Browse documentation

Models

Choose the model an agent runs on: NGT Auto, or models from your own provider keys.

Every agent runs on a language model. By default that model is NGT Auto, which NGT provides and pays for. A workspace can also offer models from Anthropic, Google and OpenAI on its own API keys, and choose exactly which of those models its agents may use.

NGT Auto

NGT Auto picks the best model for each message: a fast one for quick questions, and a stronger one that thinks longer for analysis and multi-step work. Before every turn, a small model classifies how demanding the message is, and NGT Auto runs the turn on the model for that class.

What NGT Auto gives you:

  • No key to manage. It runs on NGT's account, never on a workspace's key, so it works with or without a vault.
  • EU hosting. It runs Google Gemini models on Google Cloud in the EU (the Models page names it: Gemini Enterprise Agent Platform's EU endpoint), under NGT's Google Cloud agreement, so your data stays in the EU.
  • OCR included. Reading text and tables from scanned documents and images, when a file is extracted, is part of NGT Auto. It runs on Mistral OCR at Mistral AI's EU endpoint, the one model call of an extraction that runs outside Google Cloud.

New agents, templates and the built-in Spark agent all start on NGT Auto. In usage reporting, its calls are grouped under NGT Auto rather than the model each turn ran on.

Open the Models page

Models is part of Agent Studio. Open the Settings tab and choose Models in the Agent Studio group, or click the Models card on the Ecosystem page. You can also open it from an agent's Model section with Manage models.

The page has three parts:

  • Enterprise guarantees: what NGT's Google Cloud agreement covers for providers marked as running on Google Cloud in the EU: Google Cloud, EU data residency, and Enterprise security, no training on your data, with links to Google's pages (Data residency, Security controls, Zero data retention).
  • Active: the providers your agents can use now. NGT comes first; open it to see Auto and OCR.
  • Bring your own: the providers you can switch on with your own API key.

A provider's row shows its logo, how many models it offers, and, for a provider that runs on your own key, a key icon: dark with the tooltip Key added, or muted with Needs your key. Hosting marks on a row, such as NGT's Google Cloud and EU marks, say where its calls run; hover them for details.

ModelsThe model providers and models your agents can use.Enterprise guaranteesGoogle Cloud, EU data residencyMarked providers run on Gemini Enterprise Agent Platform's EU endpoint, so your data stays in the EU.Data residencyEnterprise security, no training on your dataGoogle Cloud's security controls apply, and Google doesn't train or fine-tune its models on your data.Security controlsZero data retentionActive (2)NGT1AAnthropic8Bring your own (2)GGoogle4OOpenAI4
The Models page: the guarantees, the providers in use, and the ones you can bring your own key for.

Use your own provider keys

To run agents on your own Anthropic, Google or OpenAI account:

  1. On the Models page, find the provider under Bring your own, open its ⋯ menu and choose Activate (or open the provider and click Activate).
  2. In Add your Provider API key, choose the Vault to store the key in and paste the API key. Click Skip for now if you will add the key later.
  3. On each agent that should use these models, pick that vault in its Vault section. The provider's models then appear in the Model picker.

You can also add a key from a vault itself: open the vault and use Add provider key in its LLM provider keys section. See Vaults.

An agent bound to a vault runs on that vault's keys instead of the platform's. A model you pick by hand runs on your key and goes straight to that provider. Even a Gemini model picked by hand runs on Google's Gemini API, which makes no EU promise, unlike NGT Auto.

The Model picker only lists a provider's models once the provider is active and the agent's vault holds its key. Without a vault, the picker holds only the models NGT provides. A note at the foot of the picker says how to add more.

Choose which models your team can use

Open a provider (click its row, or Manage models in its ⋯ menu) to see its models. Each model shows what it is for and its class, with a switch to offer it to your agents. Every model is on by default, and each switch saves at once.

To use a model the list does not have yet, type its ID under Add a model. It is offered to the team straight away and marked custom. Hover a custom model to Remove it; agents already on it keep running on it until they are given another model, but no one can pick it again.

Switching a provider off is Deactivate in its ⋯ menu (a provider under Bring your own offers Manage models and Activate). The NGT provider stays active, but its NGT models dialog gives Auto its own switch (Offer Auto), so a workspace can stop offering it to agent builders. OCR reads text and tables from scanned documents and images when a file is extracted, on Mistral OCR at Mistral AI's EU endpoint, outside Google Cloud; its switch is on and cannot be changed yet.

Pick a model for an agent

Open the agent and use its Model section. The picker is searchable and grouped by provider. If a model is no longer available, the section says why and how to fix it: pick another model, turn it back on in Models, or add the provider's key to the agent's vault.

Under Advanced, Thinking sets Low, Medium or High reasoning effort when the chosen model supports it. NGT Auto sets thinking itself for each turn.

Where to go next