Every agent in its own machine.
TensorBoot boots each agent run, review, and build into its own microVM — a real virtual machine with its own kernel and its own isolated network, started in milliseconds. Other vendors run background agents only in their cloud; in the customer-operated deployment TensorBoot will run them on servers you control, and if you choose the EU-hosted option we will operate those servers for you and say so plainly.
The execution fabric under the whole suite.
Agents run arbitrary code, so where they run matters. Containers share a kernel with their host; TensorBoot gives every run a hardware-virtualised machine of its own, scheduled across your servers with the primitives of a modern control plane. It's how fifty agents work in parallel overnight without any of them being able to touch each other — or anything else.
Isolation without the wait.
Isolation
Kernel-level isolation
Firecracker and Cloud Hypervisor microVMs — each sandbox gets its own kernel, not a shared one with namespaces.
Per-tenant networks
eBPF-enforced network isolation per team and per run. An agent sees exactly the endpoints its policy allows, and nothing else.
Performance & scheduling
Milliseconds, not minutes
Warm pools restore sandboxes from snapshots in ~150ms (the design target) while capacity is available; beyond it, runs queue transparently. Capacity scales with your licensed developers.
Fleet scheduling
Declarative, control-plane-style scheduling across your servers. Add a machine; the fleet absorbs it.
Deployment
Background agents, sovereign
In the self-hosted deployment, overnight and parallel agent fleets run on your hardware — never in a third-party cloud your security team didn't approve.
Air-gap on the roadmap
The design tolerates disconnection within the platform's offline grace window, with self-hosted models keeping inference on your hardware. Permanently air-gapped operation is on the roadmap.
Included with the platform.
Sandbox capacity for TensorCheck PR reviews is part of the platform subscription and scales with your licensed developers — there is no separate capacity product to buy for those review sandboxes, and in the customer-operated deployment the hardware is yours. Background cloud agents dispatched via TensorCode are billed per agent-hour, metered to the second, and free when self-hosted. See pricing for the details.
Be first in line when TensorBoot ships.
Request early access and you'll get TensorBoot early access, included in the platform price — no separate fee.