmail, and a dozen other messaging channels. You message it with a trigger word (default @Andy), and it can reply, run scheduled jobs, search the web, or access files you explicitly mount. The whole thing installs with one bash script that bootstraps Node, Docker, and your first channel.

The interesting bit

The author built this because OpenClaw’s 500K lines, 53 config files, and shared-memory architecture felt too complex to trust with full access to their life. NanoClaw’s answer is radical minimalism: one Node host process, a handful of source files, and each agent running in its own Linux container with filesystem isolation. Communication between host and container happens through two SQLite files (inbound.db and outbound.db) — no IPC, no stdin piping, exactly one writer per file. It’s an architecture that treats “small enough to understand” as a security feature.

Key highlights

• Agents run in Docker containers with explicit mounts, not application-level permission checks; credentials never enter the container (routed through OneCLI’s Agent Vault)
• Channels and alternative providers (OpenAI Codex, Ollama, OpenRouter) install on demand via /add-<name> skills — trunk stays lean, your fork only gets what you ask for
• Per-agent workspaces with separate CLAUDE.md, memory, and container config; flexible isolation lets you share one agent across channels or keep them strictly separated
• AI-native setup: nanoclaw.sh handles deterministic steps, hands off to Claude Code when judgment is needed; same for debugging (“describe the problem in chat”)
• No config files — customization means Claude Code modifies the source directly, which the author considers safe because the codebase is small enough to reason about

Caveats

• Requires Claude Code installed for setup, customization, and all /add-* skills; not a self-contained tool
• macOS/Linux/WSL2 only; Windows native is unsupported
• The “no config files” philosophy is elegant until you want to track changes or revert; the README doesn’t mention version control guidance for your fork

Verdict

Worth a look if you want a multi-channel AI assistant and OpenClaw’s scope makes you nervous. Skip it if you need a polished SaaS with a dashboard, or if “just have Claude modify the code” sounds like a maintenance strategy you’d regret in six months.