Choose Your Sandbox
Compare the five sandboxes on agentstead.dev and find the right one for your workflow.
Not sure where to start? Pick OpenClaw — it's the most general-purpose sandbox and works well for almost every workflow.
Every sandbox gives you the same platform foundation: a persistent Linux desktop, public HTTPS URL, secrets management, backups, and optional Tailscale networking.
The difference is the agent daemon that runs alongside the desktop.
Pick your sandbox
OpenClaw
Full desktop + OpenClaw gateway. Best for Claude Code, Codex, OpenCode, and general-purpose AI coding workflows.
OpenFang
Rust-based agent with REST API surface. Works without a provider key (Ollama fallback). Includes ffmpeg.
NullClaw
Zig-based sandbox with a minimal footprint. Choose this if you specifically want the Zig execution model.
ZeroClaw
Rust-based sandbox with pairing-code desktop auth. Performance-oriented.
NanoBot
Ultra-lightweight Python agent. Connect to Telegram, Discord, Slack, and more. MCP and ClawHub skills included.
Quick comparison
| OpenClaw | OpenFang | NullClaw | ZeroClaw | NanoBot | |
|---|---|---|---|---|---|
| Type | Node.js gateway | Rust binary | Zig binary | Rust binary | Python agent |
| Port | 18789 | 4200 | 42618 | 42617 | 18790 |
| Token type | Gateway Token | API Key | Session Token | Session Token | Gateway Token |
| RAM | 4 GB | 4 GB | 4 GB | 4 GB | 2 GB |
| Channels | — | — | — | — | Telegram, Discord, Slack, … |
| MCP support | — | — | — | — | Yes |
| Best for | General AI coding | API-oriented + media | Minimal/Zig | Rust workloads | Channel bots |
What every sandbox includes
Every AgentStead sandbox ships with an identical platform layer. See Pre-installed Tools for setup guides.
Hardware profile
| Sandbox | RAM | vCPU |
|---|---|---|
| OpenClaw | 4 GB | 2 |
| OpenFang | 4 GB | 2 |
| NullClaw | 4 GB | 2 |
| ZeroClaw | 4 GB | 2 |
| NanoBot | 2 GB | 1 |
Switching sandboxes
There's no in-place conversion between sandbox types. To switch:
- Create a backup of your current agent's workspace
- Deploy a new agent of the target type
- Download and re-upload any workspace data you need
- Destroy the old agent when done