Sandboxes for
AI agents.
Let agents do the work.
Decide what they can reach.
An agent needs a workspace.
Your systems need boundaries.
Running code is only part of the job. The harder questions are what it can access, which credentials it gets, and where the work happens.
Too much network access.
An agent calling one API shouldn't need an open path to every other service.
Start closed. Allow what matters.
Oatbox uses operator-configured network policies enforced outside the sandbox. Approved destinations are reachable; everything else stays denied.
Credentials inside the workspace.
A long-lived API key in an environment variable gives generated code a secret it can read and copy.
Keep secrets at the boundary.
Scoped credential bindings let the configured egress gateway attach credentials to requests, rather than putting the underlying secret in the sandbox.
No clear execution location.
Where your agent runs matters when you are working with customer data or internal services.
Choose where the work runs.
Restrict new sandbox placement to a region. Execution location is explicit, rather than an infrastructure detail you have to guess.
Regional execution is not a data-residency guarantee. APIs, logs, backups, and snapshots each need their own location controls.
What is your agent
allowed to do?
Tell us what you're building, the services your agents need, and the boundaries you want to set.
Join early accessOatbox is in early development.
For evaluation, not production or sensitive workloads.