// the find
nanocoai/nanoclaw
A lightweight alternative to OpenClaw that runs in containers for security. Connects to WhatsApp, Telegram, Slack, Discord, Gmail and other messaging apps,, has memory, scheduled jobs, and runs directly on Anthropic's Agents SDK
NanoClaw is a self-hosted AI assistant that sandboxes agents in Docker containers and connects them to messaging platforms like WhatsApp, Telegram, Slack, and Discord. It positions itself as a lightweight, auditable alternative to OpenClaw, small enough that you can actually read the source. The target user is a developer who wants a personal AI assistant they can modify and trust.
Container-level isolation is a real security win over the typical 'allowlist in userland' approach — bash access that only touches the container filesystem is meaningfully safer. The two-SQLite architecture (one inbound, one outbound, one writer each) is a clean design that avoids the usual shared-state headaches at the cost of complexity you'd get from IPC or message queues. The skills-over-features contribution model keeps trunk genuinely small and avoids the feature bloat that killed similar projects. Credential injection via OneCLI's Agent Vault rather than mounting secrets into containers is the right call and prevents a common foot-gun.
Docker-on-macOS has persistent performance overhead and Docker Desktop licensing concerns for commercial use — this is a real friction point the README waves away with 'just run the script'. The setup script handing off to Claude Code for failure recovery is clever in theory but creates a dependency on a paid Anthropic subscription just to finish installation, which will surprise users who thought this was free to run locally with Ollama. The branch-based skill distribution model (channels branch, providers branch) is architecturally tidy but will create maintenance drag as the project ages — a channels branch that diverges for 18 months from trunk is a fork in everything but name. No mention of what happens to in-flight messages when you upgrade the agent container, which is exactly the kind of question you need answered before running anything important on this.