NanoClaw Review

NanoClaw is described as an AI assistant that runs agents securely in their own containers, lightweight and built to be understood and customized. The author explains the motivation in the README: OpenClaw is impressive, but nearly half a million lines of code, 53 config files, and 70-plus dependencies, with security at the application level (allowlists, pairing codes) rather than OS-level isolation, made them uncomfortable giving it access to their life. NanoClaw offers the same core functionality as one process and a handful of files, with agents in Linux containers that only see what is explicitly mounted, so Bash is safe because commands run inside the container. The README lists philosophy points: small enough to understand, secure by isolation, built for the individual user, customization equals code changes, AI-native hybrid install, skills over features (channels live on a long-lived channels branch and providers on a providers branch, added by commands such as /add-telegram), and best harness with Claude Code via the Claude Agent SDK, with other providers such as Codex and OpenCode as drop-in options. A newer Slack feature provisions one Slack app per agent.

Quick start is to clone the repository (the README names a nanoclaw-v2 folder) and run bash nanoclaw.sh. The script installs Node, pnpm, and Docker if missing, installs a credential gateway and registers your Anthropic credential, builds the agent container, and pairs your first channel. If a step fails, Claude Code is invoked automatically to diagnose it. A migrate-v2.sh script moves an existing v1 install's state, group folders, scheduled tasks, and channel auth into a v2 checkout without flipping the system service until you choose to. You must run the scripts directly, not from inside a Claude session. We did not run them.

There is deliberately no dashboard or debugging UI; you talk to the assistant in a chat channel and ask Claude Code to explain or change the system. Each spawned agent can have its own bot identity, container, and memory, with shared rooms. If you like conversational control and reading code, this is a feature; if you want a graphical admin panel, it is a gap. This review is based on the README only.

• Each agent runs in its own Linux container with filesystem isolation
• Channels for Slack, Telegram, Discord, WhatsApp, iMessage, and CLI
• Per-agent Slack apps, identities, and memory
• Claude Code through the Claude Agent SDK by default, with Codex and OpenCode as options
• Scripted install with automatic Claude Code diagnosis on failures
• v1 to v2 migration script
• Skills-based extension instead of a plugin framework

• Genuinely small and readable, which makes audit and customization realistic
• Container-level isolation is a stronger security boundary than permission prompts alone
• Fast-growing, with about 31,000 stars and daily commits
• MIT license

• Customization means editing code, which assumes you are comfortable with that or happy to delegate it to Claude Code
• Defaults are tied to Anthropic credentials and Claude Code, and other providers are add-on branches
• Requires Docker and a terminal-based install
• No management UI, so visibility depends on chat and logs

NanoClaw is free under the MIT license. You pay the model provider (by default Anthropic) for usage and run the host yourself.

OpenClaw is a much larger personal-agent project with many integrations and application-level permissions. NanoClaw trades breadth for a small codebase and container isolation, so you get fewer built-in features but a system you can read and sandbox. Choose based on whether you value feature coverage or auditability.

• Security-minded users who want an agent they can fully read and audit
• Developers comfortable forking and modifying their own assistant
• People who live in Slack, Telegram, or WhatsApp

• Users who want a point-and-click setup and admin dashboard
• Teams that need a large catalogue of ready-made integrations
• Anyone unwilling to run Docker