Coming from NanoClaw
Coming from NanoClaw
NanoClaw and RemoteClaw both connect AI agents to messaging channels, but they take fundamentally different approaches. If you’re coming from NanoClaw, here’s what changes.
The key difference
NanoClaw builds a new agent using the Anthropic Agent SDK and runs it in containers. RemoteClaw bridges your existing CLI agent as a subprocess — your ~/.claude, ~/.gemini, or other agent config is preserved as-is.
What you gain
- Multi-runtime support. NanoClaw is Claude-only via the Anthropic Agent SDK. RemoteClaw supports 4 CLI runtimes: Claude Code, Gemini CLI, Codex CLI, and OpenCode. Switch runtimes via config, no rebuild needed.
- 20+ messaging channels. NanoClaw supports WhatsApp, Telegram, Slack, Discord, and Gmail. RemoteClaw inherits OpenClaw’s full adapter library.
- Your existing config. RemoteClaw uses your
~/.<agent>directly. No need to reconfigure your agent inside a new framework. - Subscription compatibility. RemoteClaw spawns the CLI binary, so your existing Claude subscription or API keys work as-is.
- 50+ MCP tools. Infrastructure-bound tools for sessions, messaging, cron, gateway, and more.
What you trade off
- Container isolation. NanoClaw runs agents in containers with strong sandbox boundaries. RemoteClaw uses process isolation — the agent runs as a subprocess on your host. If container-level sandboxing is a hard requirement, NanoClaw may be the better fit.
- Bring-your-own-agent not required. NanoClaw provides a pre-configured Claude agent. RemoteClaw requires you to have a CLI agent installed and configured first.
Architecture side by side
RemoteClaw (middleware)├── ChannelBridge (message routing + session tracking)├── CLI subprocess (claude, gemini, codex, or opencode)│ ├── LLM interaction (handled by the CLI)│ ├── Tool execution (handled by the CLI)│ └── Conversation management (handled by the CLI)└── MCP server (injected into subprocess for gateway access)The CLI agent owns the agentic loop. RemoteClaw handles session persistence, message delivery, and MCP tool bridging.
NanoClaw (containerized agent)├── Anthropic Agent SDK (Claude-only orchestration)├── Container isolation (container-based sandboxing)├── Built-in tool use (via Agent SDK)└── Channel adapters (WhatsApp, Telegram, Slack, Discord, Gmail)NanoClaw runs a new Claude agent inside a container. The agent is built from the Anthropic SDK, not from your existing CLI setup.
Key differences at a glance
| Dimension | RemoteClaw | NanoClaw |
|---|---|---|
| Agent model | CLI subprocess (your agent, your config) | Anthropic Agent SDK (new agent instance) |
| Runtime support | 4 runtimes (Claude, Gemini, Codex, OpenCode) | Claude only |
| Config | Uses your ~/.<agent> directly | Agent configured within NanoClaw |
| Authentication | Your existing CLI subscription or API keys | API keys (Anthropic policy for Agent SDK) |
| Isolation | Process isolation (subprocess boundary) | Container isolation (container boundary) |
| Channels | 20+ (inherited from OpenClaw) | WhatsApp, Telegram, Slack, Discord, Gmail |
| MCP tools | 50+ infrastructure-bound tools | Agent SDK built-in tools |
Anthropic restricts using the Agent SDK with consumer subscription plans (Max/Pro). This affects any project built on the Agent SDK, including NanoClaw. RemoteClaw spawns the CLI binary, which works with your existing subscription or API keys — whichever you have configured.
Next steps
- Get started with RemoteClaw — install and configure
- Personal Assistant Setup — full setup guide
- Middleware Architecture — how RemoteClaw’s architecture works
If RemoteClaw isn’t the right fit, NanoClaw is a great choice if you want a lightweight, security-hardened Claude agent with container isolation and minimal setup.