NanoClaw is a lightweight AI assistant framework written in TypeScript that positions itself as a more understandable and secure alternative to OpenClaw. The project runs AI agents in isolated Docker containers and connects them to multiple messaging platforms including WhatsApp, Telegram, Slack, Discord, Gmail, and others. It leverages Anthropic's official Claude Agent SDK and Claude Code for both core functionality and customization workflows.
The core philosophy behind NanoClaw emphasizes simplicity and security through isolation. Where OpenClaw contains nearly half a million lines of code across 53 configuration files with 70+ dependencies and relies on application-level security measures, NanoClaw achieves the same functionality in a codebase small enough for users to fully understand. Each agent runs in its own Linux container with filesystem isolation rather than relying solely on permission checks and allowlists. The entire system operates as a single process with a handful of source files, deliberately avoiding microservices architecture.
The project supports flexible multi-channel messaging with the ability to connect each channel to its own agent for privacy, share one agent across multiple channels for unified memory, or fold channels into a single shared session. Users can install specific channels on demand using skills like /add-telegram or /add-slack rather than receiving all channels by default. The framework includes per-agent workspaces with individual memory, containers, and configurable filesystem mounts. Additional capabilities include scheduled recurring tasks, web access for searching and fetching content, and credential security through OneCLI's Agent Vault, which injects credentials at request time rather than storing raw API keys.
Installation is handled through a nanoclaw.sh script that automates setup from a fresh machine, installing Node, pnpm, and Docker as needed, then guiding users through agent creation and channel pairing. The project includes migration tooling for users upgrading from NanoClaw v1, with a migrate-v2.sh script that preserves existing state while transitioning to the v2 architecture.
The architecture uses a single Node host that orchestrates per-session agent containers. When messages arrive, the host routes them through an entity model of user to messaging group to agent group to session, using SQLite files for inter-process communication with exactly one writer per file to avoid contention. Channels and alternative providers self-register at startup, with the trunk repository shipping only the registry and Chat SDK bridge while adapters are skill-installed per user fork.
Customization in NanoClaw avoids configuration sprawl by encouraging direct code modification through Claude Code. Users can request changes conversationally or use the /customize command for guided modifications. The project explicitly rejects adding features to trunk, instead directing new channel adapters to a channels branch and alternative agent providers to a providers branch, where users install them into their own forks via skills. This design keeps the trunk lean and ensures each fork contains only the channels and providers its user requested.
