Agent Access in Cask Film
Cask Film has two ways in for an agent. The agent panel runs your own Claude Code or Codex beside the editor. Agent Access connects an outside MCP client instead. Both reach the same tools. This page covers the panel and what any agent sees once it is driving Cask Film.
The agent panel
Section titled “The agent panel”Click the sparkle in the editor toolbar, or press ⇧⌘A, to open the panel. ⇧⌘N starts a new conversation. ⌘Return sends, and Esc stops the current turn.
The panel drives the Claude Code or Codex you have installed and signed in to, on your own subscription. It needs Node as well. If several copies of the CLI are installed, Cask Film uses the newest one it finds, so the panel’s model list matches the CLI you keep updated. A missing or signed-out CLI shows the install command instead of a conversation.
Your own skills are off
Section titled “Your own skills are off”The panel runs on Cask’s own guidance. It leaves out the skills, plugins, instructions, hooks and MCP servers installed in your Claude Code or Codex.
To bring them in, turn on Settings > General > Agent Panel > Use my own agent skills. The switch applies to conversations you open next. Open ones keep the setting they started with. With the switch off, Codex also ignores your ~/.codex/config.toml, while keeping your sign-in.
How much it decides
Section titled “How much it decides”The mode picker sets how far the agent goes on its own:
- Read-only: looks and reports, changes nothing.
- Guided (the default): suggests choices and waits for your review.
- Full auto: chooses and continues through reviews. It can spend credits, export and delete without asking.
It works like an assistant director
Section titled “It works like an assistant director”You direct. The agent knows film craft (story, coverage, film grammar, sound, dialogue, genre) and works one stage at a time: idea, story, look, shot list, storyboard, test shots, generation, edit and finish. At each choice it offers two or three options and recommends one.
It opens by asking about the idea, the feeling, the audience, the length, your experience and the budget, three questions at most at a time. Use Suggestions lets it answer for you. A beginner gets a default workflow with each step explained.
The brief and each version of the film’s documents are saved with the project. When a stage is ready, a review card appears in the conversation with Read Version, Approve and Request Changes. Only you can approve. Earlier versions stay, and changing the brief marks later work as out of date.
The agent proposes paid generations one at a time, with a budget for the whole film up front. After it starts a generation it waits for it to finish, then looks at the result against the prompt and the shots around it, and tells you plainly whether it holds up. It proposes a paid retake rather than starting one. For audio, it asks you to listen.
Each tool step in the conversation reads as plain words, for example 2 cards · frame rate 24, and a run of steps folds into one line.
What the agent sees
Section titled “What the agent sees”203 tools, 61 of which only read. They cover the canvas, the timeline, the library, captions, voices, scenes, generation and export, in about the same shape as the editor itself. An agent can build and render a scene in Scene Explorer, and use your saved voices, but it can’t make or delete a voice.
observe is the call to make first. It reports whether a project is open, its summary, the current selection, entity counts and undo depth, and it is the cheapest thing that proves the whole path works.
The one rule an agent gets wrong
Section titled “The one rule an agent gets wrong”Generating a shot, or importing a file, puts it on the canvas and in the library. It does not put it in the video. Export reads the timeline and nothing else.
An agent that generates four shots and reports the scene finished has produced four canvas cards. The clips have to be added to a timeline track before any of it is in an exported file. See the canvas and the timeline.
The agent can look at its work
Section titled “The agent can look at its work”Two tools return an actual picture into the model’s context. One renders the composed timeline frame at a time you name, exactly as export would produce it. The other renders a single canvas card.
A render of the timeline is the check that matters, because it is the only one that shows what export will produce.
A cloud project can refuse every write
Section titled “A cloud project can refuse every write”A cloud project open on another Mac is read-only here, because that Mac holds the edit lease. Every write tool refuses and says so, while reads and export keep working. Take Back Editing in the app moves the lease.
Watching and undoing
Section titled “Watching and undoing”Every change an agent makes is one undo entry, and your ⌘Z and its edits share one stack. You can take over mid-task with no handover step.