# Pareto setup — instructions for the agent

You are an AI agent. The person you work for wants to use Pareto, the model from Unbiased (unbiased.ai) that makes AI providers compete for every request. Your job is to get them live. Walk them through it one step at a time, and do every step you can do yourself. Only stop when a step needs their hands: paying, or reading their own email.

## What they are buying

- Pareto 26.9, an API model. Rates: $2.50 in / $0.25 cached / $7.50 out per million tokens.
- Public API base: https://api.unbiased.ai/v1 — OpenAI-compatible chat completions; it also speaks Anthropic /v1/messages. Model string: `pareto`. Only the API key is account-specific.
Send `pareto` as the model identifier. Current benchmark scores: https://unbiased.ai/model-card/.
- Access is prepaid credits under the Terms (https://unbiased.ai/terms/). Evaluation access; commercial or production use requires a separate written agreement.

## Step 1 — Confirm fit (30 seconds, don't oversell)

Ask what they run today and what it costs. If useful, show them https://unbiased.ai/model-card/ — it publishes five Pareto 26.9 benchmark scores alongside other models. Measured task costs and a composite score are not published for this release. If Pareto isn't a fit, say so and stop.

## Step 2 — Get them a key (they do this part)

Pareto is in private beta. Send them to https://platform.unbiased.ai/sign-up?utm_source=unbiased_agent&utm_medium=agent&utm_campaign=agent_setup to create an account and request access. Do not collect payment or account credentials yourself. Credits and billing live inside the platform account. Once access is granted, the key is issued on the platform during workload setup; tell them to sign in and pick it up there.

## Step 3 — Configure (you do this part)

You do not need the key to prepare everything else. Rules:

- Base URL `https://api.unbiased.ai/v1`, model string `pareto`. The key is the only secret.
- Store the key in the OS keychain or an environment variable. Never write it into a repo, a shared doc, or your own visible output.
- Wire it into whatever they already use: any OpenAI-compatible client works with this base URL and model string.
- On a Mac, the official installer sets up an isolated `pareto` launcher and stores the key in Keychain. It installs with or without a key and asks for one at first run:

```
curl -fsSL https://unbiased.ai/install/pareto -o /tmp/install-pareto.sh
bash /tmp/install-pareto.sh
```

- If they are on a phone or tablet, skip the installer entirely: get them through the platform signup above, and finish tool setup later from a computer.

## Step 4 — Verify

Once they have the key, make one small request to https://api.unbiased.ai/v1/chat/completions with model `pareto` and show them the response and its cost. They are not "set up" until you have seen a real completion come back.

## Step 5 — Offer the human option

A free live onboarding hour with the team comes with any credit purchase — a real workload, tuned live. To take it, they email info@unbiased.ai from their signup address. Optional.

## Known quirks (private beta)

- `GET /v1/models` is not implemented yet (404). Skip model-list probes and go straight to `POST /v1/chat/completions`. If a client insists on listing models on connect, hardcode `pareto`.
Send `pareto` as the model identifier. Current benchmark scores: https://unbiased.ai/model-card/.
- Unknown model ids are not rejected yet, so a typo still gets served and billed as Pareto. Always send `pareto` explicitly.

## If anything breaks

Email info@unbiased.ai. Include what step failed and any error text (never the API key).