Models

Model picks are personal. Settings > Models writes models.* in the user .vixl settings.json. A project folder cannot override them. Runtime still reads the provider key from the OS keychain.

Until a provider exists, this section is blocked. The chat input picker is disabled in the same case and offers Add a provider. Send requires a model.

Saved values are providerId::modelId. Pick a model to persist it.

Set the default

Default (models.default) is the fallback every other role uses when that role has no override. Set this first. Without a resolved Default (or Agent, which falls back to Default), starting a chat from a plan is blocked until you pick one here.

Role models

Each role has its own picker. Roles other than Default can fall back to Default. Use default clears that role and its reasoning override.

Ask, Plan, Agent, and Orchestrator are the chat mode defaults. Agent is also used for single-agent plan builds. Orchestrator is the parent for Orchestrator mode and for Orchestrate from a plan, with nested Parent and Subagent pickers.

Subagent (models.subagent) is the default for nested spawn_subagent runs when an agent file does not set its own model. If Subagent is unset, resolution uses Agent, then Default.

Title (models.title) generates short titles for new chats. Auto-title (chat.autoTitle, on by default) lives on that row. Turning it off disables the Title picker. If Title is still using Default, Settings warns you to prefer a small, low-cost model for that background task.

Resolution (when the chat has no picker override): the role's own setting if set; for Subagent, then Agent, then Default; for chat modes and Title, Default. A per-chat picker override wins when present.

How modes consume these defaults is on Models and roles.

Catalog providers list live /models from that endpoint (Ollama and LM Studio use their default base URLs). You do not maintain a static list for those. Custom endpoint model rows (import, id, capabilities, pricing) are edited on the provider, not here.

Per-model options

The extras panel on a model picker writes models.catalogOptions for that providerId::modelId:

• Allowed in chat hides the model from chat pickers when off.
• Fast when the model has a fast sibling.
• Reasoning when the model supports it: Default, None, Minimal, Low, Medium, High, Extra high, Max.
• Context window and Max output when the catalog reports those limits.

Role reasoning overrides are separate (models.<role>Reasoning) and are covered with the rest of role behavior on Models and roles.

First-run setup is Set up providers and models. To attach tools from outside the model, add MCP servers.