Install CrashLab

CrashLab 0.1 supports macOS on Apple Silicon and Intel. The supported sandbox is local Docker.

Requirements

• Homebrew
• Git
• Docker Engine or Docker Desktop with Buildx and a running daemon
• the credentials required by the agent being evaluated

You do not need to install Python, uv, Node.js, or Bun separately. The formula owns CrashLab's Python runtime, installs uv for locked evaluation packages, retains the exact CrashLab wheel used by their fresh runtimes, and installs the Ink workspace with its formula-managed Node runtime.

Install with Homebrew

brew install crashlabsai/tap/crashlab

Verify the complete local toolchain:

crashlab --version
crashlab --help
docker version
docker buildx version

Run bare crashlab from an agent repository to open the interactive workspace. Explicit commands such as crashlab run --json remain available for coding agents and CI.

Install the coding-agent setup workflow

CrashLab bundles a managed setup-crashlab skill for Codex, Claude Code, Cursor, and OpenCode:

crashlab skills install

To install for every supported agent location:

crashlab skills install --agent all

Managed copies update idempotently. CrashLab refuses to overwrite an unmanaged or user-modified skill. Restart the coding-agent client if it does not discover new skills immediately.

Update or uninstall

brew update
brew upgrade crashlab

brew uninstall crashlab

Upgrading the CLI does not delete project-owned evaluation source or retained .crashlabs/runs/ history. During the 0.1 alpha, read the release notes before upgrading because authoring contracts may change.

Credentials stay on the host

CrashLab runs the target in Docker while the host keeps the world, verifiers, expected answers, and provider credentials outside the target container. CrashLab does not automatically load a repository .env file. Read Local security and secrets before connecting a provider credential.

Build from source

Source installation is a contributor workflow, not the supported customer distribution. Contributors need Python 3.11+, uv, and Bun 1.1.31+ to build and test both executables; follow the repository's contributing guide.

Next, set up your agent repository.