diff --git a/app.go b/app.go
index eabc3b5..62aa73c 100644
--- a/app.go
+++ b/app.go
@@ -22,6 +22,7 @@ import (
 	"github.com/mmmy/snapgo/internal/infrastructure/config"
 	"github.com/mmmy/snapgo/internal/infrastructure/display"
 	"github.com/mmmy/snapgo/internal/infrastructure/hotkey"
+	llmpkg "github.com/mmmy/snapgo/internal/infrastructure/llm"
 	"github.com/mmmy/snapgo/internal/infrastructure/oss"
 	"github.com/mmmy/snapgo/internal/infrastructure/screencapture"
 	sshpkg "github.com/mmmy/snapgo/internal/infrastructure/ssh"
@@ -96,6 +97,15 @@ type CaptureActionResult struct {
 	Path      string `json:"path,omitempty"`
 }
 
+// OperationStatusPayload drives the frontend status HUD and mirrors the
+// native macOS HUD state.
+type OperationStatusPayload struct {
+	Operation string `json:"operation"`
+	Phase     string `json:"phase"`
+	Message   string `json:"message"`
+	State     string `json:"state"`
+}
+
 // NewApp creates a new App with collaborators already initialised.
 func NewApp() *App {
 	store, err := config.NewFileStore()
@@ -110,6 +120,7 @@ func NewApp() *App {
 			slog.Warn("config load failed, using defaults", "err", lerr)
 		}
 	}
+	cfg.Normalize()
 	return &App{
 		cfg:        cfg,
 		configFile: store,
@@ -297,7 +308,13 @@ func (a *App) runUploadPipeline(provider domain.OSSProvider, pngBytes []byte) er
 		FallbackDir: filepath.Join(userPicturesDir(), "SnapGo"),
 		PathPrefix:  cfg.S3.PathPrefix,
 	}
-	return svc.ExecuteWithBytes(a.ctx, pngBytes)
+	a.emitOperationStatus("upload", "上传中", "正在上传截图到 S3", "running")
+	if err := svc.ExecuteWithBytes(a.ctx, pngBytes); err != nil {
+		a.emitOperationStatus("upload", "上传失败", err.Error(), "error")
+		return err
+	}
+	a.emitOperationStatus("upload", "上传完成", "链接已复制到剪贴板", "success")
+	return nil
 }
 
 func (a *App) runCopyImagePipeline(pngBytes []byte) error {
@@ -332,6 +349,7 @@ func (a *App) runSaveRemotePipeline(pngBytes []byte) error {
 		err := fmt.Errorf("SSH host/user is not configured")
 		slog.Warn("save-remote rejected: ssh not configured",
 			"host", cfg.SSH.Host, "user", cfg.SSH.User)
+		a.emitOperationStatus("save-remote", "需要配置 SSH", err.Error(), "error")
 		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
 		return err
 	}
@@ -355,7 +373,78 @@ func (a *App) runSaveRemotePipeline(pngBytes []byte) error {
 		Notifier:  &runtimeNotifier{ctx: a.ctx},
 		Cfg:       cfg.SSH,
 	}
-	return svc.ExecuteWithBytes(a.ctx, pngBytes)
+	a.emitOperationStatus("save-remote", "保存中", "正在保存截图到远端 SSH", "running")
+	if err := svc.ExecuteWithBytes(a.ctx, pngBytes); err != nil {
+		a.emitOperationStatus("save-remote", "保存失败", err.Error(), "error")
+		return err
+	}
+	a.emitOperationStatus("save-remote", "保存完成", "远端路径已复制到剪贴板", "success")
+	return nil
+}
+
+func (a *App) runSummaryPipeline(pngBytes []byte) error {
+	a.mu.RLock()
+	cfg := a.cfg
+	a.mu.RUnlock()
+
+	if !cfg.IsS3Configured() {
+		err := fmt.Errorf("请先在 S3 配置页填写可用的对象存储配置")
+		a.emitOperationStatus("summary", "需要配置 S3", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	if !cfg.IsLLMConfigured() {
+		err := fmt.Errorf("请先在 LLM 配置页填写 API Key 和模型")
+		a.emitOperationStatus("summary", "需要配置 LLM", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+
+	s3Provider, err := oss.NewS3Provider(cfg.S3)
+	if err != nil {
+		a.emitOperationStatus("summary", "S3 配置错误", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	providerID, llmCfg, _ := cfg.ActiveLLMProvider()
+	visionClient, err := llmpkg.NewVisionClient(providerID, llmCfg)
+	if err != nil {
+		a.emitOperationStatus("summary", "LLM 配置错误", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	svc := &application.CaptureSummaryService{
+		Provider:   s3Provider,
+		Summarizer: visionClient,
+		Clipboard:  a.clip,
+		Prompt:     cfg.LLM.Prompt,
+		PathPrefix: cfg.S3.PathPrefix,
+	}
+
+	a.emitOperationStatus("summary", "上传中", "正在上传截图供模型读取", "running")
+	uploaded, err := svc.UploadImage(a.ctx, pngBytes)
+	if err != nil {
+		a.emitOperationStatus("summary", "上传失败", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+
+	a.emitOperationStatus("summary", "识别中", "图片已上传，正在请求多模态模型", "running")
+	summary, err := svc.Summarize(a.ctx, uploaded.URL)
+	if err != nil {
+		a.emitOperationStatus("summary", "识别失败", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	if err := svc.CopySummary(a.ctx, summary); err != nil {
+		a.emitOperationStatus("summary", "复制失败", err.Error(), "error")
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+
+	a.emitOperationStatus("summary", "识别完成", "总结已复制到剪贴板", "success")
+	wruntime.EventsEmit(a.ctx, "upload:success", "summary copied to clipboard")
+	return nil
 }
 
 func (a *App) consumePendingCapture() (*pendingCapture, error) {
@@ -408,6 +497,7 @@ func (a *App) GetConfig() domain.AppConfig {
 // SaveConfig persists the supplied configuration and re-registers the hotkey
 // if it changed.
 func (a *App) SaveConfig(cfg domain.AppConfig) error {
+	cfg.Normalize()
 	a.mu.Lock()
 	prev := a.cfg
 	a.cfg = cfg
@@ -705,6 +795,35 @@ func (a *App) SaveRegionToRemote(result CaptureResult) error {
 	return a.runSaveRemotePipeline(cropped)
 }
 
+// SummarizeRegion uploads the selected screenshot to S3, sends the public URL
+// to the configured multimodal LLM, and copies the resulting summary.
+func (a *App) SummarizeRegion(result CaptureResult) error {
+	pc, err := a.consumePendingCapture()
+	if err != nil {
+		slog.Warn("SummarizeRegion: no pending capture", "err", err)
+		return err
+	}
+	defer func() {
+		a.capturing.Store(false)
+		a.dismissOverlay()
+	}()
+
+	a.dismissOverlay()
+	flushFrame()
+
+	slog.Info("SummarizeRegion: capturing region",
+		"x", result.Rect.X, "y", result.Rect.Y,
+		"w", result.Rect.W, "h", result.Rect.H,
+		"annotations", len(result.Annotations))
+	cropped, err := a.captureSelectedPNG(result, pc)
+	if err != nil {
+		slog.Error("SummarizeRegion: capture failed", "err", err)
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	return a.runSummaryPipeline(cropped)
+}
+
 // SaveNativeRegionToRemote is the macOS-native overlay equivalent of
 // SaveRegionToRemote. The AppKit panel is already closed by the time this
 // runs (see saveRemoteSelection in native_overlay_darwin.go), so we only
@@ -735,6 +854,33 @@ func (a *App) SaveNativeRegionToRemote(result CaptureResult) error {
 	return a.runSaveRemotePipeline(cropped)
 }
 
+// SummarizeNativeRegion is the macOS-native overlay equivalent of
+// SummarizeRegion. The AppKit panel is already closed before this runs.
+func (a *App) SummarizeNativeRegion(result CaptureResult) error {
+	pc, err := a.consumePendingCapture()
+	if err != nil {
+		slog.Warn("SummarizeNativeRegion: no pending capture", "err", err)
+		return err
+	}
+	defer func() {
+		a.capturing.Store(false)
+		hideDockIcon()
+	}()
+
+	flushFrame()
+	slog.Info("SummarizeNativeRegion: capturing region",
+		"x", result.Rect.X, "y", result.Rect.Y,
+		"w", result.Rect.W, "h", result.Rect.H,
+		"annotations", len(result.Annotations))
+	cropped, err := a.captureSelectedPNG(result, pc)
+	if err != nil {
+		slog.Error("SummarizeNativeRegion: capture failed", "err", err)
+		wruntime.EventsEmit(a.ctx, "upload:failure", err.Error())
+		return err
+	}
+	return a.runSummaryPipeline(cropped)
+}
+
 func parseNativeAnnotations(raw string) []application.Annotation {
 	if raw == "" {
 		return nil
@@ -784,6 +930,27 @@ func (a *App) QuitApp() {
 	os.Exit(0)
 }
 
