# OpenHands tasks in a sandbox

Coming soon. OpenHands tasks are in development and not yet tested in a live sandbox. They are not available yet.

## Use case: delegate a bounded task

Once enabled, ask OpenHands to add tests for one module under a fixed per-task AI budget and workspace spending caps. Tasks will need a supported workspace AI provider. Tasks are planned for OpenAI-compatible chat providers. Native Anthropic and Responses providers will not be supported. The design keeps provider keys outside the sandbox. A task is meant to stop when its budget runs out. If a call’s usage is unknown, Runzivo plans to count the full reserved amount. Review changes and test results before keeping them.

## Task controls after enablement

Planned workflow: start the sandbox, open its Agent panel and select Start Agent Server. When task access is available, enter a prompt and select Submit task. Watch events, send a follow-up, pause, resume or cancel. Pause will wait for the current model call. Cancel will stop the server. Closing the panel will leave the task running. Restart the server for another task after completion.

## Access and lifetime

Submit, steer, pause and resume will need sandboxes:write and ai_chat:use. Only the task owner will steer, pause or resume a task. Sandbox writers will be able to cancel it. Task events will need sandboxes:write. Readers will see status only. Anyone able to run sandbox commands, including through sudo, will have the active task’s AI access up to its budget. Share command access only with trusted people and code. Planned limits: sessions last at most one hour and end when the sandbox stops or expires.

## Billing

No price has been announced for OpenHands tasks. Existing sandbox compute billing still applies. Installed Codex and Claude Code are separate tools with their own authorized logins.

The panel will not estimate remaining spend.

## Related pages and language

[Français (Canada)](https://runzivo.dev/docs/sandboxes-openhands-fr-ca/)

On this page and page tools