AGENT SANDBOX
Let your agent write code, run commands, and work with files. Give every task a place to run — and a lifecycle you control.
ILLUSTRATIVE LIFECYCLE
Choose a template. Open a workspace.
Execute commands. Work with files.
Bring the results into your app.
End the sandbox when work is done.
WHY SANDBOX
Agents turn ideas into code. Give that code a dedicated environment, a clear task, and an explicit end — so execution becomes part of your product.
Give generated code a dedicated place to work. Send in the files it needs, run the task, and bring the results back to your application.
Choose the network controls your task needs. Available settings and enforcement depend on the selected provider and template.
Create a sandbox for a task, collect the results, then shut it down. Pause and resume where supported when the work needs more time.
COMPARE
Familiar building blocks, a different workflow. A sandbox gives agent applications an API for the environment where their work happens.
Compare environments
| At a glance | Cloud PC | Virtual machine | Docker | Agent sandbox |
|---|---|---|---|---|
| Typical user | People | Ops & engineers | Developers | Agent applications |
| Workspace | Interactive desktop | Whole machine | Container | Task environment |
| Lifecycle | Usually persistent | Managed by you | Managed by you | API-managed |
| Interface | Screen & keyboard | Machine & shell | Container tools | SDK & API |
| Your focus | Remote work | Infrastructure | Packaging code | Running agent tasks |
PRODUCT VS CHASSIS
A finished AI product gives you a ready-made workflow. An execution layer gives you the building blocks to make the experience your own.
FINISHED PRODUCT
SANDBASE
WHERE IT FITS
Keep the product, the agent loop, and the runtime working together. SandBase connects your application to sandbox execution through a familiar interface.
The experience your users see. Your workflows, your interface, your choice of model.
Connect model decisions to tools, execution, and the next step in a task.
An E2B-compatible gateway for creating and managing execution environments.
The runtime and isolation technology behind the selected provider and template.
USE CASES
Different teams, one pattern: give your agent a dedicated place to turn a plan into an output.
01 / AI product teams
Give code execution its own environment. Turn generated scripts into useful results for your application.
02 / Platform engineers
Connect sandbox lifecycles to your existing workflow with a familiar SDK and your SandBase API key.
03 / Solo builders
Try a command, work with files, and collect an output without building the execution plumbing from scratch.
SECURITY BOUNDARIES
A sandbox is part of your security design. Isolation, network controls, and persistence depend on the selected provider and template.
Transfer only the files a task needs. A working directory alone is not a security control.
Check the network controls your provider and template support before running sensitive workloads.
Collect outputs and stop environments when work is complete. Use timeouts as part of your workflow.
Review template privileges and isolation guarantees. Do not assume every runtime has the same defaults.
FAQ
The things to know before your agent runs its first command.
A dedicated execution environment where an agent can run commands and work with files. Your application creates it, sends in work, collects results, and ends the environment when the task is complete.
Docker is a building block for running containers. SandBase provides an E2B-compatible gateway for managing sandbox lifecycles and accessing commands and files through an SDK. The underlying isolation depends on the provider and template.
Create a fresh environment for independent tasks and collect the results before stopping it. For longer work, use pause and resume where the selected provider supports them. Confirm state-retention behavior before relying on persistence.
Yes. Follow the Sandbox quickstart to configure the SDK for SandBase, select a template, and run your first command using your SandBase API key.
Sandbox execution is separate from model selection. Your application chooses the model and sends the resulting commands or files to the sandbox.
No. A directory path or a prompt instruction does not enforce isolation. Review the selected provider’s runtime, networking, template permissions, and data-handling behavior for your workload.
See the current quickstart and your workspace terms for availability and pricing. Check the resources and provider you plan to use before running your workload.
