Cua
By Cua AI
YC-backed computer-use infrastructure, MIT-licensed with around 26,000 GitHub stars, that provisions disposable Linux, Windows, macOS, and Android desktops for AI agents to operate through screenshots and clicks, alongside a local driver for automating native apps on your own machine and small specialist models for scoring interface decisions; the open-source driver and local sandbox are free, hosted cloud desktops bill by vCPU and memory per hour.
Best for
- running computer-use agent benchmarks and evaluations at scale
- giving an agent a disposable desktop instead of your own machine
- automating native desktop apps on macOS, Windows, or Linux
- generating verified training data from real GUI interactions
- testing an agent across multiple operating systems in parallel
- Sep 23, 2026Added to the sandbox category: Y Combinator-backed computer-use infrastructure, MIT-licensed, roughly 26,000 GitHub stars, that provisions disposable Linux, Windows, macOS, and Android desktops for agents to operate through screenshots and clicks, with a free open-source driver and local sandbox, and usage-billed hosted cloud desktops. Sandbox was the thinnest category in the catalogue before this addition.
Other Sandboxes & code execution
E2B
Sandboxed cloud VMs purpose-built for AI agents — spin up a Linux environment in milliseconds, run untrusted code, throw it away.
Vercel Sandbox
Firecracker-based microVMs from Vercel for running untrusted code at the edge. Strong fit if you're already deploying on Vercel.
Modal
Serverless compute platform built for AI workloads — GPU jobs, ephemeral functions, and code sandboxes triggered from agents or webhooks.
Daytona
Cloud dev environments designed for AI agents. Each agent gets its own isolated, reproducible workspace it can install tools into.
Runloop
Secure code-execution sandboxes for AI agents with evaluation tooling and snapshot/checkpoint support, built for enterprise scale.
Blaxel
Infrastructure platform for autonomous agents using microVMs that boot in milliseconds and resume in ~25ms with $0 idle cost.
Northflank
Deployment platform offering secure sandboxes and microVMs (200ms spin-up) for running untrusted and AI-generated code at scale.
Beam
On-demand AI compute platform for running agent sandboxes, task queues and GPU workloads, with bring-your-own-cloud support.