Agent Access in Cask Flow
Cask Flow is built to be driven by an agent. Connecting one is the same job in both apps, and Agent Access covers it. This page is what is different once a client is connected to Cask Flow.
What the agent sees
Section titled “What the agent sees”The wire is the same model you see: one composition, one tree of nodes, composition-local seconds, one camera. An agent addresses a node by id and a property by name. There are no clips and no clip-local time, so instructions written against Cask Film’s tools do not apply here.
76 tools cover it, 26 of which only read. Your client shows the list, and each tool carries a plain-language title plus flags saying whether it only reads, whether it destroys something, whether a retry is safe, and whether it reaches the backend.
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 a single frame at a time you name, exactly as export would produce it.
- The other renders up to twelve frames across a range as one captioned contact sheet, so a move can be judged from its frames side by side in one call.
Images ride in the message content only. A schema-checking client never sees a megabyte of base64 inside the structured result.
The agent is briefed before it starts
Section titled “The agent is briefed before it starts”Alongside the tool list, the app ships prose: the mental model, the workflow, and six deep-dive topics covering nodes, keyframes, the camera, export, generation and workflow. Your client injects the overview into the model’s system prompt, and the agent can pull any topic mid-task.
That prose comes from the running app rather than from the bridge binary, so it tracks the app version you actually have open. With the app closed, the bridge answers with a short fallback that says the real guidance arrives when you open the app.
Sign-in gates almost everything
Section titled “Sign-in gates almost everything”While nobody is signed in, every tool except the two that report capabilities and guidance is refused. The capability tool reports that no one is signed in, which is the call an agent makes first to find out.
Signing in is a human action in the app window. No tool can do it.
Watching and undoing
Section titled “Watching and undoing”The right pane’s Agent tab shows the connection, the clients attached, and the activity as it happens. The Changes tab is the undo history itself, newest first, with Undo on each step.
One tool call is one undo entry, labelled with the tool that made it. Your ⌘Z and the agent’s edits share one stack, so taking over by hand needs no handover.
Checkpoints
Section titled “Checkpoints”An agent can snapshot the whole project, list its snapshots and restore one. That is a coarser tool than undo, and it is what to reach for before a change that would be tedious to unpick step by step.