OpenClaw rebuilt for people who read code before they run it

What it does

NanoClaw is a self-hosted AI assistant that ingests messages from WhatsApp, Telegram, Slack, Discord, and about a dozen other channels, then routes them to per-agent Linux containers running the Anthropic Claude SDK. Each agent gets its own SQLite-backed queues, explicit filesystem mounts, memory store, and scheduled tasks, all orchestrated by a minimal Node host that keeps raw API keys out of the containers entirely.

The interesting bit

The project treats its own source code as the configuration layer: there are no config files, and the intended way to change behavior is to ask Claude Code to edit the repo directly. Channels and alternative AI providers live on separate Git branches and are copied into your fork only when you run an /add-<skill> command, so trunk stays lean and your install stays bespoke.

Key highlights

• Agents are sandboxed in Docker (or optional micro-VMs / Apple Container) with explicit filesystem mounts rather than application-level permission checks
• Credentials stay outside containers; OneCLI’s Agent Vault injects them at request time with per-agent rate limits and policies
• Supports recurring scheduled tasks, web search, and per-agent memory that can span multiple messaging channels or stay isolated
• Native Anthropic Agents SDK with optional drop-in providers: OpenAI Codex, OpenRouter, and local Ollama
• Architecture uses two SQLite files per session—one writer each—avoiding IPC complexity

Caveats

• Setup, customization, and channel installation all require Claude Code; there is no standalone admin UI
• The README truncates mid-warning with “You should still review what you’re run,” suggesting container isolation is a baseline, not a complete safety blanket
• macOS, Linux, or WSL2 only; native Windows is out

Verdict

Ideal for developers who want a hackable, self-hosted assistant they can understand end-to-end. Not for anyone who wants a polished, zero-code SaaS experience.