Crucible
Grades what an AI agent just wrote or drew, and names the fix.

An MCP server and CLI that checks game content an agent just generated, then hands back a specific named fix so it can try again. It never calls a model to do that. One tier is deterministic pattern and pixel analysis; the other hands the calling agent a named rubric item and records the verdict it comes back with.
An agent writes a line of dialogue or draws a sprite, and that's usually where it stops, because "make it better" isn't an instruction a model can act on. Crucible turns it into a loop instead. Generate, check against named craft knowledge, get a specific fix, regenerate.
The prose side reads plain dialogue and flags the usual tells: said-bookisms, adjacent lines saying the same thing twice, stock phrases like "delve into" and "a tapestry of", and the balanced contrast template you start seeing everywhere once you look for it. One of those on its own is fine and gets logged as information. Three in the same file promotes every instance to a failure, because at that point it has stopped being a rhetorical choice and become a tic.
Sentence rhythm is the check I like most. It takes the coefficient of variation of words per sentence and warns when every line takes the same beat to read, which is what generated writing does. Under four sentences it says nothing at all rather than drawing a conclusion from too little data.
The visual side takes pixel art authored as SVG and rasterizes it with anti-aliasing off, which is the only thing that makes one SVG unit really be one pixel. Without that the detectors would be reading blur instead of a grid. Then it goes looking for banding, jaggies, stray fragments and sealed holes, and reports them at pixel coordinates. The banding check segments regions and measures how straight a shared border is, because a band and a border following a curved silhouette look identical if all you do is count colors.
The rubrics are JSON and they cite where they came from: Derek Yu on pixel art, Thomas and Johnston, James Gurney. There is no default style profile, deliberately. Ship one and every project that installs this starts sounding the same.
What it does
- Deterministic tier needs no API key; the rubric tier the agent judges
- Catches stock phrasing, said-bookisms and balanced-contrast templates
- Sentence rhythm scored statistically rather than by keyword
- Pixel art rasterized with anti-aliasing off, so detectors read a real grid
- Agents can't raise their own iteration budget once a loop has started
Screens


