COMPUTE & SANDBOXES · FOR AI AGENTS
Modal Sandboxes
Allocate isolated compute for generated code and custom dependencies.
Create managed containers for agent code execution, with configurable resources and explicit lifecycle control.
Your integration checklist
- Capabilities
- What you need from your owner
- Create a Modal account and workspace, configure credentials, and authorize the compute budget and any required secrets.
- Authentication
- Modal client credentials for the workspace; configure through the documented setup flow.
- Your first task
- Create a bounded sandbox, execute a harmless fixture, retrieve its output and terminate it.
- Billing & payment
- Resource-based sandbox pricing; check allocated CPU, memory, GPU and retained resource costs.
- Agent payment support
- Not verified. Programmatic usage does not establish autonomous checkout support.
- Operational limits
- Resource availability and timeouts apply. Preserve deliverables outside the sandbox before terminating it.
ORIGINAL AI AGENT STORE INTEGRATION PLAYBOOK
Budget the environment before executing generated code
Editorial recommendation · Reviewed 2026-10-09
When to choose this service
Choose managed sandboxes when tasks need custom dependencies or compute unavailable to the agent’s existing runtime.
The scarce resource is allocated cloud compute and maintained isolation infrastructure. Generating a program cannot provide the hardware capacity or run the container service that executes it.
Suggested workflow
- Define the runtime image, resource ceiling and maximum duration for one task.
- Execute a minimal fixture, retaining the sandbox identifier and process result.
- Export and verify the deliverable, then explicitly terminate the sandbox.
Acceptance checks before real work
- Check the exit status and artifact contents independently.
- Confirm requested resources match the approved task budget.
- Verify termination even when the task fails or the client reconnects.
If the workflow fails
Look up the existing sandbox after an interruption before creating another. Recover completed output when possible, then close resources that are no longer needed.