AI Agent Sandbox for Secure AI Code Execution and Automation
Secure, programmable sandboxes for AI-generated code, browser workflows, computer use, and long-running tasks.
From agent execution to evals and RL
Support the full range of AI-native workloads—from coding agents and browser automation to evaluations, reinforcement learning environments, and long-running workflows.
Coding Agents
Run agents that write, execute, and iterate on code in isolated sandbox environments.
Browser Automation
Power agents that browse the web, complete multi-step tasks, and extract structured results.
Computer Use
Enable agents to interact with full desktop environments for GUI-based workflows and human-like actions.
Evaluations & RL
Run repeatable evaluation and reinforcement learning workloads in controlled environments for testing, training, and iterative optimization.
AI-Powered CI/CD
Run code review, test generation, and validation workflows in isolated sandboxes across your CI/CD pipelines.
Long-running Workflows
Keep multi-step agent tasks running across extended sessions without breaking execution flow.
Fast startup. Strong isolation. Massive concurrency.
Launch sandbox instances in milliseconds, isolate every task, and scale to thousands of concurrent executions for real-world agent workloads.
Sub-second startup
Launch sandbox instances in under 200ms on average, optimized for bursty and high-frequency workloads.
Secure isolation
Each task runs in an isolated sandbox with system-level separation to reduce data leakage and unauthorized access risks.
High concurrency at scale
Run thousands of sandbox instances in parallel with consistently low latency for high-throughput agent workloads.
Core runtime capabilities
Everything agents need to execute code, interact with systems, and stay in control across real-world workflows.
Code Execution
Run Python, JavaScript, C++, and other languages in a secure sandbox.
External APIs
Allow agents to access external APIs and online data as needed.
Browser Use
Automate web navigation, form filling, and data extraction inside the sandbox.
Computer Use
Control full desktop environments for typing, clicking, scrolling, screenshots, and other GUI actions.
Persistent Sessions
Pause and resume long-running agent tasks without losing state.
Live Session View
View and stream sandbox sessions with VNC for interactive monitoring and debugging.
Run agents in a few lines of code
Install the SDK, set your API key, and start a sandbox for a coding agent or a full desktop.
pip install novita-sandbox
npm i novita-sandbox
1import os
2from novita_sandbox import Novita
3
4novita = Novita(api_key=os.environ["NOVITA_API_KEY"])
5
6# The codex template ships with Codex pre-installed
7sandbox = novita.sandbox.create(
8 "codex",
9 timeout=3600,
10 envs={"CODEX_API_KEY": os.environ["CODEX_API_KEY"]},
11)
12
13try:
14 # Run Codex non-interactively: take a prompt, print the result, exit
15 execution = sandbox.commands.run(
16 'codex exec --full-auto --skip-git-repo-check "Write a FizzBuzz script and run it"',
17 on_stdout=lambda data: print(data, end=""),
18 timeout=0,
19 )
20 print(execution)
21finally:
22 sandbox.kill()
231from novita_sandbox import Novita
2
3novita = Novita() # reads NOVITA_API_KEY from the environment
4
5# Start an isolated virtual desktop for your computer use agent
6desktop = novita.desktop.create()
7desktop.stream.start()
8
9# Interactive VNC URL for your agent UI
10print(desktop.stream.get_url())
11
12# View-only URL for monitoring
13print(desktop.stream.get_url(view_only=True))
14
15# Clean up when the task is done
16desktop.stream.stop()
17desktop.kill()
18Run Codex non-interactively in an isolated sandbox: it reads and edits code, runs commands, and returns the result.
Codex guideFlexible, Usage-Based Pricing
Per-second billing based on vCPU and memory allocation. No plans, no lock-ins, no hidden costs. Learn more
vCPU
Calculated per allocated vCPU-second
Memory
Calculated per allocated GiB-second
Storage
Measured hourly and charged daily. First 60 GB included
Usage-Based Pricing Examples
1 vCPU + 512 MiB RAM
Short-lived agent task (5 min)
2 vCPU + 1 GiB RAM
Code execution job (1 hr)
8 vCPU + 8 GiB RAM
Multi-agent or RL workload (1 hr)
* Free tier quota: 5 concurrent sandboxes, 1-hour max session. Top up your balance and use the sandbox to unlock paid tier (100 concurrent, 24-hour session) automatically.
Start building with Agent Sandbox
Get started with quickstarts, reusable templates, and isolated sandbox storage.
File system
Each sandbox includes an isolated filesystem for code, data, and session state.
Enterprise
For advanced workloads that need higher resource limits, custom deployment regions, or dedicated infrastructure.
Frequently asked questions
What is Novita Agent Sandbox?
Novita Agent Sandbox is a secure cloud runtime for AI agents. Each sandbox is an isolated environment where an agent can execute code, use a browser, control a desktop, and run long-running tasks without touching your own machine or other workloads.
Which coding agents can I run in a sandbox?
Claude Code and Codex have ready-made templates with the agent pre-installed, so you can create a sandbox and run the agent headless in a few lines of code. You can also install any other agent or tool in a custom template.
Does Agent Sandbox support computer use?
Yes. A desktop sandbox gives your computer use agent a full virtual desktop for typing, clicking, scrolling, and screenshots. You can stream it over VNC as an interactive session or as a view-only feed for monitoring.
How fast do sandboxes start, and how are they isolated?
Sandboxes start in under 200ms on average. Every task runs in its own sandbox with system-level separation, which reduces the risk of data leakage and unauthorized access between tasks.
Which languages and SDKs are supported?
Agents can run Python, JavaScript, C++, and other languages inside a sandbox. You control sandboxes with the novita-sandbox SDK for Python 3.9+ and Node.js 20+, or from the terminal with the Novita Sandbox CLI.
Can a sandbox keep state for long-running tasks?
Yes. Persistent sessions let you pause and resume long-running agent tasks without losing state, and each sandbox has its own isolated filesystem for code, data, and session files.
How is Agent Sandbox priced?
You pay per second for the vCPU and memory you allocate, with no upfront commitment. New accounts start on a free tier; top up your balance to unlock the paid tier with more concurrent sandboxes and longer sessions. See the pricing section above for current rates.
Everything you need to build production AI.
200+ models, on-demand GPUs, and secure agent runtimes — unified under one API. Free to start, scales as you grow.