Blender

Two ways in, and they are not interchangeable.

Interactive, through the MCP server, talking to an add-on inside a running Blender. The user can see everything you do. Their viewport, their unsaved edits, their file.

Headless, blender -b file.blend --python script.py -- args. Its own process, no UI, nothing shared. This is where renders and long sweeps belong.

The rule that decides which: does this touch state the user is looking at? Use the live MCP for inspection and user-authorized edits to the open scene. Use headless Blender on an explicit copy for long renders or batch processing; never replace the user's open file with a headless result without reconciling live edits. Choose the route by the owned document and task, not merely by whether it writes a file.

For longer film work, load the relevant reference file before changing the assembly or render plan:

• Film production and delivery: integrating scene candidates, story and behaviour clocks, camera movement, lighting, sourced assets, sound continuity, frozen inputs, resumable jobs and complete-film review.
• Render performance: separating elapsed time from rendering work, controlled resolution/cache tests, device verification, preview settings, motion quality and measured estimates.

Give one lead ownership of the assembled scene, route and timing contract, and one process ownership of the GPU. Prove the smallest unresolved action from the actual film camera, then review the complete encoded journey. Keep a requested sequencing/sound approval before the expensive final render. A Blender film does not imply permission to build a game-engine version.