film-bench

film-bench is a canvas for making short AI films one shot at a time. Boards are shots, cards are takes, and one prompt bar at the bottom generates stills and clips through OpenRouter, across several models and styles in one press. It needs Python 3 and ffmpeg on the PATH, and nothing else: no pip install, no node, no build step.

The tool addresses the problem that generating video from a shell is not an iteration loop, and form-based UIs separate the picture from the controls. film-bench applies the shape of tools like Recraft to short films, where the unit of work is a shot and the hard problem is keeping a character or an object the same from one clip to the next. Boards are shots, left to right in film order, and the order is saved because the order of the boards is the cut. Cards are takes, newest first, and frame slots at the top of each board allow dragging any still onto FIRST or LAST so the next clip is generated between those exact pixels.

Style and content are separate: a shot's prompt says what happens, and a style says what it looks like. The bench folds the two together at call time and shows the result under composed. Stills are generated first because a still costs cents and a clip costs dollars, so every look decision is made on stills. The bench starts disarmed, and while it is, the Generate button says Dry run: the exact request body and a cost estimate go to the log, and nothing is charged.

An MCP server (mcp.py) allows driving the bench from an agent, forwarding calls to the running bench's HTTP API. The server exposes 18 tools, and the arm switch is not a tool: while disarmed, every generate an agent asks for is a dry run. The gallery shows every take the bench has ever made, newest first, across every film, and an A|B comparison allows judging two takes of the same prompt side by side.