nanocoai/nanoclaw

A lightweight personal AI assistant that runs each agent in its own Linux container and connects to your messaging apps.

What it is

NanoClaw is a small TypeScript AI assistant that positions itself as an easier-to-understand, isolation-first alternative to OpenClaw. It runs agents inside per-session Linux containers with filesystem isolation rather than application-level permission checks, connects them to messaging channels, and gives them memory and scheduled jobs. It is built for individual users who want to fork a compact codebase and have Claude Code customize it to their exact needs instead of configuring a large generic framework.

Key features

Container isolation: each agent group runs in its own Docker container and can only see explicitly mounted directories, so bash access runs inside the container, not on the host.
Multi-channel messaging installed on demand via /add-<channel> skills — WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, and email via Resend — with flexible per-channel isolation.
Per-agent workspace: each group has its own CLAUDE.md, memory, container, and mounts, plus reusable agent templates stamped via ncl groups create --template.
Scheduled recurring tasks (with optional script gates), web search and fetch, and a scripted install flow that hands off to Claude Code when a step needs judgment.
Credential security through OneCLI's Agent Vault: agents never hold raw API keys, and outbound requests get credentials injected at the proxy with per-agent policies and rate limits.
Runs on Anthropic's Claude Agent SDK by default, with drop-in alternative providers (/add-codex, /add-opencode, /add-ollama-provider) configurable per agent group.

Tech stack

TypeScript, Node.js >=20, pnpm (pnpm@10.33.0); root package version 2.1.53.
SQLite via better-sqlite3 11.10.0 (two files per session, one writer each), cron-parser 5.5.0 for scheduling, and @clack/prompts for the interactive installer.
OneCLI SDK (@onecli-sh/sdk 2.2.1) for credential injection and a chat dependency (4.29.0) for the messaging bridge.
Agent-runner inside the container uses Bun and the Claude Agent SDK; Docker is the runtime (macOS, Linux, Windows via WSL2).
Tooling: Vitest, ESLint (including eslint-plugin-no-catch-all), Prettier, tsx, Husky; MIT-licensed. Exposes an ncl CLI binary.

When to reach for it

Running a personal assistant that reaches you across several messaging apps while keeping each agent sandboxed.
Wanting a codebase small enough to read and modify, customized through code changes rather than configuration files.
Needing scheduled agent jobs (briefings, reviews, reminders) triggered from chat.
Isolating agents and credentials so an agent can only touch what you explicitly mount and never holds raw keys.

When not to reach for it

You cannot or don't want to run Docker; container isolation is central and Docker is the default runtime.
You prefer configuration files and a stable, feature-complete framework over forking and modifying source per user.
You want a batteries-included assistant where new channels and providers ship in the trunk; here they live on separate branches and are installed per-fork via skills.
You need a large multi-tenant deployment; the design is oriented around individual, bespoke installs.

Maturity signal

NanoClaw is young (created early 2026) but moving fast, with commits pushed the same week as this writing, a large following, and releases already in the 2.x line. Its fork count is exceptionally high relative to stars, which fits the stated design: users are expected to fork and customize their own copy. The high open-issue count is consistent with rapid, active development and heavy real-world use rather than neglect. MIT licensing keeps forking and customization low-risk. Treat it as an actively developed, deliberately minimal project whose base is intended to stay small.

Alternatives

OpenClaw — use instead when you want the more full-featured, batteries-included assistant and are comfortable with a much larger codebase and application-level permissioning.
Anthropic's Claude Agent SDK, used directly — build on it yourself when you want a bespoke agent and don't need NanoClaw's multi-channel routing, container isolation, and scheduling scaffolding.
n8n — use instead when your need is app-to-app workflow automation rather than a conversational, memory-carrying assistant.

Notes

Contribution policy is unusual: only security fixes, bug fixes, and clear improvements are accepted to the base; new channels and providers must land on the channels/providers registry branches and are installed per-fork.
The architecture uses two SQLite files per session (inbound.db and outbound.db), each with exactly one writer, to avoid IPC and cross-mount contention between the host and container.
There are no configuration files by design — customization means asking Claude Code to change the code — and the uninstaller tags each checkout so it removes only what that copy created.

Wiki: nanocoai/nanoclaw