# Chat and your agent

CutZal doesn't run an AI model of its own. Connect the Claude Code or Codex you already use, and chat and Auto cut run on that subscription. Usage comes out of your subscription, and CutZal doesn't charge you for it. Without a connection, you can still use the default cut, keyboard review and export.

## Connect

1. Install Claude Code or Codex in the terminal and sign in. Claude Code needs a paid Claude subscription (Pro or higher), and Codex needs a ChatGPT account. The free Claude plan can't use Claude Code, and API key connections aren't supported.
2. Check the connection status under **Agent** in CutZal **Settings**. It checks as soon as you open settings, and when it's connected it shows something like "Claude Code · subscription connected". If it isn't connected, a note and **Open settings** appear below the chat input. Once it has connected, chat checks again on its own the next time you open CutZal, so there's nothing more to do.
3. If it isn't installed, the version is too old or you need to sign in, CutZal tells you why and shows the install, update or sign-in command. Copy it with **Copy command** and paste it into the terminal. For more detailed installation steps, **Install guide** opens the official docs.

CutZal never receives your sign-in details or keys. It just runs the program you signed in to in the terminal, as it is. Checking the connection doesn't send video or transcripts.

## Provider terms and your account

Chat and Auto cut run on your own Claude or ChatGPT plan, so each company's terms apply as they are.

- Usage comes out of your plan's usage limits. CutZal doesn't pay for you or resell usage.
- Claude Code is covered by Anthropic's [Consumer Terms](https://www.anthropic.com/legal/consumer-terms) and [Usage Policy](https://www.anthropic.com/legal/aup), and Codex by OpenAI's [Terms of Use](https://openai.com/policies/terms-of-use/) and [Usage Policies](https://openai.com/policies/usage-policies/). Anthropic says the usage limits of its Pro and Max plans are set for ordinary individual use.
- You only sign in on each company's official sign-in screen. CutZal never asks you to enter your Claude or ChatGPT account details in its own screens.

### Why CutZal has no sign-in window

Anthropic's [legal and compliance page](https://code.claude.com/docs/en/legal-and-compliance) doesn't allow other companies' apps to take Claude account sign-ins themselves or to store sign-in details. The same page says it doesn't stop people from signing in to an unmodified Claude Code with their own subscription and using it. So CutZal leaves signing in to the official program and runs the program you signed in to as it is, without changes. Codex connects the same way.

### If a provider's policy changes

- If the current way of connecting is no longer allowed, CutZal turns that connection off in an update and shows why under **Agent** in **Settings** and below the chat input.
- Even with the connection off, you can keep using the default cut, keyboard review and export. If the other provider is connected, you can keep chatting with its models. Auto cut only runs on Claude.
- Any action on your account follows each company's terms.

## Hand off edits in chat

Type what you want in the chat panel and send it. For example: "Trim the repeated explanations in the selected range and keep the complete sentences."

- Scope: select transcript lines or timeline blocks before you send, and chat works on just that range. With nothing selected, it works on the whole video. The chip above the input shows **Whole video** or the range you selected. You can also add a range with **Attach this range to chat** in the right-click menu of a final timeline block.
- What you can hand off: cutting and restoring, placing sound effects, changing the layout, making motion graphics and going back to a checkpoint. Chat can't transcribe, export or sign in.
- Model: pick one from the model chip below the input. Models in the Claude group run on Claude, and models in the GPT group run on your connected Codex. Each group marks its recommended model. Use the **Reasoning effort** chip next to it to set how hard the model thinks. A new conversation starts at each model's default effort, and switching models switches to that model's default. A session stays with the provider of the model you first sent with, so to switch providers, open a **New session**.
- Images and references: send up to 4 reference images along with your message using **Attach images** under **Attach**. You can also drag them onto the input or paste them. Connect files or folders such as a plan to the conversation with **Connect references**, and the AI can read them.
- Long instructions: click the expand button on the input, and the same input opens wide in the middle of the screen so you can see and edit all of it. Your text, attachments and scope stay as they are. Press `Esc`, click the same button or click outside to put it back. If you send while it's expanded, it collapses after sending. Instructions can be up to 2,000 characters, and the expanded input shows the character count.

The line below the input shows which subscription is in use, and **Session info** under **More** at the top of the chat shows what is sent where. Conversations are saved in the project folder, so they carry on after you reopen the app.

## Collect spots and send them together

While you review, things to fix turn up here and there. Instead of sending each one separately, you can collect them in the **Send list** and send them all at once.

