CelestoAI/celesto
Sandboxes for coding agents, CI runners and RL training
What it solves
Celesto provides isolated, secure computing environments for AI agents to execute code, browse the web, and interact with desktop applications. By running these tasks in lightweight virtual machines rather than directly on your host system, it creates a strong security boundary for untrusted code while ensuring agents have a consistent, persistent workspace.
How it works
Celesto uses lightweight virtual machines that start in approximately 500 milliseconds. You can manage these "computers" via a CLI or Python SDK, either locally on your machine or in the Celesto Cloud. The system supports various runtimes, including minimal environments for code execution, browser sessions for web automation, and full desktop environments for GUI-based tasks. It allows for file sharing between your host and the virtual machine, network restriction policies, and state persistence across sessions.
Who it’s for
It is designed for developers building AI agents, CI/CD pipelines, or reinforcement learning workflows who need a safe, reproducible, and fast way to give their agents access to a computer.
Highlights
- Fast startup: Virtual machines boot in about 500ms without needing to pull images every time.
- Strong isolation: Uses virtual machines to separate agent tasks from your host system.
- Flexible runtimes: Supports Linux, Windows, and macOS environments, including browser and desktop sessions.
- Consistent API: Use the same Python code for local development and cloud-based production workloads.
- Network control: Options to restrict outbound internet access or allow specific IP ranges.
- Persistent state: Files and environment state can be saved and reused across sessions.
Related
- Project
- Project
- Project
- Project
- Project