+func (a *App) emitOperationStatus(operation, phase, message, state string) {
+	if a.ctx != nil {
+		wruntime.EventsEmit(a.ctx, "operation:status", OperationStatusPayload{
+			Operation: operation,
+			Phase:     phase,
+			Message:   message,
+			State:     state,
+		})
+	}
+	switch state {
+	case "success":
+		showOperationStatus(phase, message, operationStatusSuccess)
+		hideOperationStatusAfter(1400 * time.Millisecond)
+	case "error":
+		showOperationStatus(phase, message, operationStatusError)
+		hideOperationStatusAfter(2600 * time.Millisecond)
+	default:
+		showOperationStatus(phase, message, operationStatusRunning)
+	}
+}
+
 // runtimeNotifier emits success / failure events via the Wails runtime.
 type runtimeNotifier struct{ ctx context.Context }
 
diff --git a/frontend/src/App.vue b/frontend/src/App.vue
index a1d3e4b..7794ea8 100644
--- a/frontend/src/App.vue
+++ b/frontend/src/App.vue
@@ -3,9 +3,7 @@
  * App shell — switches between two distinct UI modes that share the same
  * Wails window:
  *
- *   • "settings" : full settings UI (self-drawn title bar + sidebar + form).
- *                  Self-drawn because the window is now Frameless to make
- *                  the overlay paint edge-to-edge.
+ *   • "settings" : full settings UI (native title bar + sidebar + form).
  *   • "overlay"  : Snipaste-style region picker that fills the whole
  *                  primary display.
  *
@@ -24,6 +22,7 @@ import {
   CopyRegionImage,
   SaveRegionImage,
   SaveRegionToRemote,
+  SummarizeRegion,
   CancelRegion,
 } from '../wailsjs/go/main/App'
 
@@ -39,7 +38,7 @@ const mode = ref<Mode>('settings')
 // `SettingsTab` is hoisted to the App shell so the sidebar (which lives
 // here) and the inner SettingsView (which renders the matching card) can
 // share a single source of truth without an event bus.
-type SettingsTab = 'general' | 's3' | 'ssh'
+type SettingsTab = 'general' | 's3' | 'ssh' | 'llm'
 const activeTab = ref<SettingsTab>('general')
 
 // Sidebar entries are declarative so adding a destination type later is
@@ -48,6 +47,7 @@ const sidebarItems: Array<{ id: SettingsTab; label: string }> = [
   { id: 'general', label: 'General' },
   { id: 's3', label: 'S3' },
   { id: 'ssh', label: 'SSH' },
+  { id: 'llm', label: 'LLM' },
 ]
 
 interface OverlayPayload {
@@ -62,6 +62,14 @@ const capturing = ref(false)
 const toast = ref<{ kind: 'success' | 'error'; text: string } | null>(null)
 let toastTimer: number | undefined
 
+const operationStatus = ref<{
+  operation: string
+  phase: string
+  message: string
+  state: 'running' | 'success' | 'error'
+} | null>(null)
+let operationTimer: number | undefined
+
 function showToast(kind: 'success' | 'error', text: string) {
   toast.value = { kind, text }
   window.clearTimeout(toastTimer)
@@ -70,6 +78,21 @@ function showToast(kind: 'success' | 'error', text: string) {
   }, 3000)
 }
 
+function updateOperationStatus(payload: {
+  operation: string
+  phase: string
+  message: string
+  state: 'running' | 'success' | 'error'
+}) {
+  operationStatus.value = payload
+  window.clearTimeout(operationTimer)
+  if (payload.state !== 'running') {
+    operationTimer = window.setTimeout(() => {
+      operationStatus.value = null
+    }, payload.state === 'success' ? 1600 : 3200)
+  }
+}
+
 async function retryHotkey() {
   try {
     await RetryRegisterHotkey()
@@ -168,6 +191,28 @@ async function onOverlaySaveRemote(rect: {
   }
 }
 
+async function onOverlaySummarize(rect: {
+  rect: {
+    x: number
+    y: number
+    w: number
+    h: number
+  }
+  annotations: Array<{
+    tool: string
+    color: string
+    points: Array<{ x: number; y: number }>
+  }>
+}) {
+  mode.value = 'settings'
+  overlayPayload.value = null
+  try {
+    await SummarizeRegion(rect as any)
+  } catch {
+    /* Surfaced via upload:failure */
+  }
+}
+
 async function onOverlayCancel() {
   mode.value = 'settings'
   overlayPayload.value = null
@@ -194,7 +239,12 @@ onMounted(() => {
     capturing.value = false
   })
   EventsOn('upload:success', (url: string) => {
-    showToast('success', `Copied: ${url}`)
+    showToast(
+      'success',
+      url === 'summary copied to clipboard'
+        ? 'Summary copied to clipboard'
+        : `Copied: ${url}`
+    )
   })
   EventsOn('upload:failure', (reason: string) => {
     showToast('error', reason)
@@ -205,6 +255,7 @@ onMounted(() => {
   EventsOn('hotkey:error', (reason: string) => {
     hotkeyStatus.value = { state: 'error', reason }
   })
+  EventsOn('operation:status', updateOperationStatus)
 })
 
 onUnmounted(() => {
@@ -216,6 +267,7 @@ onUnmounted(() => {
   EventsOff('upload:failure')
   EventsOff('hotkey:ready')
   EventsOff('hotkey:error')
+  EventsOff('operation:status')
 })
 </script>
 
@@ -229,6 +281,7 @@ onUnmounted(() => {
     @copy="onOverlayCopy"
     @save="onOverlaySave"
     @save-remote="onOverlaySaveRemote"
+    @summarize="onOverlaySummarize"
     @cancel="onOverlayCancel"
   />
 
@@ -264,6 +317,20 @@ onUnmounted(() => {
     </main>
 
     <Toast v-if="toast" :kind="toast.kind" :text="toast.text" />
+    <div
+      v-if="operationStatus"
+      class="operation-hud"
+      :class="operationStatus.state"
+    >
+      <span v-if="operationStatus.state === 'running'" class="spinner" />
+      <span v-else class="status-mark">
+        {{ operationStatus.state === 'success' ? 'OK' : '!' }}
+      </span>
+      <div>
+        <strong>{{ operationStatus.phase }}</strong>
+        <p>{{ operationStatus.message }}</p>
+      </div>
+    </div>
   </div>
 </template>
 
@@ -371,6 +438,66 @@ onUnmounted(() => {
   min-width: 0;
 }
 
+.operation-hud {
+  position: fixed;
+  right: 18px;
+  bottom: 18px;
+  z-index: 20;
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  min-width: 260px;
+  max-width: 360px;
+  padding: 12px 14px;
+  color: #f9fafb;
+  background: rgba(17, 24, 39, 0.94);
+  border-radius: 8px;
+  box-shadow: 0 12px 34px rgba(15, 23, 42, 0.28);
+}
+.operation-hud strong {
+  display: block;
+  margin-bottom: 2px;
+  font-size: 13px;
+}
+.operation-hud p {
+  margin: 0;
+  color: #d1d5db;
+  font-size: 12px;
+  line-height: 1.35;
+}
+.spinner {
+  width: 20px;
+  height: 20px;
+  border: 2px solid rgba(255, 255, 255, 0.25);
+  border-top-color: #60a5fa;
+  border-radius: 50%;
+  animation: spin 0.75s linear infinite;
+  flex: 0 0 auto;
+}
+.status-mark {
+  display: grid;
+  place-items: center;
+  width: 24px;
+  height: 24px;
+  border-radius: 50%;
+  font-size: 11px;
+  font-weight: 700;
+  flex: 0 0 auto;
+}
+.operation-hud.success .status-mark {
+  color: #052e16;
+  background: #86efac;
+}
+.operation-hud.error .status-mark {
+  color: #450a0a;
+  background: #fca5a5;
+}
+@keyframes spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+
 @media (prefers-color-scheme: dark) {
   .app-root {
     background: #1c1d22;
diff --git a/frontend/src/views/CaptureOverlay.vue b/frontend/src/views/CaptureOverlay.vue
index 3fdf4e0..468e0ad 100644
--- a/frontend/src/views/CaptureOverlay.vue
+++ b/frontend/src/views/CaptureOverlay.vue
@@ -51,6 +51,10 @@ const emit = defineEmits<{
     e: 'save-remote',
     payload: { rect: Rect; annotations: Annotation[] }
   ): void
+  (
+    e: 'summarize',
+    payload: { rect: Rect; annotations: Annotation[] }
+  ): void
   (e: 'cancel'): void
 }>()
 
@@ -112,7 +116,7 @@ const sizeLabel = computed(() => {
 })
 
 const MARK_TOOLBAR_W = 190
-const ACTION_TOOLBAR_W = 180
+const ACTION_TOOLBAR_W = 216
 const TOOLBAR_GROUP_GAP = 8
 
 const rightToolbarPos = computed(() => {
@@ -363,7 +367,11 @@ function onSaveRemote() {
   emitAction('save-remote')
 }
 
