AI requests and models

Ask, the setup agents, and every other AI step in Crux call models server-side through OpenRouter — requests never run on a browser-side provider key, and nothing about your project is exposed to the model beyond the context needed for the turn.

Choose a model for the Agent

The model control sits in the Agent composer, to the left of the attach and web-search controls. It shows the model in use; click it to switch.

Fast is the default and the right answer most of the time: it streams quickly enough that you can keep editing while it thinks. Reach for Sharp when a Fast answer feels shallow rather than wrong, and for Capable when the question is genuinely hard and you would rather wait for a better answer.

A few things worth knowing:

• The choice is remembered per project, so a project where you mostly do heavy synthesis can sit on Capable while another stays on Fast.
• You can switch mid-conversation. The next turn uses the new model; the thread carries on.
• Every option reads the same context. Your graphs, cited items, connected sources, and web results work identically whichever model is running — only the reasoning changes.
• Credits follow the work. Fast makes your allowance go furthest; Sharp draws slightly more per turn and Capable noticeably more, because you are asking for more thinking. See AI credits.
• On a Free plan, Capable is shown but locked, marked Pro. If a plan changes and the selected model is no longer available, Crux quietly falls back to Fast rather than failing the turn.

Everything else is managed for you

Outside the Agent composer, Crux picks the model so you don't have to. Generate on a detail field, the research and experience agents during project setup, and the skills behind slash actions all run on the managed default, and supporting jobs use models chosen for the specific task: text embeddings for search and citation matching, a vision model for extracting text from images and scanned PDFs, and image models for persona portraits and diagrams.

That split is the point. You get a dial where a harder question is worth more thinking, and no configuration anywhere else — no keys, no per-feature settings, and no need to re-test your workflows when a better model ships, because Crux moves the managed default forward after checking the prompts behind each feature.

Web search turns

Turning Web search on for a turn attaches live web results to the request, on whichever model you have selected. The reply cites web sources separately from your project's own evidence, so you can always tell which claims came from outside your graphs. Web-search turns are metered as their own operation type.

Credits

Every AI operation — an Ask turn, a Generate on a field, an agent run during project setup — debits AI credits from your plan allowance, and only after a successful response. The balance is visible in the sidebar. See AI credits for how allowances, pooling, and top-ups work.

Quota or rate limit

If you see a message about an AI quota or rate limit, the provider throttled or rejected the call rather than Crux failing:

1. Wait a short time — rate limits clear on their own.
2. Try the request again, or switch the Agent to Fast for that turn.
3. If it keeps happening, open Help in the sidebar and use Chat with us.

Other AI failures

For temporary provider errors or unexpected failures:

1. Retry once — most transient errors clear.
2. Check that the item you are asking about has finished processing; a file still being parsed has no text to answer from yet.
3. Open Help in the sidebar and use Chat with us so we can diagnose it with you.

Help also includes the docs and an agent grounded in this documentation.

Next

For how Ask uses these requests day to day, see Using Ask.