AI assistant setup

Current status

The native desktop app includes assistant UI and provider settings, but the model bridge is still being ported to the current native-media line. In builds with the bridge disabled, the app will tell you the assistant is not wired up yet.

The intended assistant model is deliberately conservative: it produces normal Cutlass edit commands, the same kind the UI runs, so changes remain inspectable and undoable.

Provider settings

Cutlass does not ship a model. Point it at an OpenAI-compatible endpoint, either local or hosted.

Common local setup:

[ai]
base_url = "http://localhost:11434/v1"
model = "qwen3:14b"

For cloud providers, set either a literal key or an environment variable reference:

[ai]
base_url = "https://api.openai.com/v1"
model = "gpt-4.1-mini"
api_key_env = "OPENAI_API_KEY"

Config file

The settings file lives at:

~/.cutlass/config.toml

The desktop Settings dialog writes the same file and preserves comments and hand edits where possible.

Reviewable edits

Assistant edits should behave like user edits:

• The assistant plans an edit.
• The edit lowers into structured timeline commands.
• The engine validates the command against the current project.
• The resulting change appears on the timeline.
• Undo and redo continue to work through normal history.

Dry-run and review flows matter because smaller local models can be less reliable at tool calls.