Models and routing

cmdk is not tied to one model. The catalogue comes from the AI gateway and includes the current models from OpenAI, Anthropic, Google and others, alongside self-hosted options.

Auto mode

By default the composer is in Auto. Each message is routed to the model that fits it: a quick question goes to a fast model, a long document or a multi-step task goes to a stronger one. Trivial messages such as a greeting skip the router entirely, so the answer starts at once.

Auto has three pools:

Media has its own picks. Under Image, Video, Music and Speech you choose which model handles each kind of generation.

Choosing a model yourself

Open the model selector in the composer and search. The chat keeps that model until you change it or start a new chat.

Three other places pin a model:

• Skills. A skill can carry a model in Edit details. When the skill is loaded in a chat, the rest of that run uses the skill's model; if several skills are loaded, the most recent one wins.
• Automations. The brief on an automation's Settings tab has a model picker for the whole plan.
• Agents. Each agent has a model in its settings, used for the steps it takes.

Free models and depleted credits

Two small models cost no credits. When a user's monthly credits run out, chat and automation runs fall back to the free model instead of stopping, so nothing is lost. See Credits and usage.

What the model sees

Every message is sent with the workspace Instructions and Glossary, the memory the agents keep, a project's Escalations when you are in one, the skills and tools in scope, and the relevant pieces of knowledge retrieved for that message. The stable part of this context is cached, which keeps repeated turns fast and cheap. The date is included; the time of day is not.