# Set up the agent

> Choose Orcabase AI or your own OpenRouter or Vercel AI Gateway key, pick the models your team can use, and set a step limit per question.

Source: https://tini.so/docs/agent/setup

The agent needs an AI model to think with. An admin chooses how it gets one, once, for the whole organization, under **Settings → Agent** “Admins”. Everyone in the organization can then use it.

## Two ways to run it

|   | Orcabase AI | Your own key |
| --- | --- | --- |
| What it is | Runs on Orcabase’s AI account | Runs on your OpenRouter or Vercel AI Gateway account |
| Setup | Pick models. There’s no key to paste | Paste an API key, then pick models |
| Models | The list Orcabase offers | Hundreds: anything the gateway offers |
| Who pays for AI | Covered by a credit for your organization | You, directly to the gateway, at its prices |
| Usage shown as | A percentage of your credit | Tokens and dollars |

Orcabase AI appears first when it’s available for your organization. You can switch between the two at any time; a key you saved stays saved, so switching back doesn’t mean pasting it again.

OpenRouter and Vercel AI Gateway are **gateways**: one account and one API key reach models from Anthropic, OpenAI, Google and others, on one bill.

## Set up your own key

1. **Get a key**

   Create an API key in your OpenRouter or Vercel AI Gateway account. Give it a spending limit there: it’s a second safety net on top of the limits in Orcabase.

2. **Paste it**

   In **Settings → Agent**, pick **OpenRouter** or **Vercel AI Gateway** as the **Provider** and paste the key into **API key**.

3. **Pick models**

   Choose a **Default model**, and up to eight **More models to try** that people can switch to. The picker is searchable, shows each model’s price per million tokens and how much text it can take in, and lists popular models first. It hides models that can’t use tools, since the agent needs them; you can also type any model id.

4. **Test and save**

   Click **Test and save**. Orcabase makes one tiny request with that key and each model before saving, so a wrong key or an unavailable model is caught now. *Saved. The key works.* means you’re done.

> **Note: How the key is kept**
>
> The key is encrypted, used only to call the gateway, and never shown again: the page shows its last four characters. **Remove key** deletes it, and the agent stops until a new one is added.

## Set up Orcabase AI

Pick **Orcabase AI** as the provider, choose a default model and any more models to try from Orcabase’s list, and save. There’s no key and nothing to test. Usage draws on a credit shared by your whole organization; see [Orcabase AI credit](https://tini.so/docs/agent/privacy-and-usage#orcabase-ai-credit).

## Steps per question

Each call to the model is one step. **Steps per question** caps how many one question can take: between 5 and 100, 30 by default. At the limit the agent pauses and asks whether to keep going, so the limit controls cost without cutting answers off. Changing only this number saves without testing the models again.

## Usage

The **Usage** card on the same page shows this month’s model calls, tokens and cost (or share of credit), broken down by model and by member. On Orcabase AI it also shows how much of today’s, this week’s and this month’s credit is used. See [AI privacy and usage](https://tini.so/docs/agent/privacy-and-usage).

---

Previous: [Data Agent](https://tini.so/docs/agent.md) · Next: [AI privacy and usage](https://tini.so/docs/agent/privacy-and-usage.md)

All docs: https://tini.so/llms.txt