Tensorlake Sandbox

license

Apache-2.0(OSI)

source

public(runtime+SDK, BYOC self-host in your own AWS/GCP/Azure)

managed-tier

Tensorlake hosted cloud + proprietary Lattice scheduler

core-gated

gated(the MicroVM runtime and the Lattice cluster scheduler that creates sandboxes exist in no public repo - the Apache-2.0 tree publishes the SDK, the tl CLI and client crates, which authenticate to cloud.tensorlake.ai with an API key)

Apache-2.0 SDK and CLI with a proprietary managed runtime. The core is gated on a named mechanism: not a carve-out inside the repository, but a component that is in no repository at all. The 2,350-entry tree is entirely client side - the Python SDK around cloud_client.py, the tl CLI, and nine crates that are all clients, codecs or bindings - and the README attributes sub-second creation to "Lattice, a dynamic cluster scheduler", along with snapshots, cloning, auto suspend/resume and live migration, none of which appear in any of the org’s 58 public repos. Onboarding is an API key from cloud.tensorlake.ai. One tension is worth stating plainly: this record describes the source as runtime plus SDK with BYOC self-hosting in your own AWS, GCP or Azure account, and that is not supported - the runtime is not published, and BYOC means the vendor’s closed control plane runs in your account rather than open self-hosting, which would argue for a more restrictive reading than open core. The tension is stated here rather than resolved.

• https://github.com/tensorlakeai/tensorlake recorded 2026-07-09

  Apache-2.0 license; sandbox runtime + SDK (Python + Rust); BYOC self-host

• https://www.tensorlake.ai/ recorded 2026-07-09

  sandbox-native cloud for AI agents; run sandboxes in your own cloud account

• https://api.github.com/repos/tensorlakeai/tensorlake/git/trees/main?recursive=1 recorded 2026-08-12

  Full untruncated recursive tree, 2,350 entries, one governing LICENSE plus five vendored third-party ones. Everything in it is client side: src/tensorlake/ is the SDK around cloud_client.py, and all nine crates/ are cli, cloud-sdk, function-agent-core, gsvc-codec, gsvc-fs-client, gsvc-mount and the node/py bindings. platform/ holds only macos/tlfs, a FUSE mount helper. No scheduler, allocator, VMM or control plane, and no ee/ or enterprise/ path.

• https://raw.githubusercontent.com/tensorlakeai/tensorlake/main/README.md recorded 2026-08-12

  The capabilities the product is sold on are server side and named as such: "Sandboxes created in under a second via Lattice, a dynamic cluster scheduler", plus snapshots and cloning, auto suspend/resume, live migration between machines, and "up to 5 million sandboxes in a single project". Getting started is "Sign up at cloud.tensorlake.ai and get your API key", export TENSORLAKE_API_KEY, tl login. The repo documents no way to run any of it yourself.

• https://api.github.com/orgs/tensorlakeai/repos?per_page=100 recorded 2026-08-12

  All 58 public repos in the org. Beyond the SDK they are skills, cookbooks, examples, benchmarks, integration plugins and forks of third-party projects (cloud-hypervisor, vllm, arrow-rs, moka, fuser). The former open compute engine survives only as archived clients - indexify-typescript-client and indexify-extractor-template, both archived. Nothing implements Lattice or the sandbox runtime.

• https://api.github.com/repos/tensorlakeai/tensorlake recorded 2026-08-12

  Repo metadata: language Python, license spdx_id Apache-2.0, 986 stars, default branch main.

986 GitHub stars on tensorlakeai/tensorlake, which falls under 1K and reads at level 1 on a star-based band - capped in any case, because a star is not a use. No download, install or customer figure is published for this product, so stars remain the only honest signal and the level is directional.

• https://api.github.com/repos/tensorlakeai/tensorlake recorded 2026-08-12

  stargazers_count = 986 for tensorlakeai/tensorlake

Firecracker microVM isolation plus durable-execution orchestration and snapshot/resume make for a rich agent-runtime feature set, level with the microVM peers: tensorlake.ai leads on Firecracker isolation with each tool call and harness in its own microVM, snapshots with pause/fork/resume, versioned mountable volumes, durable serverless functions with queues, timers and retries, and cloud dev boxes. The measured half of that picture is in tension with the score. Independent ComputeSDK TTI benchmarks put Tensorlake at 1.35s median time-to-interactive, 1.44s P95, composite 86.1, 15th of 24 providers - roughly three times the ~460ms and 12th-of-19 reading this band was written against, and well short of the vendor’s own sub-second creation claim. Note that 460ms is also the number the vendor site animates in its demo, so the vendor claim and the leaderboard are not two independent measurements. Nothing here asserts sub-second latency as a measured fact: the vendor claim and the measured figure are recorded separately, which is what the two sources support. The 4 stands on isolation and feature depth, and whether a 1.35s time-to-interactive should pull it down is an open question rather than one settled here.

• https://www.tensorlake.ai/ recorded 2026-08-14

  Firecracker microVM sandboxes, snapshot/suspend/resume, durable execution, scale to thousands.

• https://www.computesdk.com/benchmarks/sandboxes/ recorded 2026-08-14

  ComputeSDK TTI leaderboard, independent, 100 iterations per run, last run 7 August 2026, 24 providers on 4 vCPU / 16GB in Northern Virginia. Tensorlake records 1.35s median, 1.44s P95, 1.44s P99, 100% success, composite 86.1 - 15th of 24 by composite and 17th of 24 by median.