NanoClaw is a personal AI assistant framework that runs Claude agents inside their own Linux containers. It positions itself as a minimal alternative to OpenClaw: where OpenClaw is a large codebase with application-level security (allowlists, pairing codes) running in a single shared Node process, NanoClaw keeps the same core functionality in one process and a handful of source files, and relies on OS-level filesystem isolation — agents can only see directories explicitly mounted into their container, and shell commands run inside the container rather than on the host. Credentials never enter the container; outbound API requests are routed through a proxy vault that injects authentication at request time and enforces per-agent policies and rate limits.
It is built for individual users who want a bespoke assistant rather than a configurable platform. There are no configuration files: customization means editing the code, typically by asking Claude Code to make the change, and the codebase is deliberately kept small enough for that to be safe. The trunk repository ships only the registry and infrastructure; channel adapters (WhatsApp, Telegram, Discord, Slack, Teams, iMessage, Matrix, email, and others) and alternative model providers (OpenCode-routed backends, Ollama for local models, Codex) are installed on demand into your own fork via /add- skills, so each install carries only what its owner asked for.
Architecturally, a single Node host process routes inbound messages from channel adapters into a per-session SQLite inbox, wakes the corresponding agent container (running the Claude Agent SDK on Bun), and delivers the agent's responses back out through the channel. Each agent group gets its own workspace, memory, instructions, and container, with flexible isolation options — a dedicated agent per channel, one agent shared across channels, or multiple channels folded into a single session. It also supports scheduled recurring tasks, web access, and reusable agent templates for stamping out pre-configured agents.
