What I asked the agent: I connected invideo's MCP server to Claude and directed it cut by cut: "turn the inside of this door into a portal", "use the real shape of the tunnel hole, not a rectangle", "open all the arch gaps under the bridge", "make the train body the portal, not the sky", "add the same effect to the locomotive". For each one it captured timeline frames, measured the shape and wrote a custom shader transition on the cut: the next clip appears inside the shape, plays, then the camera pushes through. Moving shapes (the approaching locomotive, the drone drift on the bridge) are keyframed, not tracked.