NanoClaw

Core Positioning#

NanoClaw is a minimalist alternative to OpenClaw, delivering full personal AI assistant capabilities in roughly 3,900 lines of code (15 source files). Targeted at individuals or small teams, it provides no Web UI or admin dashboard—all interaction happens through messaging channels or Claude Code.

Security & Isolation Model#

• Container Isolation: Agents run inside Docker containers with filesystem-level OS isolation, not application-layer permission checks
• Optional Isolation: Docker Sandboxes micro-VM isolation, macOS native Apple Container (details TBD)
• Credential Safety: Agents never hold raw API keys; outbound requests inject credentials via OneCLI Agent Vault with per-agent policy and rate limiting

Multi-Channel Messaging#

13+ messaging channels installed on demand: WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, Email (Resend). Channel adapters installed via /add-<channel> skills.

V2 Flexible Isolation: Each channel can connect to an independent agent (full privacy), a shared agent (unified memory), or fold into a single shared session.

Agents & Workspaces#

• Each agent group has its own workspace: CLAUDE.md, memory, container, mount points
• Multi-Model: Native Claude Code (Anthropic Agent SDK); optional /add-codex (OpenAI Codex), /add-opencode (OpenRouter/Google/DeepSeek), /add-ollama-provider (local open-source models). Providers configurable per agent group
• Scheduled Tasks: Set recurring Claude tasks that auto-reply results via messaging channels
• Web Access: Agents can search and scrape web content

Skills System#

The trunk only publishes a registry and infrastructure. Channel adapters and alternative providers are installed on demand into user forks via /add-<name> skills, with self-registration at startup.

Architecture#

Data Flow:

Messaging App → Host Process (Router) → inbound.db → Container (Bun, Claude Agent SDK) → outbound.db → Host Process (Delivery) → Messaging App

Core Decisions:

• Single-Process Host: One Node.js process orchestrates agent containers for all sessions—no microservices, no message queues
• Dual SQLite Model: Two SQLite files per session (inbound.db / outbound.db), each with a single writer, avoiding cross-mount contention and IPC overhead
• Entity Model: User → Messaging Group → Agent Group → Session
• 60s Sweep: Host process sweeps every 60 seconds for expiration, due message wake-up, and recurring tasks

Key Modules:

AI-Native Experience#

• Zero Config Files: No YAML/JSON/TOML configs—all customization via Claude Code directly modifying code
• Hybrid Install: nanoclaw.sh auto-installs deps, registers credentials, builds containers, pairs channels; on failure, hands off to Claude Code for diagnosis and recovery
• Self-Diagnostics: /debug command lets Claude Code diagnose issues itself
• Default trigger word @Andy, modifiable via Claude Code

Requirements#

• macOS / Linux / Windows (WSL2)
• Node.js 20+, pnpm 10+, Docker Desktop or Docker Engine, Claude Code
• Primary language: TypeScript (89.3%), Shell (9.1%)
• MIT License (Copyright (c) 2026 Gavriel, year TBD)

Scale Comparison#

Unconfirmed Items#

• OneCLI Agent Vault independent repo/docs link not specified
• Apple Container technical details and availability not detailed
• v1 → v2 migration details not elaborated
• Media references on official site (VentureBeat, Fortune, etc.) not individually verified
• Actual maturity and test coverage of each channel adapter not stated