diff --git a/AGENTS.md b/AGENTS.md
index eb07e04..ac7604a 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -19,20 +19,29 @@ generation endpoint and serves a small vanilla HTML/JS playground.
   `noEmit: true`, so plain `bunx tsc` works too).
 - Tests / lint / formatter: none configured. If adding tests, use `bun test`.
 
-The server binds `0.0.0.0` (see `index.ts:6`), so it is reachable from other
+The server binds `0.0.0.0` (see `index.ts:61`), so it is reachable from other
 hosts on the network when running locally — be mindful when entering API keys.
 
 ## Architecture
 
 - `index.ts` — the entire backend. One `Bun.serve` instance with:
   - `/` serves `index.html` via Bun's HTML import (`import index from "./index.html"`).
-  - `POST /api/generate` accepts `{ baseURL, apiKey, model, prompt, size }`,
-    builds an OpenAI-compatible provider with `@ai-sdk/openai-compatible`, and
-    calls `generateImage` from `ai`. Images come back as base64 and are
-    returned as `data:` URLs in `{ images: string[] }`.
+  - `POST /api/generate` accepts
+    `{ baseURL, apiKey, model, prompt, size, referenceImages? }`. It returns
+    `{ images: string[] }` where each entry is a `data:` URL (base64).
+  - Two code paths inside the handler:
+    1. No `referenceImages` → uses `@ai-sdk/openai-compatible` + `generateImage`
+       from `ai`.
+    2. `referenceImages` present → hand-rolled `multipart/form-data` POST to
+       `${baseURL}/images/edits` (see `generateWithReference`). The AI SDK
+       does not currently expose image edits for OpenAI-compatible providers,
+       so this path bypasses it on purpose. The edits endpoint is gpt-image
+       series only (see UI hint in `index.html`).
 - `index.html` — self-contained UI: inline CSS, plain DOM JS, no build step.
-  Settings (baseURL, apiKey, model, size, prompt) persist in `localStorage`
-  under the `aip:<field>` prefix. There is no React code despite
+  Text fields (`baseURL`, `apiKey`, `model`, `size`, `prompt`) persist in
+  `localStorage` under the `aip:<field>` prefix. Reference images are kept
+  in an in-memory `refImages` array as base64 data URLs and are **not**
+  persisted — refreshing the page drops them. There is no React code despite
   `react` / `react-dom` / `@types/react*` being in `package.json` — treat
   those deps as latent. Do not invent a React frontend unless asked.
 - No router, no DB, no auth. API key is supplied per-request by the browser
@@ -59,3 +68,7 @@ hosts on the network when running locally — be mindful when entering API keys.
 - The AI SDK image type is loose; the current handler casts to
   `{ mediaType?: string; base64?: string }`. Mirror that pattern rather than
   trusting field presence.
+- For anything the AI SDK does not cover (e.g. image edits, masks, variations),
+  follow `generateWithReference`: build `FormData` with `Blob`s decoded from
+  the incoming data URLs and `fetch` the upstream endpoint directly with the
+  caller's `Authorization: Bearer <apiKey>`.