Run agents in isolated environments with persistent storage and network controls
Blaxel supplies infrastructure for AI agents that need an isolated computer, persistent working state and controlled access to external services. Its platform brings compute, storage and networking together so application developers can focus on agent behavior rather than assemble those capabilities separately. It is a development platform, not a consumer assistant that performs tasks without an application built around it.
Sandboxes run in individual microVMs and can suspend when idle while retaining the state needed for a later resume. Batch jobs provide another route for distributing many independent tasks, and MCP servers can be hosted as workloads. The official site also lists a separate Agent Runtime as coming soon; that planned product should not be confused with the currently advertised sandbox and job capabilities. Developers still need to define workload lifecycles and the conditions under which a run is considered complete.
Blaxel offers local filesystem snapshots, durable volumes and Agent Drive for shared files across agents. These options serve different purposes: restoring a sandbox is not the same as sharing a dataset among several workers. Networking controls include outbound allow-lists, dedicated egress options and a proxy that can inject credentials without placing them directly in the sandbox. A model gateway provides an integration route to external model providers.
Pricing is usage-based rather than a mandatory base subscription. Active compute, stored snapshots, images, volumes and certain optional services have separate rates. Suspended compute can stop accruing active-runtime charges while retained state still has a storage cost. Several newer network and shared-storage features are described as free during beta, which is not a permanent pricing commitment. A useful evaluation therefore measures a realistic task's active time, retained data and required concurrency together. Before production use, the application team should also test permissions, cleanup behavior and recovery from failures instead of relying only on a fast sandbox-start demonstration.
