Performance · Benchmarks

Among the fastest sandboxes to boot — and it shows.

ComputeSDK runs an independent, public benchmark on Time to Interactive (TTI) — the time from creating a sandbox to running your first command inside it. It's the number that decides whether an agent waits or just goes. And it isn't a single cold start: TTI is measured across three launch patterns — sequential, staggered (200ms apart), and a concurrent burst — at 100 sandboxes each, so the numbers reflect real load, not a lucky one-off.

Tensorlake sits in the top tier — trading the #1 spot day to day — booting a sandbox to interactive in well under a second, with a 100% success rate. The leaderboard shifts daily; see today's standings on the ComputeSDK sandbox benchmarks.

Deep dive · Multi-cloud scheduling

How we start fast, everywhere — the scheduler behind the numbers.

Those startup numbers aren't luck — they're the scheduler. We wrote up how Tensorlake places sandboxes across clouds and regions, keeps capacity warm, and routes each request to the fastest path to launch — the architecture behind consistent, low-latency starts at scale.

Read it on the Tensorlake blog.

Dev workflow · Native SSH

Use a sandbox as your dev machine.

Sandboxes now support native SSH and interactive PTY sessions — a real terminal into your sandbox, plus everything that comes with it: VS Code Remote-SSH, port forwarding, long-running shells. Spin up a fast, disposable, fully isolated dev environment in seconds and code in it like it's local.

Read about SSH and PTY sessions and remote dev in our docs.

Images · OCI support

Bring any OCI image.

Sandboxes now support OCI images. Point us at a container image you already build and ship, and run it as a Sandbox — no rewriting your environment, no bespoke image format. Your existing tooling and dependencies come along for the ride.

See custom images in our docs.

Terminal agents · Harbor

A dedicated guide for running Harbor on Tensorlake.

Tensorlake is an official Harbor environment runtime for evaluating CLI agents. There's now a dedicated, step-by-step page showing exactly how to wire Sandboxes into your Harbor stack — a clean MicroVM per run, strong isolation, easy cleanup.

Follow the Harbor guide in our docs.

Community · From the community

Building a stateful research agent inside a Sandbox.

A developer built a stateful research agent on Tensorlake Sandboxes and wrote up the experience for Towards AI — a hands-on look at what the workflow actually feels like, with real numbers on how it performed.

Read it on Medium.

Try Sandboxes

We're offering free credits to developers who try Tensorlake Sandboxes and share their feedback. Whether you're building CLI agents on Harbor, computer-use agents, RL environments, or scaling parallel developer workflows, we want to hear what you build and what we can improve.