Skip to content
Ajutant Docs

Models and providers

Client Admin Manage models Since 0.1.0

You control which models the platform uses, what they cost, and how assistants choose them.

The model registry showing models with capabilities, pricing, and lifecycle.

The model registry

Section titled “The model registry”

The registry is the source of truth for available models, with their capabilities, pricing, and lifecycle. You can view model details, pricing history, and which assistants use each one.

Tiers and environment defaults

Section titled “Tiers and environment defaults”

Assistants usually pick a tier (a quality/cost level) rather than a named model. You map each tier to a real deployment for this environment, so a “value” assistant runs on the model you’ve chosen for that tier. You also set the deployments used by platform functions (such as chat-title generation).

Connecting providers

Section titled “Connecting providers”

Add provider credentials, Azure OpenAI, AWS Bedrock, Anthropic, Google Vertex: with the key or role held securely in your vault, then import their models with a guided pick-list. Vertex and Bedrock image models are supported too.

What each model accepts

Section titled “What each model accepts”

Some models reject request settings that others take for granted. OpenAI’s reasoning models and current Claude models, for example, reject temperature. Rather than guess, the platform checks each model when it’s added, by sending it a tiny test request for each setting, and records the answer. Models that were already installed are checked automatically within a few minutes of an upgrade.

From then on, a setting a model rejects is simply left out of its requests, for every assistant, task and developer-API call that uses it. Nothing else changes.

On the Models page:

  • Omits temperature (or another setting) means the model rejects it, so it’s being left out. This is information, not a fault.
  • Not yet checked means the check hasn’t run yet. If a request is sent before then and the model rejects a setting, the request is retried once without it, and the answer is recorded.
  • Parameter check failed means the check couldn’t run, for example because the provider was busy. It’s retried automatically. Hover to see why.

To check a model now, after fixing its credential for example, choose Check parameters from its menu. Test connection runs the same check once the connection succeeds.

The model is checked again automatically if its model version changes.