Run Crabbox Test Suites in a Tensorlake Sandbox

Crabbox is an open source CLI from OpenClaw with a simple loop: warm a box, sync the diff, run the suite. Since May 2026 it has shipped a Tensorlake provider in its Firecracker family, so one command puts your test run inside an isolated microVM:

crabbox run --provider tensorlake --tensorlake-image tl-crabbox -- pnpm test

Who does what

The division of labor is clean. Crabbox owns the local workflow: config, repo claims, sync manifests, and guardrails. Tensorlake owns the microVM and command transport. Under the hood, Crabbox shells out to the tensorlake CLI, calling sbx create, cp, exec, and terminate to drive the sandbox lifecycle.

tl-crabbox is a public image we publish specifically for Crabbox: the standard Ubuntu base plus a writable /workspace (Crabbox's default workdir) and pnpm preinstalled.

Setup

Install both CLIs:

brew install openclaw/tap/crabbox
curl -fsSL https://tensorlake.ai/install | sh

Export your API key, then add a .crabbox.yaml at your repo root so every run picks up the right image without flags:

provider: tensorlake
tensorlake:
  image: tl-crabbox

The image line matters. Crabbox defaults to /workspace/crabbox as its workdir, and in Tensorlake's standard images commands run as tl-user, which cannot create /workspace. Pinning tl-crabbox avoids that failure. If you prefer a standard image, set tensorlake.workdir: /home/tl-user/crabbox instead.

One requirement to know up front: run Crabbox from inside a git repository. It builds its sync file list from git ls-files, so a plain directory fails.

Why run tests in a microVM

A test suite is arbitrary code execution by design. Running it in a sandbox means a flaky postinstall script, a test that shells out, or a dependency you have not audited executes in a disposable VM with its own kernel, not on your machine. And because the environment comes from a pinned image, the suite runs against the same toolchain every time, on every machine.

The full provider reference, including CPU, memory, disk, and namespace configuration, is in the Crabbox integration docs.