Choose the agent you already use and a model suited to the review. Your repository’s review context stays available when you switch.
Switch the active agent
Settings → Models has four tabs: Claude Code, Cursor, Codex and Copilot. By default, CodeCrab selects the first ready agent in that order. The agent in use is tagged Active.
- Open the tab of the agent you want.
- Click Use Claude Code, Use Cursor, Use Codex or Use Copilot. Viewing a tab alone changes nothing.
- Select the model for that agent.
You can switch at any time. Review Profiles, skills, reviews and deep dives carry over.
One model per agent
Each agent tab has one model selection for reviews, private skills, combining multiple review passes, Investigate, deep dives and Pre-push. Local default uses whatever that agent’s setup or account already uses.
| Claude Code option | Review trade-off |
|---|---|
| Local default | Use the CLI’s configured model. |
| Opus | Deepest analysis; slower and more quota. |
| Sonnet | Balanced default. |
| Haiku | Faster, with shallower findings. |
The existing Preferences guide covers language and appearance.
Cursor model list
Choose Auto, Local default or a model available in your Cursor account. The list has one current option per model family; speed and effort variants are folded into that family’s default variant.
Refresh after Cursor adds models. A previously selected model remains shown and selected if it is no longer in the current list. Exact model names depend on availability; examples are not guaranteed options.
Codex model list
The list comes from Codex for your account. Choose Local default or a listed model. If the list is empty, open Codex once in its CLI or app, then refresh. The Windows tab is marked Beta.
Copilot model list
Pick Auto (Copilot picks the model for each run) or a model your plan allows. Click refresh to load your account's models. Copilot Free offers Auto only. A previously selected model your plan no longer allows is marked Unavailable. Profiles and skills use the strongest model your plan allows at high reasoning effort (Auto on Copilot Free).
Strong models for profiles and skills
Review Profiles and generated skills are reused across reviews, so their generation uses a strong model regardless of the review model selection.
| Agent | Profile and skill generation |
|---|---|
| Claude Code | Opus; local default if your plan does not include Opus. |
| Cursor | The strongest model in your available list. |
| Codex | The top model at high reasoning effort. |
Cursor, Codex and Copilot mark this model with Profile & skills. Generation happens occasionally, not on every review. Hover a Reviewed chip on a PR card to see which agent and model produced that review.
Control review effort
Select only the skills you need. For Claude Code, Sonnet or Haiku can reduce review effort; for Cursor or Codex, choose an appropriate lighter model from your available list. Profile and skill generation still uses a strong model.