-function emitAction(action: 'confirm' | 'copy' | 'save' | 'save-remote') {
+function onSummarize() {
+  emitAction('summarize')
+}
+
+function emitAction(action: 'confirm' | 'copy' | 'save' | 'save-remote' | 'summarize') {
   if (!rect.value) return
   const payload = {
     rect: { ...rect.value },
@@ -373,6 +381,7 @@ function emitAction(action: 'confirm' | 'copy' | 'save' | 'save-remote') {
   if (action === 'copy') emit('copy', payload)
   if (action === 'save') emit('save', payload)
   if (action === 'save-remote') emit('save-remote', payload)
+  if (action === 'summarize') emit('summarize', payload)
 }
 
 function onCancel() {
@@ -593,8 +602,8 @@ onUnmounted(() => {
       />
       <button
         class="action-btn"
-        data-tip="复制图标"
-        aria-label="复制图标"
+        data-tip="复制图片"
+        aria-label="复制图片"
         @click="onCopy"
         v-html="copyIcon"
       />
@@ -612,6 +621,19 @@ onUnmounted(() => {
         @click="onSaveRemote"
         v-html="saveRemoteIcon"
       />
+      <button
+        class="action-btn"
+        data-tip="复制总结"
+        aria-label="复制总结"
+        @click="onSummarize"
+      >
+        <svg viewBox="0 0 24 24" aria-hidden="true">
+          <path d="M6 3h8l4 4v14H6z" />
+          <path d="M14 3v5h4" />
+          <path d="M9 12h6" />
+          <path d="M9 16h5" />
+        </svg>
+      </button>
       <button
         class="action-btn primary"
         data-tip="上传云端"
@@ -706,7 +728,7 @@ onUnmounted(() => {
   width: 190px;
 }
 .action-toolbar {
-  width: 180px;
+  width: 216px;
 }
 
 .icon-btn,
@@ -762,6 +784,16 @@ onUnmounted(() => {
   fill: currentColor;
   pointer-events: none;
 }
+.action-btn > svg {
+  width: 18px;
+  height: 18px;
+  fill: none;
+  stroke: currentColor;
+  stroke-width: 2;
+  stroke-linecap: round;
+  stroke-linejoin: round;
+  pointer-events: none;
+}
 .action-btn::after {
   content: attr(data-tip);
   position: absolute;
diff --git a/frontend/src/views/SettingsView.vue b/frontend/src/views/SettingsView.vue
index 4bcd2ca..733f896 100644
--- a/frontend/src/views/SettingsView.vue
+++ b/frontend/src/views/SettingsView.vue
@@ -8,6 +8,7 @@
  *   • "general" — the global hotkey / capture parameters.
  *   • "s3"      — S3-compatible object-storage credentials.
  *   • "ssh"     — SSH/SCP destination for the "save remote" button.
+ *   • "llm"     — multimodal screenshot summary provider settings.
  *
  * State flow: load() pulls config from Go on mount, save() pushes back.
  */
@@ -22,7 +23,7 @@ import {
 
 // Tab discriminator shared with the App shell. Kept as a string union so
 // the parent can pass the value without importing a type from this view.
-type TabId = 'general' | 's3' | 'ssh'
+type TabId = 'general' | 's3' | 'ssh' | 'llm'
 
 const props = defineProps<{
   tab: TabId
@@ -51,12 +52,44 @@ interface SSHConfig {
   connectTimeoutSecs: number
 }
 
+interface LLMProviderConfig {
+  label: string
+  baseUrl: string
+  apiKey: string
+  model: string
+  maxTokens: number
+  temperature: number
+  timeoutSecs: number
+}
+
+interface LLMConfig {
+  activeProvider: string
+  prompt: string
+  providers: Record<string, LLMProviderConfig>
+}
+
 interface AppConfig {
   hotkey: string
   s3: S3Config
   ssh: SSHConfig
+  llm: LLMConfig
 }
 
+const llmProviderOptions = [
+  { id: 'qwen', label: '阿里千问多模态' },
+  { id: 'doubao', label: '火山方舟豆包多模态' },
+  { id: 'openai', label: 'ChatGPT / OpenAI-compatible' },
+]
+
+const DEFAULT_SUMMARY_PROMPT = `你是一个截图内容总结助手。请阅读截图，并用中文输出一段适合直接粘贴到聊天、工单、Issue 或 PR 中的说明。
+
+要求：
+- 先说明截图中最重要的信息和用户可能想表达的意图。
+- 如果截图包含报错、异常状态、表格、代码、界面控件或关键数字，请准确提取。
+- 如果截图内容不足以判断，不要编造；直接说明不确定点。
+- 输出尽量简洁，默认 3 到 6 条要点。
+- 不要输出 Markdown 标题，不要寒暄。`
+
 // `defaultConfig` keeps the initial render in sync with the Go-side
 // DefaultAppConfig so the UI never flashes empty fields before load()
 // completes. Centralising the defaults also avoids subtle drift between
@@ -85,6 +118,39 @@ function defaultConfig(): AppConfig {
       knownHostsPath: '',
       connectTimeoutSecs: 10,
     },
+    llm: {
+      activeProvider: 'openai',
+      prompt: DEFAULT_SUMMARY_PROMPT,
+      providers: {
+        qwen: {
+          label: '阿里千问多模态',
+          baseUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
+          apiKey: '',
+          model: 'qwen-vl-plus',
+          maxTokens: 600,
+          temperature: 0.2,
+          timeoutSecs: 60,
+        },
+        doubao: {
+          label: '火山方舟豆包多模态',
+          baseUrl: 'https://ark.cn-beijing.volces.com/api/v3',
+          apiKey: '',
+          model: '',
+          maxTokens: 600,
+          temperature: 0.2,
+          timeoutSecs: 60,
+        },
+        openai: {
+          label: 'ChatGPT / OpenAI-compatible',
+          baseUrl: 'https://api.openai.com/v1',
+          apiKey: '',
+          model: 'gpt-5.5',
+          maxTokens: 600,
+          temperature: 0.2,
+          timeoutSecs: 60,
+        },
+      },
+    },
   }
 }
 
@@ -109,6 +175,15 @@ const sshPathDisplay = computed({
   },
 })
 
+const activeLLMProvider = computed(() => {
+  const id = config.value.llm.activeProvider || 'openai'
+  if (!config.value.llm.providers[id]) {
+    config.value.llm.activeProvider = 'openai'
+    return config.value.llm.providers.openai
+  }
+  return config.value.llm.providers[id]
+})
+
 function flash(kind: 'ok' | 'err', text: string) {
   message.value = { kind, text }
   window.setTimeout(() => {
@@ -125,6 +200,11 @@ async function load() {
     Object.assign(merged, cfg)
     merged.s3 = { ...merged.s3, ...(cfg as any).s3 }
     merged.ssh = { ...merged.ssh, ...(cfg as any).ssh }
+    merged.llm = { ...merged.llm, ...(cfg as any).llm }
+    merged.llm.providers = {
+      ...defaultConfig().llm.providers,
+      ...((cfg as any).llm?.providers ?? {}),
+    }
     config.value = merged
   }
 }
@@ -368,6 +448,103 @@ onMounted(load)
       </div>
     </section>
 
+    <!-- LLM tab — multimodal screenshot summary provider settings. -->
+    <section v-if="props.tab === 'llm'" class="card">
+      <h2>LLM screenshot summary</h2>
+      <div class="grid">
+        <label class="field full">
+          <span>Provider</span>
+          <select v-model="config.llm.activeProvider">
+            <option
+              v-for="provider in llmProviderOptions"
+              :key="provider.id"
+              :value="provider.id"
+            >
+              {{ provider.label }}
+            </option>
+          </select>
+        </label>
+        <label class="field full">
+          <span>Base URL *</span>
+          <input
+            v-model="activeLLMProvider.baseUrl"
+            placeholder="https://api.openai.com/v1"
+            spellcheck="false"
+          />
+        </label>
+        <label class="field">
+          <span>API Key *</span>
+          <input v-model="activeLLMProvider.apiKey" type="password" />
+        </label>
+        <label class="field">
+          <span>Model *</span>
+          <input
+            v-model="activeLLMProvider.model"
+            placeholder="gpt-5.5 / qwen-vl-plus / Ark endpoint ID"
+            spellcheck="false"
+          />
+        </label>
+        <label class="field">
+          <span>Max tokens</span>
+          <input
+            v-model.number="activeLLMProvider.maxTokens"
+            type="number"
+            min="64"
+            max="4096"
+            placeholder="600"
+          />
+        </label>
+        <label class="field">
+          <span>Temperature</span>
+          <input
+            v-model.number="activeLLMProvider.temperature"
+            type="number"
+            min="0"
+            max="2"
+            step="0.1"
+            placeholder="0.2"
+          />
+        </label>
+        <label class="field">
+          <span>Timeout (sec)</span>
+          <input
+            v-model.number="activeLLMProvider.timeoutSecs"
+            type="number"
+            min="10"
+            max="300"
+            placeholder="60"
+          />
+        </label>
+        <label class="field full">
+          <span>Prompt</span>
+          <textarea
+            v-model="config.llm.prompt"
+            class="prompt-textarea"
+            spellcheck="false"
+          />
+        </label>
+      </div>
+      <p class="hint">
+        The "copy summary" screenshot action first uploads the image through the
+        S3 configuration, then sends the public image URL to this multimodal
+        provider. Keep S3 configured before using it.
+      </p>
+      <p v-if="config.llm.activeProvider === 'doubao'" class="hint">
+        Volcengine Ark usually uses the inference endpoint ID as the model
+        value; paste the endpoint/model identifier from your Ark console.
+      </p>
+      <p v-else-if="config.llm.activeProvider === 'qwen'" class="hint">
+        DashScope OpenAI-compatible mode is used here. If your workspace uses a
+        regional endpoint, replace Base URL with that compatible-mode URL.
+      </p>
+
+      <div class="actions">
+        <button class="btn primary" :disabled="saving" @click="save">
+          {{ saving ? 'Saving…' : 'Save' }}
+        </button>
+      </div>
+    </section>
+
     <transition name="fade">
       <div
         v-if="message"
@@ -460,13 +637,13 @@ kbd {
 .field input[type='text'],
 .field input[type='number'],
 .field input:not([type='checkbox']),
-.field select {
+.field select,
+.field textarea {
   border: 1px solid #d1d5db;
   border-radius: 6px;
   padding: 7px 10px;
   font-size: 13px;
   line-height: 1.4;
-  height: 36px;
   background: #fff;
   color: inherit;
   outline: none;
@@ -474,8 +651,20 @@ kbd {
   width: 100%;
   box-sizing: border-box;
 }
+.field input[type='text'],
+.field input[type='number'],
+.field input:not([type='checkbox']),
+.field select {
+  height: 36px;
+}
+.field textarea {
+  min-height: 168px;
+  resize: vertical;
+  font-family: inherit;
+}
 .field input:focus,
-.field select:focus {
+.field select:focus,
+.field textarea:focus {
   border-color: #3b82f6;
   box-shadow: 0 0 0 3px rgba(59, 130, 246, 0.18);
 }
@@ -598,15 +787,20 @@ kbd {
     background: #1f2937;
     border-color: #374151;
   }
-  .field input {
+  .field input,
+  .field textarea {
     background: #111827;
     border-color: #374151;
-    color: #e5e7eb;
+    color: #fff;
   }
   .field select {
     background: #111827;
     border-color: #374151;
-    color: #e5e7eb;
+    color: #fff;
+  }
+  .field input::placeholder,
+  .field textarea::placeholder {
+    color: #9ca3af;
   }
   .field input:disabled {
     background: #1f2937;
diff --git a/frontend/wailsjs/go/main/App.d.ts b/frontend/wailsjs/go/main/App.d.ts
index d6c5fbe..2afcccc 100755
--- a/frontend/wailsjs/go/main/App.d.ts
+++ b/frontend/wailsjs/go/main/App.d.ts
@@ -37,6 +37,10 @@ export function SaveRegionToRemote(arg1:main.CaptureResult):Promise<void>;
 
