NanoClaw

NanoClaw is a free, open-source personal AI agent built by NanoCo on top of Anthropic's official Claude Agent SDK. Instead of running one giant Node process with shared memory (the OpenClaw model), NanoClaw spins up a separate Linux container per agent group, mounts only the folders you explicitly allow, and routes credentials through OneCLI's Agent Vault so raw API keys never enter the container. The whole trunk is one process plus a handful of source files, so you can read the entire codebase in an afternoon: 28,691 GitHub stars and 122 active subscribers in 2026 say a lot of self-hosters wanted exactly that.

What NanoClaw does:

• •Container-isolated agents every agent group runs in its own Docker container (Apple Container on macOS as an opt-in, Docker Sandboxes micro-VM as an extra layer) so a compromised agent cannot read your home directory or other agents' state.
• •13+ messaging channels on demand WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, and email via Resend, each installed from the CLI by running /add-telegram, /add-whatsapp, etc. inside Claude Code. Trunk stays clean; you only carry channels you actually use, and a local nanoclaw CLI ships alongside the bash installer for direct interactive use.
• •Flexible per-channel isolation connect each channel to its own agent for hard privacy, share one agent across many channels for unified memory with separate threads, or fold multiple channels into one shared session. Set per channel via /manage-channels.
• •Scheduled jobs that talk back recurring tasks like "every weekday at 9am, summarise the sales pipeline from my Obsidian vault and send it to Telegram" are first-class, no cron file to edit.
• •Hybrid AI-native install bash nanoclaw.sh runs a deterministic scripted path (Node, pnpm, Docker, OneCLI, container build, channel pairing). When a step fails, control hands off to Claude Code automatically to diagnose and resume from the broken step.
• •Multi-provider per agent group Claude is the default via the Agent SDK, but /add-codex adds OpenAI Codex (ChatGPT subscription or API key), /add-opencode adds OpenRouter, Google, DeepSeek and others via OpenCode, and /add-ollama-provider adds local open-weight models. Each agent group can run on a different backend.
• •Per-agent CLAUDE.md and memory every group has its own CLAUDE.md, its own SQLite memory, its own container, and only the mounts you allow. Nothing crosses unless you wire it.
• •Customise by editing code, not config there are no settings files. To change behaviour you tell Claude Code ("change the trigger word to @Bob", "make replies shorter") and it modifies your fork. The codebase is small enough that this is safe.
• •Two-database session architecture inbound.db and outbound.db per session, each with exactly one writer, no IPC, no stdin piping. The host process polls outbound.db and delivers via the channel adapter.
• •Web search and fetch built in the agent-runner inside the container ships with web tools so the AI can pull live information without you wiring an MCP server.

Pricing:

• •NanoClaw itself $0, MIT licensed forever. Clone, run bash nanoclaw.sh, you're done.
• •Claude API or subscription required to actually run the agents. Anthropic's Claude API pay-as-you-go (about $3 per million input tokens for Sonnet in 2026) or a Claude Pro/Max subscription via Claude Code.
• •Optional providers bring your own OpenAI key for Codex, OpenRouter key for OpenCode, or run Ollama locally for free.

Limitations:

• •Docker familiarity required self-hosting expects basic Docker, Linux or WSL2 skills and a working Docker Desktop install. Non-technical users will struggle without a developer friend.
• •Setup needs Claude Code installed /customize, /debug, error recovery during setup, and every /add-<channel> skill run inside Claude Code, so you must install it from claude.ai/download first.
• •Windows is WSL2-only there is no native Windows runtime. You run NanoClaw inside WSL2, which adds an install layer for Windows-first users.
• •No monitoring dashboard there is no web UI for logs, metrics, or scheduled-task status. Debugging is done by asking Claude Code in chat, which is fast for AI-native users but a hard left turn if you expect a Grafana panel.
• •Trunk only accepts security and bug fixes new channels and providers must land on the channels or providers branches as skills. If you need a feature merged into mainline, expect a no.