The Chat Interface

The chat panel on the right is where you direct the agent, in plain English. Type / for a dropdown of commands (arrow keys, Enter). The agent streams its response, shows a badge for each tool it runs, says what it did and why, and takes follow-ups: actually make it 2 dB higher instead.

What makes a good prompt

• Be specific about the track when you have multiple: normalize track 1 to -14 LUFS.
• Use time references: cut from 1:30 to 2:00 or remove the first 5 seconds.
• Chain operations in one message; the agent plans the whole chain.
• Reference markers you have set: cut everything before the chorus marker.

Big sessions

Each turn the assistant gets a short summary of the session: tracks, buses, master effects and markers, capped at 8,000 bytes for the hosted providers and 2,300 for Ollama. A list that is too long is cut, the summary says how many were left out, and the assistant reads the rest a page at a time with the list_session tool. It only reads, so Plan first never holds it for approval. With Ollama the tool is always sent when a list was cut.

Plan first

With the Plan toggle above the chat box on, no edit runs until you approve it. The toggle keeps its setting across restarts.

• The agent shows a plan card. Run approves it, Discard ends the turn without applying anything, and Edit lets you reword a step before you run it.
• If the agent answers without a plan, the card shows the first edit it is about to make, with its exact tool and settings (for example reverse, track: 0). Approve it and the rest of that turn runs without asking again. Discard it and nothing runs, its badge reads not run, and the agent is told you declined.
• If you reword a step instead, nothing runs: the agent proposes again and you approve that.
• Questions that only read the session (tempo, loudness, spectrum) run without asking. So do hearing an effect before you add it and rendering a preview: they change nothing in your project, and what they save is a disposable copy cleared as it grows.