A personal generation factory for image, 3D, video, voice and music, running on one laptop. Claude calls the MCP tools itself, a machine gate screens the output, and an agent looks at the thumbnail and rules on it.
This page opens up one of the five axes: image. The others run on the same kernel, through the same gate and the same review.
Every image below was generated by this studio. None of it is product UI.
Instead of a grid of outputs, here is one real job laid open. I asked for an image to use on this page; the factory produced four candidates, approved two and rejected two. Every value below is what the machine wrote down at the time.
The subject is a stamp for a reason. What this factory does all day is rule on things — approved or rejected — so the subject and the argument say the same thing.
It starts with one sentence. A local krea2 produces the candidates and a machine gate checks resolution and pixel variance before anything reaches a reviewer. The image below is the exact thumbnail the agent was handed when it ruled.
A pipeline that only shows its successes is a gallery, not a pipeline. Two of the four candidates in this batch were rejected, and the reasons are still on them in the library.
RejectedNear-frontal rather than three-quarter, so it is a poor 3D source. The background also carries ghosted letterform artifacts that would fight background removal.
RejectedA frontal view on a round base, so the top face is barely visible — the least usable geometry of the four.
Claude calls the MCP tools directly. Nobody is pasting prompts into a web UI.
Resolution, pixel variance and similar measures drop the obvious failures first. What doesn't need a human never reaches one.
The response carries the thumbnail back as an image. The agent looks at it and rules on prompt adherence and artifacts.
Approved or rejected, with the reason, goes into SQLite. The agent's judgment stays in the library and becomes material for the next job.
This isn't a demo rig. The cover images on the blog you're reading come out of it too.
One shared kernel with a per-modality MCP server on top of it. GPU arbitration, a SQLite asset store, thumbnails and a common server skeleton live in the kernel, and several repositories were folded into a single monorepo through history-preserving subtree merges. Development runs across two machines, so machine-specific values are isolated into profiles and a six-layer fingerprint — code, Python, ComfyUI, models, runtime, machine — lets a machine answer whether the two are actually in the same state.