 export function ShowWindow():Promise<void>;
 
+export function SummarizeNativeRegion(arg1:main.CaptureResult):Promise<void>;
+
+export function SummarizeRegion(arg1:main.CaptureResult):Promise<void>;
+
 export function TestConnection(arg1:domain.S3Config):Promise<void>;
 
 export function TestSSHConnection(arg1:domain.SSHConfig):Promise<void>;
diff --git a/frontend/wailsjs/go/main/App.js b/frontend/wailsjs/go/main/App.js
index 14fecd8..d51dce7 100755
--- a/frontend/wailsjs/go/main/App.js
+++ b/frontend/wailsjs/go/main/App.js
@@ -70,6 +70,14 @@ export function ShowWindow() {
   return window['go']['main']['App']['ShowWindow']();
 }
 
+export function SummarizeNativeRegion(arg1) {
+  return window['go']['main']['App']['SummarizeNativeRegion'](arg1);
+}
+
+export function SummarizeRegion(arg1) {
+  return window['go']['main']['App']['SummarizeRegion'](arg1);
+}
+
 export function TestConnection(arg1) {
   return window['go']['main']['App']['TestConnection'](arg1);
 }
diff --git a/frontend/wailsjs/go/models.ts b/frontend/wailsjs/go/models.ts
index dbebcd8..7aa1a83 100755
--- a/frontend/wailsjs/go/models.ts
+++ b/frontend/wailsjs/go/models.ts
@@ -53,6 +53,64 @@ export namespace application {
 
 export namespace domain {
 	
+	export class LLMProviderConfig {
+	    label: string;
+	    baseUrl: string;
+	    apiKey: string;
+	    model: string;
+	    maxTokens: number;
+	    temperature: number;
+	    timeoutSecs: number;
+	
+	    static createFrom(source: any = {}) {
+	        return new LLMProviderConfig(source);
+	    }
+	
+	    constructor(source: any = {}) {
+	        if ('string' === typeof source) source = JSON.parse(source);
+	        this.label = source["label"];
+	        this.baseUrl = source["baseUrl"];
+	        this.apiKey = source["apiKey"];
+	        this.model = source["model"];
+	        this.maxTokens = source["maxTokens"];
+	        this.temperature = source["temperature"];
+	        this.timeoutSecs = source["timeoutSecs"];
+	    }
+	}
+	export class LLMConfig {
+	    activeProvider: string;
+	    prompt: string;
+	    providers: Record<string, LLMProviderConfig>;
+	
+	    static createFrom(source: any = {}) {
+	        return new LLMConfig(source);
+	    }
+	
+	    constructor(source: any = {}) {
+	        if ('string' === typeof source) source = JSON.parse(source);
+	        this.activeProvider = source["activeProvider"];
+	        this.prompt = source["prompt"];
+	        this.providers = this.convertValues(source["providers"], LLMProviderConfig, true);
+	    }
+	
+		convertValues(a: any, classs: any, asMap: boolean = false): any {
+		    if (!a) {
+		        return a;
+		    }
+		    if (a.slice && a.map) {
+		        return (a as any[]).map(elem => this.convertValues(elem, classs));
+		    } else if ("object" === typeof a) {
+		        if (asMap) {
+		            for (const key of Object.keys(a)) {
+		                a[key] = new classs(a[key]);
+		            }
+		            return a;
+		        }
+		        return new classs(a);
+		    }
+		    return a;
+		}
+	}
 	export class SSHConfig {
 	    host: string;
 	    port: number;
@@ -111,6 +169,7 @@ export namespace domain {
 	    hotkey: string;
 	    s3: S3Config;
 	    ssh: SSHConfig;
+	    llm: LLMConfig;
 	
 	    static createFrom(source: any = {}) {
 	        return new AppConfig(source);
@@ -121,6 +180,7 @@ export namespace domain {
 	        this.hotkey = source["hotkey"];
 	        this.s3 = this.convertValues(source["s3"], S3Config);
 	        this.ssh = this.convertValues(source["ssh"], SSHConfig);
+	        this.llm = this.convertValues(source["llm"], LLMConfig);
 	    }
 	
 		convertValues(a: any, classs: any, asMap: boolean = false): any {
@@ -142,6 +202,8 @@ export namespace domain {
 		}
 	}
 	
