Give every agenta computer it cansafely control.
Run code, browsers, desktop GUIs, and long-lived agent sessions in isolated cloud environments that start fast, preserve useful state, and disappear cleanly.
TypeScript SDK
One environment per task
import { StateSetSandbox } from '@stateset/sandbox-sdk';
const client = new StateSetSandbox({
baseUrl: 'https://api.sandbox.stateset.app',
authToken: process.env.STATESET_SANDBOX_KEY,
});
const sandbox = await client.create({
cpus: '2', memory: '4Gi', timeout_seconds: 300
});
const result = await client.execute(sandbox.sandbox_id, {
command: 'npm test'
});Runtime
Node 22
Isolation
gVisor
Output
Streaming
State
Checkpointable
Start in milliseconds
Claim a pre-scheduled environment from the warm pool without putting Kubernetes scheduling in the agent loop.
Redis-backed warm pools • automatic cleanup • configurable lifetimes
Run untrusted work safely
Choose the isolation boundary for each workload while secure defaults constrain users, capabilities, filesystems, and networks.
gVisor • Kata • Firecracker • WASM • network policy
Give agents a real computer
Execute code, automate browsers and desktop GUIs, keep REPL state, move files, and expose preview ports through one API.
Commands • computer use • REPLs • files • tunnels
More than a shell in a container
The primitives an agent needs to act, observe, recover, and hand off work are part of the platform—not a pile of infrastructure you assemble around it.
Code execution
Run commands with captured or real-time SSE output in configurable, ephemeral environments.
Computer use
Capture screenshots and control Chromium, Firefox, mouse, and keyboard through a streaming desktop.
Stateful REPLs
Keep variables and imports alive across Python, JavaScript, TypeScript, Julia, and R calls.
Checkpoints
Capture, restore, clone, and compare workspace state without forcing an agent to start over.
Controlled networking
Apply domain and port egress rules, proxy model access, and publish token-gated preview URLs with TTLs.
MCP-native tools
Start, inspect, and stop Model Context Protocol servers inside the environment as part of the agent session.
Fast for the agent. Controlled for the operator.
Every environment moves through one observable lifecycle, from authenticated claim to execution, persistence, and clean recycling.
Provision the right boundary.
1. Create
Request CPU, memory, lifetime, template, and isolation. A warm environment can be claimed without waiting for a cold pod schedule.
- Select container, gVisor, Kata, Firecracker, or WASM isolation.
- Enforce plan limits, concurrency budgets, RBAC, and rate limits.
- Return a sandbox ID and expiration time to the calling agent.
Control flow
One runtime, from a test command to a full desktop
Use the same lifecycle, security, and observability model across coding, browser automation, research, and persistent autonomous work.
Coding agents
The workload
Coding agents need a fast place to clone repositories, install packages, edit files, run tests, and inspect failures without touching developer machines.
The runtime
Give every task an isolated workspace with streaming execution, structured file tools, checkpoints, and enforced resource limits.
What changes
- Parallelize tasks without workspace collisions.
- Checkpoint before risky migrations or dependency changes.
- Stream test and build output directly into the reasoning loop.
Agent infrastructure should not be an integration project
A production sandbox needs more than command execution. It needs a security boundary, recoverable state, agent-native interfaces, and a platform contract operators can trust.
Let agents act without giving untrusted code the keys.
Launch a secure cloud computer for every task, keep the state that matters, and enforce the boundaries your operators require.
Isolation by workload
Five runtime boundaries from containers to microVMs
Operational by default
RBAC, quotas, metering, logs, metrics, and tracing
Work that survives
Checkpoints, artifacts, cloning, and reattachment