A browser-first AI image workspace. Bring your own OpenAI-compatible API key, generate images through a chat-style interface, and keep history on-device.

• Chat-style workspace at /chat: each prompt produces an independent image, rendered inline as a message thread.
• Per-request controls: model, size, and quality are exposed as compact chip selectors above the composer.
• Local-only persistence: chat history lives in SQLite (WASM) backed by the browser's Origin Private File System (OPFS). No backend stores your prompts or images.
• Bring your own credentials: API key, base URL, and default model are kept in localStorage and sent to the server only at generation time.

pnpm install
pnpm dev

Open http://localhost:3000 and click Start creating to enter the chat workspace. Provide your API key in the Settings drawer before generating.

The postinstall / predev / prebuild hooks copy SQLite WASM runtime assets from node_modules/@sqlite.org/sqlite-wasm/dist into public/sqlite-wasm/ so the worker can load them at runtime.

The chat history database runs in the browser through @sqlite.org/sqlite-wasm with the OPFS VFS. OPFS access requires the page to be cross-origin isolated, which in turn requires these response headers on every document:

• Cross-Origin-Opener-Policy: same-origin
• Cross-Origin-Embedder-Policy: require-corp

These headers are applied globally in next.config.ts.

Consequence: any cross-origin resource the page loads must opt into being embedded (Cross-Origin-Resource-Policy: cross-origin or matching CORS). Generated images returned as URLs by the upstream model would fail to load under this policy, so /api/generate always materializes the image into a data: URI server-side before responding.

If crossOriginIsolated is false (rare browser extensions can strip COEP), the chat database falls back to an in-memory SQLite instance, history will reset on refresh, and a warning toast is shown.

Single messages table; one user row and one assistant row share a turn_id. Errors are stored as assistant rows with an error column, so failed generations remain visible after refresh.

• Multi-turn iteration / image edit API. Today every prompt is independent — there is no way to say "make the previous image warmer." Wiring OpenAI's image edit / variations endpoints would let a user reference a previous assistant message and refine it. This is the largest UX gap and the next major feature.
• Image export and per-turn delete. Right now history can only be cleared in bulk via the topbar.
• Multi-tab safety. Opening the workspace in two tabs simultaneously will cause the second OPFS open to fail; it currently falls back to :memory: silently for that tab.
• Provider auth presets. Streamline switching between OpenAI, Azure OpenAI, and other compatible endpoints.