+	
+	
 
 }
 
diff --git a/internal/application/capture_summary.go b/internal/application/capture_summary.go
new file mode 100644
index 0000000..e360557
--- /dev/null
+++ b/internal/application/capture_summary.go
@@ -0,0 +1,72 @@
+package application
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/mmmy/snapgo/internal/domain"
+	"github.com/mmmy/snapgo/internal/infrastructure/clipboard"
+)
+
+// VisionSummarizer describes a multimodal model capable of reading a public
+// image URL and returning a textual summary.
+type VisionSummarizer interface {
+	SummarizeImage(ctx context.Context, prompt string, imageURL string) (string, error)
+}
+
+// CaptureSummaryService wires screenshot bytes -> S3 public URL -> LLM
+// summary -> clipboard. It intentionally does not notify UI state itself:
+// the caller emits fine-grained "uploading / recognizing / done" progress.
+type CaptureSummaryService struct {
+	Provider   domain.OSSProvider
+	Summarizer VisionSummarizer
+	Clipboard  clipboard.Writer
+	Prompt     string
+	PathPrefix string
+}
+
+// UploadImage uploads the screenshot and returns the public URL visible to
+// the multimodal provider.
+func (s *CaptureSummaryService) UploadImage(ctx context.Context, pngBytes []byte) (*domain.UploadResult, error) {
+	if s.Provider == nil {
+		return nil, fmt.Errorf("s3 is not configured")
+	}
+	if len(pngBytes) == 0 {
+		return nil, fmt.Errorf("empty screenshot")
+	}
+	key := buildObjectKey(s.PathPrefix)
+	start := time.Now()
+	url, err := s.Provider.Upload(ctx, key, pngBytes, "image/png")
+	if err != nil {
+		return nil, fmt.Errorf("s3 put object: %w", err)
+	}
+	return &domain.UploadResult{
+		URL:      url,
+		Key:      key,
+		Provider: s.Provider.Name(),
+		Elapsed:  time.Since(start),
+	}, nil
+}
+
+// Summarize asks the configured model to summarize the uploaded image URL.
+func (s *CaptureSummaryService) Summarize(ctx context.Context, imageURL string) (string, error) {
+	if s.Summarizer == nil {
+		return "", fmt.Errorf("llm is not configured")
+	}
+	return s.Summarizer.SummarizeImage(ctx, s.Prompt, imageURL)
+}
+
+// CopySummary writes the final summary text to the clipboard.
+func (s *CaptureSummaryService) CopySummary(_ context.Context, summary string) error {
+	if summary == "" {
+		return fmt.Errorf("empty summary")
+	}
+	if s.Clipboard == nil {
+		return fmt.Errorf("clipboard is not configured")
+	}
+	if err := s.Clipboard.WriteText(summary); err != nil {
+		return fmt.Errorf("clipboard write failed: %w", err)
+	}
+	return nil
+}
diff --git a/internal/application/capture_summary_test.go b/internal/application/capture_summary_test.go
new file mode 100644
index 0000000..a808bd3
--- /dev/null
+++ b/internal/application/capture_summary_test.go
@@ -0,0 +1,86 @@
+package application
+
+import (
+	"context"
+	"testing"
+)
+
+type fakeOSSProvider struct {
+	key         string
+	data        []byte
+	contentType string
+	url         string
+}
+
+func (f *fakeOSSProvider) Upload(_ context.Context, key string, data []byte, contentType string) (string, error) {
+	f.key = key
+	f.data = append([]byte(nil), data...)
+	f.contentType = contentType
+	if f.url == "" {
+		return "https://cdn.example.com/" + key, nil
+	}
+	return f.url, nil
+}
+
+func (f *fakeOSSProvider) Name() string { return "fake" }
+
+type fakeSummarizer struct {
+	prompt string
+	url    string
+	text   string
+}
+
+func (f *fakeSummarizer) SummarizeImage(_ context.Context, prompt string, imageURL string) (string, error) {
+	f.prompt = prompt
+	f.url = imageURL
+	if f.text == "" {
+		return "summary text", nil
+	}
+	return f.text, nil
+}
+
+func TestCaptureSummaryServiceUploadsSummarizesAndCopies(t *testing.T) {
+	ctx := context.Background()
+	provider := &fakeOSSProvider{}
+	summarizer := &fakeSummarizer{}
+	clip := &fakeClipboard{}
+	svc := &CaptureSummaryService{
+		Provider:   provider,
+		Summarizer: summarizer,
+		Clipboard:  clip,
+		Prompt:     "describe screenshot",
+		PathPrefix: "snapgo/",
+	}
+
+	uploaded, err := svc.UploadImage(ctx, []byte("png"))
+	if err != nil {
+		t.Fatalf("upload image: %v", err)
+	}
+	if provider.contentType != "image/png" {
+		t.Fatalf("expected image/png content type, got %q", provider.contentType)
+	}
+	if string(provider.data) != "png" {
+		t.Fatalf("expected uploaded png bytes, got %q", string(provider.data))
+	}
+
+	summary, err := svc.Summarize(ctx, uploaded.URL)
+	if err != nil {
+		t.Fatalf("summarize: %v", err)
+	}
+	if summarizer.prompt != "describe screenshot" {
+		t.Fatalf("expected prompt forwarded, got %q", summarizer.prompt)
+	}
+	if summarizer.url != uploaded.URL {
+		t.Fatalf("expected uploaded url forwarded, got %q", summarizer.url)
+	}
+	if summary != "summary text" {
+		t.Fatalf("expected summary text, got %q", summary)
+	}
+
+	if err := svc.CopySummary(ctx, summary); err != nil {
+		t.Fatalf("copy summary: %v", err)
+	}
+	if clip.text != "summary text" {
+		t.Fatalf("expected summary copied, got %q", clip.text)
+	}
+}
diff --git a/internal/domain/config.go b/internal/domain/config.go
index c6ed8ae..5d3a217 100644
--- a/internal/domain/config.go
+++ b/internal/domain/config.go
@@ -10,12 +10,12 @@ package domain
 // endpoint, etc.).
 //
 // Field design notes:
-// - PathPrefix supports object name templating so users can group screenshots
-//   by date.
-// - PublicURLBase is optional to support CDN-fronted buckets; otherwise the
-//   adapter falls back to "{Endpoint}/{Bucket}/{Key}" (path-style).
-// - UsePathStyle defaults to true because many self-hosted MinIO / R2
-//   deployments do not support virtual-hosted-style addressing.
+//   - PathPrefix supports object name templating so users can group screenshots
+//     by date.
+//   - PublicURLBase is optional to support CDN-fronted buckets; otherwise the
+//     adapter falls back to "{Endpoint}/{Bucket}/{Key}" (path-style).
+//   - UsePathStyle defaults to true because many self-hosted MinIO / R2
+//     deployments do not support virtual-hosted-style addressing.
 type S3Config struct {
 	Endpoint        string `json:"endpoint"`
 	Region          string `json:"region"`
@@ -34,17 +34,17 @@ type S3Config struct {
 //   - Host / Port form the destination endpoint. Port defaults to 22 when 0.
 //   - User is the SSH login name.
 //   - AuthMethod selects how the connection authenticates:
-//       "password"     → password only, via the in-process Go SSH client.
-//       "key"          → ssh-agent + ~/.ssh/id_* key files (no password),
-//                         via the in-process Go SSH client.
-//       "kerberos"     → delegate to the system /usr/bin/ssh binary so the
-//                         existing Kerberos (GSSAPI) credential cache from
-//                         `kinit` is reused. Native GSSAPI is required here
-//                         because macOS stores tickets in an API: ccache
-//                         that pure-Go SSH libraries cannot read.
-//       "" / "builtin" → legacy combined behaviour (password → agent → key
-//                         files) kept only for backward compatibility with
-//                         configs written before the method was split.
+//     "password"     → password only, via the in-process Go SSH client.
+//     "key"          → ssh-agent + ~/.ssh/id_* key files (no password),
+//     via the in-process Go SSH client.
+//     "kerberos"     → delegate to the system /usr/bin/ssh binary so the
+//     existing Kerberos (GSSAPI) credential cache from
+//     `kinit` is reused. Native GSSAPI is required here
+//     because macOS stores tickets in an API: ccache
+//     that pure-Go SSH libraries cannot read.
+//     "" / "builtin" → legacy combined behaviour (password → agent → key
+//     files) kept only for backward compatibility with
+//     configs written before the method was split.
 //   - Password is optional; when empty the implementation falls back to the
 //     local SSH agent or ~/.ssh/id_* keys (i.e. the user is responsible for
 //     having password-less access already configured on this machine).
@@ -72,6 +72,29 @@ type SSHConfig struct {
 	ConnectTimeoutSecs int    `json:"connectTimeoutSecs"`
 }
 
+// LLMProviderConfig stores one OpenAI-compatible multimodal chat endpoint.
+//
+// The three built-in providers (Qwen, Doubao/Ark, OpenAI-compatible) all use
+// the same request shape: chat/completions with a user content array that
+// contains text plus an image_url item. Keeping them as provider records lets
+// users switch between credentials/models without retyping every field.
+type LLMProviderConfig struct {
+	Label       string  `json:"label"`
+	BaseURL     string  `json:"baseUrl"`
+	APIKey      string  `json:"apiKey"`
+	Model       string  `json:"model"`
+	MaxTokens   int     `json:"maxTokens"`
+	Temperature float64 `json:"temperature"`
+	TimeoutSecs int     `json:"timeoutSecs"`
+}
+
+// LLMConfig controls screenshot summarisation.
+type LLMConfig struct {
+	ActiveProvider string                       `json:"activeProvider"`
+	Prompt         string                       `json:"prompt"`
+	Providers      map[string]LLMProviderConfig `json:"providers"`
+}
+
 // SSH authentication method identifiers stored in SSHConfig.AuthMethod.
 const (
 	// SSHAuthBuiltin is the legacy combined method (password → agent → key
@@ -87,6 +110,22 @@ const (
 	SSHAuthKerberos = "kerberos"
 )
 
+// Built-in LLM provider identifiers.
+const (
+	LLMProviderQwen   = "qwen"
+	LLMProviderDoubao = "doubao"
+	LLMProviderOpenAI = "openai"
+)
+
+const DefaultSummaryPrompt = `你是一个截图内容总结助手。请阅读截图，并用中文输出一段适合直接粘贴到聊天、工单、Issue 或 PR 中的说明。
+
+要求：
+- 先说明截图中最重要的信息和用户可能想表达的意图。
+- 如果截图包含报错、异常状态、表格、代码、界面控件或关键数字，请准确提取。
+- 如果截图内容不足以判断，不要编造；直接说明不确定点。
+- 输出尽量简洁，默认 3 到 6 条要点。
+- 不要输出 Markdown 标题，不要寒暄。`
+
 // IsKerberos reports whether this config requests Kerberos/GSSAPI auth.
 func (c SSHConfig) IsKerberos() bool {
 	return c.AuthMethod == SSHAuthKerberos
@@ -108,11 +147,14 @@ type AppConfig struct {
 	// SSH holds the configuration for the optional "save to remote via scp"
 	// destination triggered by the save-remote toolbar button.
 	SSH SSHConfig `json:"ssh"`
+
+	// LLM holds provider settings for the "copy summary" screenshot action.
+	LLM LLMConfig `json:"llm"`
 }
 
