CPA Image is built for people who'd rather not hand their image workflow to a managed third-party SaaS. It's a local-first, OpenAI-compatible image console — requests are sent directly from the browser to your configured API endpoint. Workspace settings, prompt history, and generation artifacts are stored locally in the browser (localStorage / IndexedDB) and persist across browser restarts rather than automatically clearing; API keys are only persisted locally when "Remember API key" is selected. By default, connections are direct; only when cross-origin proxy is manually enabled will requests, keys, and images pass through a third-party proxy. One panel covers every output path in the OpenAI-compatible ecosystem: text-to-image via /v1/images/generations, local image edits via /v1/images/edits, and tool-call image output via /v1/responses and /v1/chat/completions — switch on demand without changing clients. Sizes are grouped by aspect ratio and pinned in exact pixels — auto, 1024x1024, 2048x2048, up to a 3840x2160 long edge — with the UI tagging the near-2048 / near-3840 tiers as 2K / 4K for legibility, all returned natively by the endpoint with no post-processing upscale; quality is selected from auto / low / medium / high. Batch and throttling are yours to define: run up to 100 at once split into independent tasks, with self-throttled concurrency and request intervals, where a single failure doesn't sink the batch and runs are auto-numbered by time. A state machine covers queued, running, done, error, and canceled. strictPrompt locks the outer semantics to keep the model from rewriting your keywords, with the lock template customizable and toggleable on a single switch; past artifacts can be pulled back as edit baselines for chained iteration across requests without re-downloading, and failed-request error types (401, 503, upstream error) are recognized and distinguished.