Skip to main content
This guide gets you from zero to a submitted benchmark run. For command and flag details, see the CLI overview.

Prerequisites

  • A Harmstack API key (HARMSTACK_API_KEY) from Vetted Medical
  • macOS (Apple Silicon), Linux (amd64), or Windows (amd64)

Install the CLI

Binaries and checksums live on GitHub: vettedmedical/harmstack-install-v0 — latest release

macOS (Apple Silicon) and Linux (amd64) with Homebrew

Manual install from a release archive

Download the asset that matches your OS (version numbers differ per release):
  • macOS Apple Silicon: harmstack_<version>_darwin_arm64.tar.gz
  • Linux amd64: harmstack_<version>_linux_amd64.tar.gz
  • Windows amd64: harmstack_<version>_windows_amd64.zip (contains harmstack.exe)
macOS or Linux — extract and put harmstack on your PATH:
Windows — unzip the file, move harmstack.exe to a folder on your PATH, open a new terminal, then run harmstack --help. Optional: verify downloads using checksums.txt from the same release. More detail, upgrade paths, and platform notes: CLI overview — Installation.

Set your Harmstack API key

For a persistent shell setup (example: zsh):

Configure your model endpoint (non-interactive)

Export the URL and API key for the model you are benchmarking:
Harmstack supports multiple provider shapes, including openai, openai_responses, gemini, and raw.

Run your first benchmark

This example skips the interactive wizard and submits a Haystack run from the root command:
Adjust --benchmark-id, --unit-count, and --provider to match your setup. See Flags for all options.

Next steps

CLI overview

Installation depth, authentication, and links to platform guides.

CLI commands

Command inventory and where to find generated metadata.