 // DefaultAppConfig returns sane zero-value defaults used on first launch.
 func DefaultAppConfig() AppConfig {
-	return AppConfig{
+	cfg := AppConfig{
 		Hotkey: "cmd+shift+a",
 		S3: S3Config{
 			PathPrefix:   "snapgo/",
@@ -124,6 +166,102 @@ func DefaultAppConfig() AppConfig {
 			ConnectTimeoutSecs: 10,
 			StrictHostKey:      false,
 		},
+		LLM: DefaultLLMConfig(),
+	}
+	cfg.Normalize()
+	return cfg
+}
+
+// DefaultLLMConfig returns the built-in provider presets. API keys and exact
+// models remain user-editable because each provider account may expose
+// different model names / endpoint IDs.
+func DefaultLLMConfig() LLMConfig {
+	return LLMConfig{
+		ActiveProvider: LLMProviderOpenAI,
+		Prompt:         DefaultSummaryPrompt,
+		Providers: map[string]LLMProviderConfig{
+			LLMProviderQwen: {
+				Label:       "阿里千问多模态",
+				BaseURL:     "https://dashscope.aliyuncs.com/compatible-mode/v1",
+				Model:       "qwen-vl-plus",
+				MaxTokens:   600,
+				Temperature: 0.2,
+				TimeoutSecs: 60,
+			},
+			LLMProviderDoubao: {
+				Label:       "火山方舟豆包多模态",
+				BaseURL:     "https://ark.cn-beijing.volces.com/api/v3",
+				Model:       "",
+				MaxTokens:   600,
+				Temperature: 0.2,
+				TimeoutSecs: 60,
+			},
+			LLMProviderOpenAI: {
+				Label:       "ChatGPT / OpenAI-compatible",
+				BaseURL:     "https://api.openai.com/v1",
+				Model:       "gpt-5.5",
+				MaxTokens:   600,
+				Temperature: 0.2,
+				TimeoutSecs: 60,
+			},
+		},
+	}
+}
+
+// Normalize fills defaults into configs written by older app versions while
+// preserving any user-provided provider fields.
+func (c *AppConfig) Normalize() {
+	if c.Hotkey == "" {
+		c.Hotkey = "cmd+shift+a"
+	}
+	if c.S3.PathPrefix == "" {
+		c.S3.PathPrefix = "snapgo/"
+	}
+	if c.SSH.Port == 0 {
+		c.SSH.Port = 22
+	}
+	if c.SSH.PathPrefix == "" {
+		c.SSH.PathPrefix = "snapgo/"
+	}
+	if c.SSH.ConnectTimeoutSecs == 0 {
+		c.SSH.ConnectTimeoutSecs = 10
+	}
+
+	defaultLLM := DefaultLLMConfig()
+	if c.LLM.ActiveProvider == "" {
+		c.LLM.ActiveProvider = defaultLLM.ActiveProvider
+	}
+	if c.LLM.Prompt == "" {
+		c.LLM.Prompt = defaultLLM.Prompt
+	}
+	if c.LLM.Providers == nil {
+		c.LLM.Providers = map[string]LLMProviderConfig{}
+	}
+	for id, def := range defaultLLM.Providers {
+		current, ok := c.LLM.Providers[id]
+		if !ok {
+			c.LLM.Providers[id] = def
+			continue
+		}
+		if current.Label == "" {
+			current.Label = def.Label
+		}
+		if current.BaseURL == "" {
+			current.BaseURL = def.BaseURL
+		}
+		if current.Model == "" && id != LLMProviderDoubao {
+			current.Model = def.Model
+		}
+		if current.MaxTokens == 0 {
+			current.MaxTokens = def.MaxTokens
+		}
+		if current.Temperature < 0 {
+			current.Temperature = def.Temperature
+		}
+		if current.TimeoutSecs == 0 {
+			current.TimeoutSecs = def.TimeoutSecs
+		}
+		c.LLM.Providers[id] = current
 	}
 }
 
@@ -141,3 +279,21 @@ func (c AppConfig) IsS3Configured() bool {
 func (c AppConfig) IsSSHConfigured() bool {
 	return c.SSH.Host != "" && c.SSH.User != ""
 }
+
+// ActiveLLMProvider returns the selected provider config plus a boolean
+// indicating whether the selection exists.
+func (c AppConfig) ActiveLLMProvider() (string, LLMProviderConfig, bool) {
+	id := c.LLM.ActiveProvider
+	if id == "" {
+		id = LLMProviderOpenAI
+	}
+	provider, ok := c.LLM.Providers[id]
+	return id, provider, ok
+}
+
+// IsLLMConfigured reports whether the selected multimodal provider has the
+// fields required to make a request.
+func (c AppConfig) IsLLMConfigured() bool {
+	_, provider, ok := c.ActiveLLMProvider()
+	return ok && provider.BaseURL != "" && provider.APIKey != "" && provider.Model != ""
+}
diff --git a/internal/infrastructure/config/file_store.go b/internal/infrastructure/config/file_store.go
index 7994004..285e5f5 100644
--- a/internal/infrastructure/config/file_store.go
+++ b/internal/infrastructure/config/file_store.go
@@ -1,12 +1,12 @@
 // Package config provides JSON-file backed persistence for AppConfig.
 //
 // Design rationale:
-// - JSON keeps the file human-readable, which is useful while the MVP has
-//   no settings UI for every option.
-// - Storage location follows os.UserConfigDir() so each platform places it
-//   in the canonical spot (macOS: ~/Library/Application Support/SnapGo).
-// - File mode 0600 / dir mode 0700 mitigate accidental leakage of access
-//   keys before we wire up the OS keychain (deferred to a later spec).
+//   - JSON keeps the file human-readable, which is useful while the MVP has
+//     no settings UI for every option.
+//   - Storage location follows os.UserConfigDir() so each platform places it
+//     in the canonical spot (macOS: ~/Library/Application Support/SnapGo).
+//   - File mode 0600 / dir mode 0700 mitigate accidental leakage of access
+//     keys before we wire up the OS keychain (deferred to a later spec).
 package config
 
 import (
@@ -65,6 +65,7 @@ func (s *FileStore) Load() (domain.AppConfig, error) {
 	if err := json.Unmarshal(data, &cfg); err != nil {
 		return domain.DefaultAppConfig(), fmt.Errorf("parse config: %w", err)
 	}
+	cfg.Normalize()
 	return cfg, nil
 }
 
