Context
Small studios lose their hours to plumbing. The creative work gets finished, then a second, technical shift begins: matching images to a brand, shipping the piece to a live site, wiring tools that were never built to talk to each other. Studio OS names that problem context collapse and removes it. It is an execution layer, not another place to track tasks. Agents own the things a studio actually delivers, and workflows follow real client work instead of checklists. The first version proved one loop end to end: start with a piece of content, finish with a live, published article, images and all, without opening a terminal or writing a prompt. It runs every day on Square Mile Design and on What If? Magazine.
What was built
Studio OS runs in three parts, each in its own place. The web app is the brain: React and TypeScript with Claude running the logic, held in the cloud, owning the interface, the agent orchestration, and every decision. The bridge is the hands: a single Go binary on a locked local port, dumb on purpose. It runs commands and reports what it sees, nothing more, so improvements ship from the cloud without ever updating the installed piece. The engine does the heavy lifting: ComfyUI makes the images and Kohya trains the models, both on a local NVIDIA RTX A6000 with 48 GB of memory. Setup is a real Claude agent that scans the machine, writes an install plan, and runs it while the user watches; a simulate mode walks the whole flow with no GPU. The Brand Studio is the edge: reference images go in, the system reads the hardware and builds a training config, and one click locks a studio's look onto every image the workspace makes after that.
What was learned
The core idea had to bend to how the models behave. Brand began as a trained file, but FLUX schnell could not carry a strong style without falling into noise, proven across two full training runs before it was called dead. The fix was a different model, Chroma, and a different premise: brand stored as a locked block of style text, with a trained model demoted to an option. The build also learned that a setup spread across two machines is fragile. One wifi reset, one closed firewall port, and a nightly job looks dead. That lesson became the next build: two dashboards that turn recovery into a status light. The full story, objectives, architecture, what broke, and where it goes next, is in the case study.