Simple, secure sandbox infrastructure for AI

From individual agents to large-scale rollouts, Syva handles it all.

Examples

CodeBackground jobsBatch RLPreviewsEvalsSnapshots

sandbox lifecyclerunning

define

image, CPU, memory, ports

hydrate

repo, files, dependencies

run

commands, agents, services

inspect

stdout, preview, artifacts

previewready

http://localhost:5173

live port attached to sandbox

agent-run.pystreaming

from syva import Sandbox

sandbox = Sandbox.create(
    image="python-3.12:base",
    cpu_cores=2,
    memory_mb=4096,
    ports=[5173],
)

result = sandbox.run_command(
    "python",
    args=[
        "rollout.py",
        "--episodes=128",
    ],
)

print(result.stdout)
sandbox.destroy()

> $ syva sandbox create --image python-3.12:base

> sandbox sbx_8f42 ready in 1.7s

> $ python eval.py --branch reward-model

> port 5173 exposed

> stdout: 128 episodes complete

New · Persistent sandboxes

Stop, resume, and fork sandboxes

Stopping a sandbox snapshots its filesystem. Resume it by name with everything intact, or fork a prepared environment into identical copies in seconds. Idle sandboxes bill storage at $0.08/GiB-month on the compressed delta — not compute.

Read the snapshots guide

sandbox = Sandbox.get_or_create(name="agent-workspace")
sandbox.stop()    # filesystem snapshotted
sandbox.start()   # resumes where it left off
sandbox.fork()    # identical copy, warm caches

Coding agents

Give interactive agents a clean repo, real commands, preview ports, and a filesystem they can change freely.

Background agents

Let longer jobs run away from your laptop with scoped secrets, mounted context, and inspectable output.

RL rollouts

Fan out isolated environments for evals and rollouts, then pull back traces, logs, and artifacts through one API.

Built with Syva

A few concrete ways teams use sandboxes as the execution layer for agent work.

All examples

Run OpenClaw in a sandbox

Install the agent CLI, pass scoped credentials, and collect JSON output from one clean run.

Batch RL coding rollouts

Fan out one sandbox per task, run the reward check, and bring traces back through the API.

Replace a CI lane

Point GitHub jobs at short-lived Syva runners with job-scoped files, logs, and usage.

Pay for what you use.

CPU and memory are metered by the second.

Estimate

$0.040/min

20 running sandboxes. Each has 2 vCPU and 4 GB memory.

CPU$0.027/min

Memory$0.013/min

Per sandbox$0.002/min

Total minute$0.040/min

Sandboxes20

20 running

vCPU per sandbox2

<$0.001/vCPU-min

Memory per sandbox4

<$0.001/GB-min

Run state stays visible.

Keep the useful parts of every sandbox: status, output, previews, and traces.

coding agentrunningoutputstdout

background fixreadyoutput:5173

rl rollout batchstreamingoutputtraces

stdout, files, preview URLs, and usage stay attached to the run after the sandbox exits.

Run your first workload in minutes.

Proudly built in Europe, hosted in European datacenters, and powered by 100% renewable energy.