1. Select a spot to fix and press `C`. If you selected transcript lines, a timeline range or a track clip, that spot is added. With nothing selected, the playhead position is added. If you select several separate spots and press it, each spot is added on its own.
2. A note field opens in the list above the input. Write something short like "keep this sentence" and press `Enter` to go back to where you were. You can also write the note later.
3. When you're done collecting, click send. Anything you write in the input becomes an instruction for the whole list, and you can leave it empty. Up to 10 spots go out in one request.

- Even if you change cuts earlier in the video, collected spots keep pointing at the same words. If a spot was cut entirely, the list shows it as "Cut".
- Turn on the eye button and the AI also looks at the frame at that time. Turn it on for notes that need the picture, such as the camera size.
- Results come back for each spot as "Done" or "Not done" with a reason. Spots it couldn't do go back to the top of the list with the reason, so you can edit the note and send again. If the request fails or is stopped, every spot you sent comes back.
- If two notes touch the same spot, the one added later wins.
- While chat is working, you can keep adding spots and writing notes. Sending waits until the current request finishes.
- One-key jobs like cutting or restoring are faster to do directly than to add to the list. Use the list for work you hand to AI, such as sound effects, layout and deciding whether to keep something.

## Review and apply

Use the permission chip below the input to choose when chat's proposals go into your edit.

- **Plan mode**: AI checks the picture and your materials and only answers with a plan and proposals. It doesn't change your edit or make motion graphics. You can put a proposal in yourself with **Apply changes**. To have it carry out the plan, switch the mode and send something like "go ahead as planned".
- **Always ask**: new sessions always start in this mode. AI shows what it will change first, and nothing goes into your edit until you click **Apply changes**. If you change the edit after seeing a proposal, that proposal can no longer be applied, so send the same request again.
- **Accept edits**: deletes, restores and sound effect placements apply right away.

In **Always ask**, when the agent places motion work items it asks once, with the list of items it wants to place. Select a row and that work item plays on the right; use the up and down arrow keys to step through them. Uncheck any you don't want (Space) and place the rest. On your phone, tap a title to preview it below the row. Turn on **Don't ask about placing in this conversation** in that window and placements in that conversation stop asking, while other edits like cuts and restores still ask. To have it ask again, turn off **Don't ask about placing in this conversation** under **Work item placement** in the permission chip menu.

### Watch the progress

While chat works, the agent's explanations and tool runs stack up in the conversation in order. Click a tool line to expand the raw request and result for that tool. The model, tokens and estimated cost are folded into **Task details** below each turn. The AI's thinking process isn't shown. Bold text, lists, tables and code in its explanations and answers are formatted for easy reading, and links open in your default browser.

When you're at the bottom, the conversation follows new content. While you scroll up to read or expand something, it stays where you are. Click the down arrow button or scroll to the bottom to follow again.

The range chat or Auto cut is working on shows as a label and outline on the timeline and as a border in the transcript. Ranges already cut from the final version are marked at the join and on the folded deletion line. Your view doesn't move when the work position changes. To go to the work position, click the status text in the timeline toolbar. You can also assign a key to **Go to agent location** in **Editing · Shortcuts** settings.

Cuts and restores that were actually applied light up briefly where they changed. With reduce motion turned on, this flash is off and only the work position shows. If more instructions are queued, the status line also shows how many are waiting.

### While chat is working

When an external agent announces an editing session, the status line shows its name, its task and when it last sent a request. If you take over with **Take over**, the agent waits, and when you click **Return to agent** after your fixes, your changes go to that same session. If you don't want it to continue, click **End external session**. A run of consecutive agent edits is undone with a single undo, and edits you make in between are kept as separate history.

While chat works, editing is locked, because if your edits and AI's changes collide, the whole chat request fails. You can still play, select ranges, listen to the original and write your next instruction.