@@ -74,6 +75,7 @@ func (s *FileStore) Save(cfg domain.AppConfig) error {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
+	cfg.Normalize()
 	data, err := json.MarshalIndent(cfg, "", "  ")
 	if err != nil {
 		return fmt.Errorf("marshal config: %w", err)
diff --git a/internal/infrastructure/llm/openai_compatible.go b/internal/infrastructure/llm/openai_compatible.go
new file mode 100644
index 0000000..d5bc3d1
--- /dev/null
+++ b/internal/infrastructure/llm/openai_compatible.go
@@ -0,0 +1,174 @@
+// Package llm contains multimodal LLM adapters.
+//
+// The built-in providers are intentionally implemented through the
+// OpenAI-compatible chat/completions shape. Qwen DashScope, Volcengine Ark,
+// and OpenAI-compatible gateways all accept the same "text + image_url"
+// message content pattern, so one small HTTP adapter is enough here.
+package llm
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/mmmy/snapgo/internal/domain"
+)
+
+// VisionClient calls one configured multimodal chat endpoint.
+type VisionClient struct {
+	providerID string
+	cfg        domain.LLMProviderConfig
+	httpClient *http.Client
+}
+
+// NewVisionClient validates cfg and returns a reusable client.
+func NewVisionClient(providerID string, cfg domain.LLMProviderConfig) (*VisionClient, error) {
+	if strings.TrimSpace(cfg.BaseURL) == "" {
+		return nil, fmt.Errorf("llm base url is required")
+	}
+	if strings.TrimSpace(cfg.APIKey) == "" {
+		return nil, fmt.Errorf("llm api key is required")
+	}
+	if strings.TrimSpace(cfg.Model) == "" {
+		return nil, fmt.Errorf("llm model is required")
+	}
+	timeout := cfg.TimeoutSecs
+	if timeout <= 0 {
+		timeout = 60
+	}
+	return &VisionClient{
+		providerID: providerID,
+		cfg:        cfg,
+		httpClient: &http.Client{Timeout: time.Duration(timeout) * time.Second},
+	}, nil
+}
+
+// SummarizeImage asks the model to summarise an image available at imageURL.
+func (c *VisionClient) SummarizeImage(ctx context.Context, prompt string, imageURL string) (string, error) {
+	prompt = strings.TrimSpace(prompt)
+	if prompt == "" {
+		return "", fmt.Errorf("llm prompt is empty")
+	}
+	imageURL = strings.TrimSpace(imageURL)
+	if imageURL == "" {
+		return "", fmt.Errorf("image url is empty")
+	}
+
+	body := map[string]any{
+		"model": c.cfg.Model,
+		"messages": []map[string]any{
+			{
+				"role": "user",
+				"content": []map[string]any{
+					{"type": "text", "text": prompt},
+					{"type": "image_url", "image_url": map[string]any{"url": imageURL}},
+				},
+			},
+		},
+	}
+	if c.cfg.MaxTokens > 0 {
+		if c.providerID == domain.LLMProviderOpenAI {
+			body["max_completion_tokens"] = c.cfg.MaxTokens
+		} else {
+			body["max_tokens"] = c.cfg.MaxTokens
+		}
+	}
+	if c.cfg.Temperature >= 0 {
+		body["temperature"] = c.cfg.Temperature
+	}
+
+	payload, err := json.Marshal(body)
+	if err != nil {
+		return "", fmt.Errorf("marshal llm request: %w", err)
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, chatCompletionsURL(c.cfg.BaseURL), bytes.NewReader(payload))
+	if err != nil {
+		return "", fmt.Errorf("build llm request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+c.cfg.APIKey)
+
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("llm request: %w", err)
+	}
+	defer resp.Body.Close()
+	data, _ := io.ReadAll(io.LimitReader(resp.Body, 4<<20))
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("llm status %d: %s", resp.StatusCode, compactBody(data))
+	}
+
+	var decoded chatCompletionResponse
+	if err := json.Unmarshal(data, &decoded); err != nil {
+		return "", fmt.Errorf("decode llm response: %w", err)
+	}
+	if decoded.Error != nil && decoded.Error.Message != "" {
+		return "", fmt.Errorf("llm error: %s", decoded.Error.Message)
+	}
+	if len(decoded.Choices) == 0 {
+		return "", fmt.Errorf("llm response has no choices")
+	}
+	text := strings.TrimSpace(contentText(decoded.Choices[0].Message.Content))
+	if text == "" {
+		return "", fmt.Errorf("llm response is empty")
+	}
+	return text, nil
+}
+
+func chatCompletionsURL(baseURL string) string {
+	base := strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	if strings.HasSuffix(base, "/chat/completions") {
+		return base
+	}
+	return base + "/chat/completions"
+}
+
+type chatCompletionResponse struct {
+	Choices []struct {
+		Message struct {
+			Content any `json:"content"`
+		} `json:"message"`
+	} `json:"choices"`
+	Error *struct {
+		Message string `json:"message"`
+		Type    string `json:"type"`
+		Code    any    `json:"code"`
+	} `json:"error"`
+}
+
+func contentText(content any) string {
+	switch v := content.(type) {
+	case string:
+		return v
+	case []any:
+		parts := make([]string, 0, len(v))
+		for _, item := range v {
+			obj, ok := item.(map[string]any)
+			if !ok {
+				continue
+			}
+			if text, ok := obj["text"].(string); ok {
+				parts = append(parts, text)
+			}
+		}
+		return strings.Join(parts, "\n")
+	default:
+		return ""
+	}
+}
+
+func compactBody(data []byte) string {
+	text := strings.TrimSpace(string(data))
+	if text == "" {
+		return "<empty body>"
+	}
+	if len(text) > 800 {
+		return text[:800] + "..."
+	}
+	return text
+}
diff --git a/internal/infrastructure/llm/openai_compatible_test.go b/internal/infrastructure/llm/openai_compatible_test.go
new file mode 100644
index 0000000..67c3845
--- /dev/null
+++ b/internal/infrastructure/llm/openai_compatible_test.go
@@ -0,0 +1,86 @@
+package llm
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"testing"
+
+	"github.com/mmmy/snapgo/internal/domain"
+)
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+func TestVisionClientSummarizeImageUsesChatCompletionImageURL(t *testing.T) {
+	var requestBody map[string]any
+	transport := roundTripFunc(func(r *http.Request) (*http.Response, error) {
+		if r.URL.Path != "/v1/chat/completions" {
+			t.Fatalf("expected /v1/chat/completions, got %s", r.URL.Path)
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer test-key" {
+			t.Fatalf("unexpected auth header %q", got)
+		}
+		if err := json.NewDecoder(r.Body).Decode(&requestBody); err != nil {
+			t.Fatalf("decode request: %v", err)
+		}
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     http.Header{"Content-Type": []string{"application/json"}},
+			Body:       io.NopCloser(bytes.NewBufferString(`{"choices":[{"message":{"content":"summary ok"}}]}`)),
+		}, nil
+	})
+
+	client, err := NewVisionClient(domain.LLMProviderQwen, domain.LLMProviderConfig{
+		BaseURL:     "https://example.test/v1",
+		APIKey:      "test-key",
+		Model:       "qwen-vl-plus",
+		MaxTokens:   321,
+		Temperature: 0.2,
+		TimeoutSecs: 5,
+	})
+	if err != nil {
+		t.Fatalf("new client: %v", err)
+	}
+	client.httpClient.Transport = transport
+
+	text, err := client.SummarizeImage(context.Background(), "describe", "https://cdn.example.com/a.png")
+	if err != nil {
+		t.Fatalf("summarize image: %v", err)
+	}
+	if text != "summary ok" {
+		t.Fatalf("expected response text, got %q", text)
+	}
+	if requestBody["model"] != "qwen-vl-plus" {
+		t.Fatalf("expected model forwarded, got %#v", requestBody["model"])
+	}
+	if requestBody["max_tokens"] != float64(321) {
+		t.Fatalf("expected max_tokens forwarded, got %#v", requestBody["max_tokens"])
+	}
+
+	messages := requestBody["messages"].([]any)
+	content := messages[0].(map[string]any)["content"].([]any)
+	if content[0].(map[string]any)["text"] != "describe" {
+		t.Fatalf("expected prompt content, got %#v", content[0])
+	}
+	image := content[1].(map[string]any)
+	if image["type"] != "image_url" {
+		t.Fatalf("expected image_url item, got %#v", image)
+	}
+	url := image["image_url"].(map[string]any)["url"]
+	if url != "https://cdn.example.com/a.png" {
+		t.Fatalf("expected image url forwarded, got %#v", url)
+	}
+}
+
+func TestChatCompletionsURLAcceptsFullEndpoint(t *testing.T) {
+	full := "https://example.com/v1/chat/completions"
+	if got := chatCompletionsURL(full); got != full {
+		t.Fatalf("expected full endpoint unchanged, got %q", got)
+	}
+}
diff --git a/native_overlay_callbacks_darwin.go b/native_overlay_callbacks_darwin.go
index f79bc58..d9ef0f7 100644
--- a/native_overlay_callbacks_darwin.go
+++ b/native_overlay_callbacks_darwin.go
@@ -74,6 +74,18 @@ func nativeOverlaySaveRemote(x, y, w, h C.int, annotationsJSON *C.char) {
 	}()
 }
 
+//export nativeOverlaySummarize
+func nativeOverlaySummarize(x, y, w, h C.int, annotationsJSON *C.char) {
+	app := consumeNativeOverlayApp()
+	if app == nil {
+		return
+	}
+	result := nativeCaptureResult(x, y, w, h, annotationsJSON)
+	go func() {
+		_ = app.SummarizeNativeRegion(result)
+	}()
+}
+
 //export nativeOverlayCancel
 func nativeOverlayCancel() {
 	app := consumeNativeOverlayApp()
diff --git a/native_overlay_darwin.go b/native_overlay_darwin.go
index bc93774..fbaa463 100644
--- a/native_overlay_darwin.go
+++ b/native_overlay_darwin.go
@@ -15,6 +15,7 @@ extern void nativeOverlayConfirm(int x, int y, int w, int h, const char *annotat
 extern void nativeOverlayCopy(int x, int y, int w, int h, const char *annotationsJSON);
 extern void nativeOverlaySave(int x, int y, int w, int h, const char *annotationsJSON, const char *dir);
 extern void nativeOverlaySaveRemote(int x, int y, int w, int h, const char *annotationsJSON);
+extern void nativeOverlaySummarize(int x, int y, int w, int h, const char *annotationsJSON);
 extern void nativeOverlayCancel(void);
 
 static NSWindow *nativeOverlayWindow = nil;
@@ -101,9 +102,8 @@ static id nativeOverlayKeyMonitor = nil;
 // as returning a +1 retained object.
 @property(strong) NSButton *clipboardButton;
 @property(strong) NSButton *saveButton;
-// `saveRemoteButton` triggers a placeholder action ("not yet supported")
-// per product spec; the icon comes from save-remote.svg.
 @property(strong) NSButton *saveRemoteButton;
+@property(strong) NSButton *summaryButton;
 @property(strong) NSButton *uploadButton;
 @property(strong) NSButton *penButton;
 @property(strong) NSButton *rectButton;
@@ -120,6 +120,7 @@ static id nativeOverlayKeyMonitor = nil;
 - (void)copySelection;
 - (void)saveSelection;
 - (void)saveRemoteSelection;
+- (void)summarizeSelection;
 - (void)cancelSelection;
 @end
 
