Orizu CLI

What is Orizu

Orizu is a platform that helps you build continually learning agents and other LLM applications. It does so in a scientific, measurable manner by first helping you build evals and then helping you hill climb on them.

How Orizu works

There are 3 key pillars to Orizu:

1. Human evaluation: We consider human evaluation to be the gold standard, particularly when the humans involved are people with deep context and understanding of the output being evaluated. This is often the cross-functional team behind a product (product managers, engineers, designers, content writers, researchers, etc) or external experts (lawyers, engineers, writers, etc). Any good evaluation system relies on the presence of such high quality judgment to compare to. We provide tools that simplify the process of collecting the feedback from this audience.
2. Auto evaluation: Ultimately, human experts can only review so much. As your data scales, you need to encode their knowledge, their opinions, their taste and judgment into auto-evaluations. We automate this for you using the data we help you collect above.
3. Optimization: We turn your evals into an asset for you by building an automated hill-climbing workflow. We've started with prompt optimization because multiple studies and our own experience with our customers has shown strong measurable lifts with even relatively small datasets. Over time, we'll introduce additional capabilities, including optimizing your skills and other contextual information, and even model weights if your model provider allows finetuning.

CLI and Agent Skills

We've created a CLI for you to use so that you can perform any task that you can on our platform – from admin stuff like creating teams and projects to uploading datasets and assigning reviews – from your terminal. The obvious reason we did this is so that you can have your coding agents do this for you.

Note: the one task we do not let you perform from your CLI is any reviews assigned to you. We want to ensure that the data being collected is from the intended human expert.

Requirements

• Node.js 20+
• Access to an Orizu web app/API
• A valid Orizu account

By default, the CLI talks to https://orizu.ai.

Install

Install the CLI globally with your package manager:

| Package manager | Command | | --- | --- | | npm | npm i -g orizu | | pnpm | pnpm add -g orizu | | Bun | bun add -g orizu |

Then run the guided setup — it signs you in, initializes the local workspace contract, installs global coding-agent skill symlinks, and shows the next command for repo-specific Orizu adoption:

orizu setup

To repair or customize skill installs later, run the companion skill installer directly:

orizu install-skill --agent claude --agent codex --yes

Experimental plugin packages for Codex and Claude Code live under plugins/ in this repository. The CLI-installed skill remains the default onboarding path for now; the CLI is the runtime either way.

# Codex
codex plugin marketplace add <owner>/<this-repo>
codex plugin add orizu@orizu-plugins

# Claude Code (inside a session)
/plugin marketplace add <owner>/<this-repo>
/plugin install orizu@orizu

The npm package ships the skill alongside the CLI. Run orizu install-skill --agent claude --agent codex --yes for an explicit global install, or use orizu install-skill --help for project-scope installs, sync modes, and advanced targets including the managed AGENTS.md section for non-workspace repos. Keep installs in sync after CLI upgrades with orizu skills status and orizu skills update.

Coding agents can also read the bundled skill directly without installing it:

orizu skills path --json

This prints the skill location plus source metadata (name, root, skillMd, source, cliVersion, skillHash), so an agent can go from npx orizu --help to reading SKILL.md in two commands and verify the guidance matches the CLI that supplied it.

Agents are first-class users of this CLI: every command supports --json (as a prefix, orizu --json teams list, or a trailing flag) and emits a single machine-readable JSON document instead of formatted text. Discover the full command surface, including global options, with orizu capabilities --json.

Authentication

Sign in

orizu login

This will open a browser tab for you to login with. You must either have an account on the Orizu platform or it will help you create one. Approving the login creates a personal access token for the CLI and stores it in your local Orizu credentials file. You can revoke CLI tokens from the Personal Tokens page in Orizu.

Check the signed-in user

orizu whoami

Sign out

orizu logout

Common Commands

Supported Data Formats

Dataset upload, append, and edit-rows commands support:

• .csv
• .json files containing an array of objects
• .jsonl files containing one object per line

Task and dataset export support:

• csv
• json
• jsonl

Interactive And Automated Usage

Many commands can prompt for missing team, project, app, dataset, or task selections in an interactive terminal. In scripts and CI, pass explicit flags instead:

orizu tasks export --task <taskId> --format jsonl --out ./labels.jsonl

Use --json on supported task and app commands when automation needs structured output.

More Documentation

• docs/cli.md: complete CLI guide and command reference
• skills/orizu-cli/references/cli-reference.md: compact command matrix and end-to-end flows