- While it's locked, the middle of the timeline toolbar shows the chat model, the current step, the elapsed time and **Editing locked**, and the timeline is hatched. Next to them are the **Take over** and **Stop** buttons. If you try to edit while it's locked (cut, restore, undo and so on), the reason and a **Take over** button appear once more.
- A border around the window shows when the agent has editing control. It clears when you take over. The progress and chat screens on your phone also show when the agent is waiting, and you can edit the script on your phone while it waits.
- Click **Take over** to pause chat and edit right away. Chat doesn't end. It waits. The toolbar shows **You're editing**, the waiting chat model and the elapsed time. If chat was in the middle of applying an edit, the button changes to **Taking over** and waits until that edit finishes.
- When you're done, click **Return to agent** in the toolbar. Chat continues the earlier request in the same conversation, treats the spots you fixed as your decisions and doesn't revert them. Writing and sending a message in the waiting conversation's input also hands it back, and what you wrote goes along as an extra note. Above the returned request, **Returned to agent** and a list of the spots you fixed stay visible.
- If you don't want it to continue, click **End**. A queue that paused while chat waited stays, and you can send it again with **Continue sending**. Sending a separate new instruction or starting Auto cut also ends the wait.
- **Stop** only stops chat and doesn't wait. While it stops, the button shows **Stopping**. Either way, edits chat already applied stay, and you can go back with a checkpoint.
- Edits from your phone and from terminal agents aren't accepted until chat finishes either. The script screen on the phone stays visible, and only the cut and confirm buttons turn off, with the reason.
- While Auto cut runs, only media and cut settings are locked. You can still listen to the cuts that came out first and restore them. To change media or settings, click **Stop auto cut**.

You can undo edits that chat applied with **Undo** in the timeline toolbar or `⌘Z`. Each time chat applies something, the moment just before it is kept as a checkpoint, so you can also go back with **Go back to before this**. Pick an earlier checkpoint from **Checkpoints** under **More** at the top of the chat. If going back would also undo edits you made by hand after that point, it asks you to confirm first.

Whatever the permission mode, some things always need your own click. These are the things that cost money or allowance, or are hard to undo.

- Starting transcription ([Transcription options](https://cutzal.ai/docs/en/transcription.md))
- Starting Auto cut ([Review loop and AI cut grades](https://cutzal.ai/docs/en/review.md#auto-cut))
- Creating AI sound effects: agreeing to the terms and clicking create
- Placing motion graphics work on the timeline
- Exporting ([Export](https://cutzal.ai/docs/en/export.md))

## Start with an instruction from the first screen

Type what you want in the input on the first screen, attach a video file and click **Send**, and the **Import media** window opens. When importing and transcription finish, the instruction you wrote is sent to chat once. This also needs your Claude Code connected and uses your subscription allowance. To keep it from being sent, click **Remove instruction** on the waiting instruction.

## Use an agent in your terminal

You can also have Claude Code or Codex in your terminal read and edit CutZal projects with the `cutzal` command. It's handy for automation or for working on many projects at once.

1. In **Settings**, under **Agent**, find **External agent connection (CLI and Skill)**.
2. Under **Install command and Skill**, click **Install** for the `cutzal` command, the **Claude Code Skill** and the **Codex Skill**. Only when you click does it create links at `~/.local/bin/cutzal`, `~/.claude/skills/cutzal` and `~/.codex/skills/cutzal`, and it never touches other files with the same name.
3. To hand work to an agent while CutZal is open, turn on the switch next to **External agent connection (CLI and Skill)**.

With the connection on, the agent's edits come in through the running CutZal, so you can undo them with undo and checkpoints just like your own edits. If CutZal has the project open and the connection is off, the agent's edits are refused. That keeps the two from overwriting the same file. When CutZal isn't running, the agent edits the project directly.

The agent can read who changed what since it last looked with `cutzal changes --since`. When a request is refused because the edit changed, it also gets the change history. The official Skill guides the agent to check the sections a person edited by hand or restored on the phone, keep those changes and plan again. If the records are too old or have gaps, it needs to check the current project again.

There are also things an agent can't do, or can only do with your confirmation.

- If it asks for an export, CutZal's export window opens, and you need to click **Start export**.
- If a request would also undo edits you made by hand, the **Go back requested by an external agent** window appears. Choose **Go back** or **Decline**.
- An agent can't start credit transcription, transcription model downloads or AI sound effect creation.

Give your agent just one address: `https://cutzal.ai/llms.txt`. The commands and the order of work are in [CutZal for AI agents](https://cutzal.ai/docs/en/agents/index.md). If you moved the CutZal app to another folder, click **Reinstall** to fix the links.

## Save edit settings as a Skill

You can save edit settings you use often as a Skill and apply them to other projects. Create one in **Skills · Plugins · MCP** in **Settings**. Then, in the chat panel's **Apply skill to project**, check what will change with **Preview settings** and click **Apply skill settings**. For now, the silence cut threshold and how B-roll is shown are applied, and edit instructions you wrote along with them are only saved.