The AI Assistant
How to open the AI assistant, ask it to edit your project, and stay in control of what it changes.
The assistant lives in the panel on the right side of the window, on the Agent tab. As of 0.4.0 it's a chat. You type a request in plain English, the assistant answers in the conversation, and every action it takes on your project shows up as a readable row you can inspect. It's wired into the same editing engine as the rest of the app, so it genuinely changes your timeline, not just talks about it.
Open the panel
Choose View → AI Assistant Panel. The panel appears on the right side of the window, next to the Properties panel. The same menu item hides it again, so if you don't see the panel, this is where to bring it back.
To make the panel wider or narrower, drag the divider on its left edge. The app remembers the panel's size and whether it was open.
The panel has two tabs: Agent, the chat, and Terminal, covered near the end of this page.
Send a request
Type into the message box at the bottom of the panel and press Enter. Shift+Enter starts a new line instead of sending.
You can attach files to a message in three ways: click the paperclip (Attach files), drag files from Explorer onto the box, or paste a screenshot straight from the clipboard. Images travel with the message so the assistant can look at them. A message takes up to 8 attachments, with images up to 10 MB each and 25 MB together.
A half-typed message isn't lost. The text you've typed is saved with the project and comes back after a restart, even after a crash. Attachments aren't saved with it, so add those again.
What you can ask it to do
Edit the timeline. Cut clips at a point, trim them, move them, duplicate them, and delete them. Delete or disable whole time ranges, with short audio fades so mid-speech cuts don't click. Add media to the timeline by name, at a time you choose.
Adjust how clips look and sound. Set clip volume, add or remove audio fades, turn clips on or off, and change a clip's position, scale, rotation, opacity, and crop. It can apply a change across a whole time range in one step, like a zoom that covers a section of the video.
Work with tracks, markers, and links. Mute, solo, lock, hide, or show tracks. Create, move, and rename markers, including range markers. Link or unlink audio and video clips.
Import media and work with files. Ask it to import files into the media bin by their location on disk. It can also read and write files on your machine, which is one of the things the approval system below watches.
Fix the transcript. It can edit transcript wording, and the pills on the timeline update just as if you'd corrected the words yourself.
Answer questions about your project. Ask what's on the timeline, what's in your transcripts, or what the AI tools found. When you ask it to dig through a lot of transcript material, it hands the search to a helper that only reads, so the main conversation stays responsive.
It has no export command. When the edit is ready, run the export yourself. See Export to Premiere Pro.
Watch it work
Every action gets its own row in the conversation, written in plain English, like Delete 14 ranges or Read timeline summary, with a spinner while it runs and a checkmark, or a red cross, when it ends. A count at the row's right edge sums up the result, for example 3 clips.
Rows with more to show carry a small arrow. Click the row to expand it: chips name the clips it touched, and the details of the call sit below them. Answers from the assistant appear as normal chat messages, and a thin line between exchanges shows when each one started.
You don't have to wait for it to finish. Type a follow-up mid-run and it shows above the message box with a queued tag, then goes to the assistant as soon as the current step finishes. That's also how you redirect it when you see it heading the wrong way. To halt it instead, click Stop, which replaces Send while a run is going.
Follow its plan
When a job takes several steps, a Plan strip pins itself under the panel header: a checklist with a count like 3 of 7, a spinner on the step it's doing now, and checkmarks on the done ones. It updates as the assistant works, across as many exchanges as the job takes. The arrow button at its right end collapses it to one line.
Approve before it acts
You decide how much freedom the assistant has. The control in the panel header (hover text Approval mode) offers three modes, and the same choice lives in Edit → Preferences… under Agent:
- Always Ask asks before every action.
- Auto Approve, the default, lets editing actions inside your project run on their own. Anything outside the project asks first.
- Never Ask runs everything without asking.
When an action needs your OK, the run pauses on a card titled Approval needed. It names the action, and anything reaching outside your open project is flagged Outside the current project folder. You can:
- Approve it as-is.
- Deny it, with an optional reason the assistant reads.
- Edit any value on the card first, then approve. The assistant runs your corrected version.
Sending a new message while a card is waiting denies those pending actions, and the card says so. Decisions stay in the conversation as approved and denied lines, and they're still there after a restart.
Answer its questions
When the assistant needs your input, its options appear as small buttons under the question. Click one and it's sent as your reply. The buttons only work while the question is the latest thing in the conversation and no run is going. Type a free-form answer instead whenever you prefer.
Undo a whole turn
After an exchange that did real work, a summary bar sits under it: how many edits it made, how long it took, the tokens it used, and what it cost. Next to those, Revert turn undoes everything that exchange did, the timeline changes and any files it wrote, in one step. It works even if your own edits landed afterwards.
If some of it can no longer be undone cleanly, nothing is half-reverted. The bar reads Couldn't revert cleanly with a count of the conflicts, and offers Revert clean subset to roll back everything except the conflicted parts. A partial revert like that fixes the timeline only. Files the exchange wrote aren't restored.
Plain Ctrl+Z still works too. Each of the assistant's edits sits in the undo history with a readable name, like any manual change.
Cost and memory
A chip in the panel header shows the model the assistant is using and what the session has cost so far, for example $0.42. Hover it for the request and token counts.
Next to it, a small meter shows how full the assistant's working memory is, as a percentage. When the conversation grows past what it can hold, the oldest exchanges drop out of its memory, and a note in the conversation tells you when that happens.
How its edits behave
The assistant's edits land on the timeline exactly like your own. The timeline redraws live while it works. Markers created by the assistant draw a small sparkle badge on the timeline ruler, so you can tell them apart from your own. Other edits carry no badge.
Locked tracks are respected. If you ask for an edit on a locked track, the assistant refuses and tells you why.
When it saves or starts a tool
Ask the assistant to save and it goes through the same save the File menu uses. The title bar shows the right project name, the unsaved-changes asterisk clears, and the status bar's Last saved: time updates. A Save As it performs renames the project to the new file's name, exactly as it would if you did it yourself. No save dialog ever pops up on you.
A tool run the assistant starts is an ordinary run. It appears in the Background Tasks popover with its progress, you can cancel it there like any other, and anything it produces for review lands in the Review tab. See Run an AI Tool.
Each project keeps its own conversation
Switching projects switches the conversation too. The panel shows the project you open, with its own history and its own saved draft, and everything is still there when you come back, even after closing the app.
The Terminal tab
Next to Agent sits Terminal, a real command line for driving the assistant the old way. It runs on Claude Code, a separate command-line tool from Anthropic, which must be installed on your machine and logged in with your own Claude account. If Claude Code isn't installed, the tab still opens, but you get a plain command prompt and a message that the Claude CLI wasn't found.
The tab's header shows the terminal's state, Running, Stopping…, Stopped, or Restarting…, with a Stop button while it's running. When it's stopped, a card reading Agent stopped covers the terminal with a Restart button and the hint "or press Enter".
To interrupt the terminal mid-task, click into it and press Ctrl+C. This only works when nothing is selected. If clips are selected, Ctrl+C copies the clips instead, so if the interrupt doesn't land, deselect first and try again.
The Rough Cut tab
While an Agentic Rough Cut run exists, the panel gains a third tab named Rough Cut, with a dot while the run is active. It's a watch-only view of that run, separate from your conversation, and you can't type into it. See Build a Rough Cut.