Andrej Karpathy recently described a shift worth paying attention to: just as LLM agents emerged as a new layer on top of LLMs, Claws are the next layer on top of agents, taking orchestration, scheduling, context, tool calls, and persistence further than agents alone ever could.
That layer needs infrastructure that stays on. For developers running OpenClaw on Novita Sandbox, that meant working around session limits until now.
Novita OpenClaw CLI closes that gap. It’s a new deployment tool on Novita Agent Sandbox that provisions a fully configured OpenClaw instance in a single command, with no runtime cap, no manual restarts, and no lifecycle management overhead.
The Problem the CLI Solves
Novita Agent Sandbox has supported OpenClaw deployment through the Sandbox Template. That path is suitable for exploration, but it carries a hard constraint: sandboxes spun up through the UI have a session lifecycle limit. After a few hours, the sandbox stops, and the agent goes offline. Developers who needed continuous availability had to work around this themselves — scheduling pauses, scripting restarts, or manually reviving instances.
Sandboxes created through novita-openclaw-cli operate under a different model entirely. There is no session cap. The instance runs 24/7 for as long as you need it. The default spec — 4× CPU, 8 GB memory — is designed for real workloads, not just demos. And because the model layer is fully configurable, you choose which LLMs power your agent, whether hosted on Novita or brought from a third-party provider.
Getting Started
Prerequisites
- Python installed
- A Novita API key — here’s how to get one:
- Log into novita.ai — sign in with Google or GitHub (a new account is created automatically on first login), or sign up with your email address
- Create an API key — go to the Key Management settings page to create or manage your API keys.
Install and Launch
pip install novita-openclaw-cli
# Launch instance novita-openclaw-cli launch --api-key YOUR_API_KEY # Or use environment variable export NOVITA_API_KEY="your-api-key" novita-openclaw-cli launch
On success, the CLI returns three values you’ll use to access and manage your agent:
- Web UI URL — the browser interface for your OpenClaw instance
- WebSocket URL — for programmatic or agent-to-agent access
- Gateway Token — your authentication credential

Open the Web UI URL and navigate to the Chat tab. Your agent is live.
Configuring Models
Your instance comes pre-configured with two Novita-hosted models:
| Model | Role |
|---|---|
moonshotai/kimi-k2.5 | Primary |
minimax/minimax-m2.5 | Fallback |
To change the models your agent uses, navigate to Settings → Config in the Web UI, switch to Raw JSON view, and update two sections:
Step 1: Register the model under your provider
Add a new object to the models array inside models.providers.novita:
"models": [
{
"id": "model-id",
"name": "display name",
"reasoning": true,
"input": ["text"],
"contextWindow": 200000,
"maxTokens": 8192
}
]
Step 2: Set it as primary or fallback
Update the model field under agents.defaults to reference your model using the provider/model-id format.
"agents": {
"defaults": {
"model": {
"primary": "novita/moonshotai/kimi-k2.5",
"fallbacks": [
"novita/minimax/minimax-m2.5"
]
}
}
}
Click Update to save. Every LLM available on the Novita platform is supported. Third-party providers can also be configured — when you bring your own LLM, you pay only for sandbox runtime, not Novita model usage.
Connecting Channels
OpenClaw supports external messaging channels so your agent is reachable outside the Web UI. Channels require configuration — they are not active by default.
Telegram
Step 1: Create a Telegram Bot
- Open Telegram and search for @BotFather
- Send
/newbotand follow the prompts to name your bot - Copy the bot token BotFather provides

Step 2: Send the bot token to Chat tab
In the Web UI, go to the Chat tab and send the bot token.
Step 3: Pair and use Telegram Bot
In Telegram, open the bot you created and send /start. Copy the code and paste it in the Chat tab.

Once connected, messages sent to your Telegram bot are routed directly to your OpenClaw agent. More channel integrations are supported — see the official OpenClaw channels guide for the full list and setup instructions.
Command Reference
launch
Create a new instance.
novita-openclaw-cli launch [OPTIONS]
| Option | Description |
|---|---|
--api-key | Novita API Key |
--gateway-token | Custom gateway token (optional, auto-generated by default) |
--timeout | Creation timeout in seconds (default: 60) |
list
List all active instances.
novita-openclaw-cli list --api-key YOUR_API_KEY
status
Check instance status.
novita-openclaw-cli status <SANDBOX_ID> --api-key YOUR_API_KEY
stop
Terminate an instance.
novita-openclaw-cli stop <SANDBOX_ID> --api-key YOUR_API_KEY
tui
Connect via terminal interface.
novita-openclaw-cli tui <SANDBOX_ID> --token <GATEWAY_TOKEN>
Programmatic Usage
For AI agents or automation scripts, use -j flag for JSON output:
novita-openclaw-cli -j launch --api-key YOUR_API_KEY
JSON output structure:
{
"success": true,
"data": {
"sandbox_id": "xxx",
"webui": "https://xxx.sandbox.novita.ai?token=xxx",
"gateway_ws": "wss://xxx.sandbox.novita.ai",
"gateway_token": "xxx"
}
}
Environment Variables
| Variable | Description |
|---|---|
NOVITA_API_KEY | API Key, eliminates need for --api-key flag |
OPENCLAW_GATEWAY_TOKEN | Preset gateway token |
Pricing
| What you use | What you pay |
|---|---|
| Sandbox runtime | Sandbox rates |
| Novita-hosted LLMs | Model rates |
| Third-party LLMs | Sandbox runtime only |
Troubleshooting
Command not found
# Use full path ~/.local/bin/novita-openclaw-cli launch # Or on macOS ~/Library/Python/3.x/bin/novita-openclaw-cli launch
Connection timeout
Increase timeout: --timeout 180
Build the Next Layer
The shift Karpathy describes — from agents to Claws — isn’t theoretical. It’s an architectural opportunity available right now, and the teams who build on persistent, always-on infrastructure will be the ones who realize it.
Novita OpenClaw CLI gives you that foundation: a production-grade agent runtime, deployed in under a minute, with full control over the model layer and channel integrations. The sandbox handles uptime. The CLI handles provisioning. What you build on top is up to you.
Start building on Novita Agent Sandbox →
Novita AI is an AI and Agent Cloud for builders and startups — providing access to 200+ LLMs, image, video, and TTS models via one API, alongside GPU instances and Agent Sandboxes that launch in seconds.
Discover more from Novita
Subscribe to get the latest posts sent to your email.





