fullauto.online

Guide

Running Claude Code, Codex and Hermes through Ori: a working setup

Install Ori, sign in once, and launch your usual agents on any OpenRouter model. The exact commands, the model flag, the controls that carry over from your organisation, and where it still trips people up.

Published
6 Sep 2026
Reading
10 min
Class
guide

half-life 45dfrom 6 Sep 2026

This is the practical companion to our note on what Ori Harness is. It assumes you already have an OpenRouter account and want to run one or more agent CLIs against it without hand-configuring each.

1. Install

On macOS, Linux or WSL:

curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash

On Windows, in PowerShell:

irm https://openrouter.ai/labs/ori/install.ps1 | iex

Read an installer before piping it to a shell. That is standard advice, and it applies even to a tool from a company you already pay.

2. Sign in

ori login

This opens a browser OAuth flow. You can also skip it and just launch an agent, which triggers the same sign-in. Your credits, models and organisation settings follow your OpenRouter login, so there is no key to create, store or rotate.

3. Launch an agent

ori claude
ori codex
ori opencode
ori hermes

Each command starts the matching harness with its OpenRouter configuration applied. If the agent's own CLI is missing, Ori installs it first.

4. Choose a model

Add --model with any OpenRouter model ID: 

ori claude --model anthropic/claude-opus-5
ori codex  --model openai/gpt-5.6-sol

Ori detects the model and adjusts the harness settings to suit it. Some harnesses go further: Pi, Prime Agent and DeepSeek Harness read your live OpenRouter catalogue, so /model lists what you can use, /fast asks for fast routing and /zdr restricts requests to zero-data-retention providers (Prime Agent names its speed toggle /speed). For other harnesses the model is validated at launch only.

5. Know what carries over

Organisation guardrails, model allowlists, workspace budgets and routing rules apply to Ori traffic exactly as they do to any other OpenRouter request, across every agent. Usage is billed as normal and shows up in usage accounting and the activity export. If you run a team, that is the real reason to use it: one policy, several agents.

6. Keep it current

ori update

Two habits

Pin a model per task rather than leaving a default: a cheap model for mechanical edits, a stronger one for hard reasoning. And check the price of the model on OpenRouter before a long autonomous run; agent loops multiply token use faster than chat does.

Where it still bites

  • It is a configurator, so agent bugs are still agent bugs. When something fails, check whether it fails outside Ori too.
  • Model quality varies by harness. A model that shines in one agent can stumble in another; expect to try two or three.
  • Provider-side features such as prompt caching depend on the route, so cost can differ from a first-party plan. Compare on a real task.

For which harness to start with, see the harness board; for choosing a model, the model leaderboard.

Sources

(p) => ` `