Skip to main content
Choose Claude Code, Codex, or OpenCode from the model menu before starting a task. The agent, or harness, provides the tools and session behavior; the model does the reasoning. Follow-ups keep the same harness. You can change its model and reasoning effort; start a new chat to use a different harness. A first cloud handoff can choose a different cloud agent.

​
Local models

The model menu uses the connected computer’s catalog:
  • Claude Code: default models and model IDs in your global Claude settings.
  • Codex: models reported by your connected account.
  • OpenCode: models from connected or API-configured providers.
If discovery fails, the menu marks its fallback list. Catalog discovery uses global configuration, so project-specific model settings may not appear. Provider account permissions and usage limits still apply.

​
Cloud models

The cloud picker lists the models configured in Twill, including provider models through OpenCode and its OpenRouter… option. Use the picker as the current list of available models. Set your cloud workspace default in Settings → Routing. Integration triggers can choose a different configuration for a new task. See Routing. Cloud runs use workspace credits unless a matching provider key is configured. Local runs use the computer’s native account instead.

​
Reasoning effort

Effort levels vary by agent, model, and local runtime. Twill adjusts an unsupported choice when you switch models.

​
Repository instructions

Harnesses read your repository’s AGENTS.md or CLAUDE.md locally and in the cloud. Cloud agents can also record learnings there as reviewable code changes. See also repository skills and commands.