@@ -144,6 +145,7 @@ static id nativeOverlayKeyMonitor = nil;
         _clipboardButton = [SnipHoverButton buttonWithTitle:@"" target:self action:@selector(copySelection)];
         _saveButton = [SnipHoverButton buttonWithTitle:@"" target:self action:@selector(saveSelection)];
         _saveRemoteButton = [SnipHoverButton buttonWithTitle:@"" target:self action:@selector(saveRemoteSelection)];
+        _summaryButton = [SnipHoverButton buttonWithTitle:@"" target:self action:@selector(summarizeSelection)];
         _uploadButton = [SnipHoverButton buttonWithTitle:@"" target:self action:@selector(confirmSelection)];
         // Tooltip strings shown on hover for each action button.
         // Localized in Chinese to match the rest of the action UI surface.
@@ -151,6 +153,7 @@ static id nativeOverlayKeyMonitor = nil;
         [_clipboardButton setToolTip:@"复制图片"];
         [_saveButton setToolTip:@"保存本地"];
         [_saveRemoteButton setToolTip:@"保存远端"];
+        [_summaryButton setToolTip:@"复制总结"];
         [_uploadButton setToolTip:@"上传云端"];
         _penButton = [NSButton buttonWithTitle:@"" target:self action:@selector(selectPen)];
         _rectButton = [NSButton buttonWithTitle:@"" target:self action:@selector(selectRect)];
@@ -171,10 +174,10 @@ static id nativeOverlayKeyMonitor = nil;
         // Add the action toolbar background BEFORE the buttons so it sits
         // behind them in the view hierarchy (AppKit z-order = subview order).
         [self addSubview:_actionToolbarBg];
-        for (NSView *view in @[_cancelButton, _clipboardButton, _saveButton, _saveRemoteButton, _uploadButton, _penButton, _rectButton, _ellipseButton, _colorButton, _undoButton, _paletteView, _sizeLabel, _hintLabel]) {
+        for (NSView *view in @[_cancelButton, _clipboardButton, _saveButton, _saveRemoteButton, _summaryButton, _uploadButton, _penButton, _rectButton, _ellipseButton, _colorButton, _undoButton, _paletteView, _sizeLabel, _hintLabel]) {
             [self addSubview:view];
         }
-        for (NSView *view in @[_cancelButton, _clipboardButton, _saveButton, _saveRemoteButton, _uploadButton, _penButton, _rectButton, _ellipseButton, _colorButton, _undoButton, _paletteView, _actionToolbarBg]) {
+        for (NSView *view in @[_cancelButton, _clipboardButton, _saveButton, _saveRemoteButton, _summaryButton, _uploadButton, _penButton, _rectButton, _ellipseButton, _colorButton, _undoButton, _paletteView, _actionToolbarBg]) {
             [view setHidden:YES];
         }
         [_sizeLabel setHidden:YES];
@@ -248,6 +251,7 @@ static id nativeOverlayKeyMonitor = nil;
     // save-remote.svg — provided by the user; its content is the same arrow
     // icon as the previous upload, so we reuse it verbatim here.
     NSImage *saveRemoteIcon = [self iconFromSVG:@"<svg viewBox='0 0 1024 1024' xmlns='http://www.w3.org/2000/svg'><path d='M481.749333 353.792c16.682667-16.64 43.690667-16.64 60.373334 0l120.917333 120.832 3.541333 4.010667a42.666667 42.666667 0 0 1-63.872 56.32L554.666667 486.954667V896a42.666667 42.666667 0 0 1-85.333334 0v-409.173333l-48.170666 48.128a42.666667 42.666667 0 0 1-60.330667 0l-3.541333-4.010667a42.666667 42.666667 0 0 1 3.541333-56.32zM512 85.333333c122.538667 0 227.84 73.813333 273.92 179.370667A213.333333 213.333333 0 0 1 725.333333 682.666667a42.666667 42.666667 0 0 1-4.992-85.034667L725.333333 597.333333a128 128 0 0 0 36.394667-250.794666l-38.144-11.264-15.914667-36.437334a213.376 213.376 0 0 0-407.296 58.026667l-7.381333 58.368-57.173333 13.824A85.418667 85.418667 0 0 0 256 597.333333h42.666667a42.666667 42.666667 0 0 1 0 85.333334H256a170.666667 170.666667 0 0 1-40.277333-336.554667A298.709333 298.709333 0 0 1 512 85.333333z' fill='white'/></svg>"];
+    NSImage *summaryIcon = [self iconFromSVG:@"<svg viewBox='0 0 1024 1024' xmlns='http://www.w3.org/2000/svg'><path d='M256 128h448l192 192v576H256V128z m64 64v640h512V352H672V192H320z m416 45.248V288h50.752L736 237.248zM384 448h384v64H384v-64z m0 128h384v64H384v-64z m0 128h256v64H384v-64zM192 256v704h512v64H128V256h64z' fill='white'/></svg>"];
     // upload.svg — newly replaced "send/cloud" icon supplied by the user.
     NSImage *uploadIcon = [self iconFromSVG:@"<svg viewBox='0 0 1024 1024' xmlns='http://www.w3.org/2000/svg'><path d='M550.528 544.128L640 539.776V678.4l178.005333-178.005333L640 322.389333v129.024l-72.618667 10.922667c-155.52 23.424-251.733333 73.557333-312.874666 164.437333 84.906667-50.602667 178.261333-76.928 296.021333-82.645333zM554.666667 629.376c-27.392 1.322667-53.034667 3.968-77.141334 7.808l-8.192 1.365333c-136.704 23.637333-224.810667 87.168-310.954666 176.128-26.154667 27.008-51.882667 59.648-51.882667 59.648-14.848 18.176-24.021333 14.037333-20.352-9.173333 0 0 4.394667-31.402667 11.690667-65.792C144.938667 577.066667 250.581333 423.68 554.666667 377.941333V209.066667a38.4 38.4 0 0 1 65.536-27.136l291.328 291.285333a38.4 38.4 0 0 1 0 54.314667l-291.328 291.285333A38.4 38.4 0 0 1 554.666667 791.637333v-162.261333z' fill='white'/></svg>"];
 
@@ -255,6 +259,7 @@ static id nativeOverlayKeyMonitor = nil;
     [self styleIconButton:_clipboardButton image:copyIcon];
     [self styleIconButton:_saveButton image:saveIcon];
     [self styleIconButton:_saveRemoteButton image:saveRemoteIcon];
+    [self styleIconButton:_summaryButton image:summaryIcon];
     [self styleIconButton:_uploadButton image:uploadIcon];
 
     // Configure hover colors for the 5 action buttons. White is the shared
@@ -267,11 +272,13 @@ static id nativeOverlayKeyMonitor = nil;
     SnipHoverButton *clipboardHB  = (SnipHoverButton *)_clipboardButton;
     SnipHoverButton *saveHB       = (SnipHoverButton *)_saveButton;
     SnipHoverButton *saveRemoteHB = (SnipHoverButton *)_saveRemoteButton;
+    SnipHoverButton *summaryHB    = (SnipHoverButton *)_summaryButton;
     SnipHoverButton *uploadHB     = (SnipHoverButton *)_uploadButton;
     cancelHB.baseColor = baseWhite;     cancelHB.hoverColor = hoverRed;
     clipboardHB.baseColor = baseWhite;  clipboardHB.hoverColor = hoverBlue;
     saveHB.baseColor = baseWhite;       saveHB.hoverColor = hoverBlue;
     saveRemoteHB.baseColor = baseWhite; saveRemoteHB.hoverColor = hoverBlue;
+    summaryHB.baseColor = baseWhite;    summaryHB.hoverColor = hoverBlue;
     uploadHB.baseColor = baseWhite;     uploadHB.hoverColor = hoverBlue;
 
     [self styleIconButton:_penButton image:[self iconFromSVG:@"<svg viewBox='0 0 1024 1024' xmlns='http://www.w3.org/2000/svg'><path d='M742.72 752.064c-29.696 28.576-42.976 48.96-42.976 77.12 0 98.4 143.776 142.464 229.728 65.536l-21.344-23.84c-66.976 59.936-176.384 26.4-176.384-41.696 0-16.832 9.28-31.104 33.184-54.08l13.44-12.736c61.024-58.016 75.328-102.72 29.312-179.84-43.84-73.44-96.8-88.288-162.784-50.56-47.232 27.04-68.64 47.456-208.32 190.4-70.624 72.288-153.92 77.088-217.344 26.784-59.712-47.328-79.872-127.36-42.176-188.64 24.512-39.84 62.656-72.896 136.64-124.48 5.952-4.192 27.04-18.816 31.104-21.664 12.16-8.448 21.536-15.104 30.4-21.536 107.552-78.272 140.8-139.136 86.048-219.904-76.992-113.472-202.304-90.016-367.744 61.472l21.6 23.616c153.088-140.224 256.896-159.616 319.68-67.104 41.632 61.408 16.896 106.688-78.4 176.032-8.64 6.304-17.92 12.832-29.888 21.184l-31.136 21.632c-77.504 54.08-117.984 89.184-145.536 133.984-46.784 76.032-22.176 173.664 49.536 230.496 76.224 60.416 177.952 54.56 260.096-29.504 136.16-139.36 158.08-160.224 201.312-184.96 50.656-28.96 84.416-19.52 119.424 39.168 36.896 61.824 27.52 91.392-23.808 140.16 0.096-0.064-10.976 10.368-13.632 12.96z' fill='white'/></svg>"]];
@@ -591,6 +598,7 @@ static id nativeOverlayKeyMonitor = nil;
     [_clipboardButton setHidden:!visible];
     [_saveButton setHidden:!visible];
     [_saveRemoteButton setHidden:!visible];
+    [_summaryButton setHidden:!visible];
     [_uploadButton setHidden:!visible];
     [_actionToolbarBg setHidden:!visible];
     [_penButton setHidden:!visible];
@@ -608,12 +616,12 @@ static id nativeOverlayKeyMonitor = nil;
     [_sizeLabel setStringValue:[NSString stringWithFormat:@"%.0f × %.0f", _selection.size.width, _selection.size.height]];
     [_sizeLabel setFrame:NSMakeRect(_selection.origin.x, MAX(0, _selection.origin.y - 26), 110, 22)];
 
-    // Action toolbar layout — 5 icon buttons, each 28x28, separated by 8px,
-    // with 4px outer padding. Total = 5*28 + 4*8 + 2*4 = 180px wide.
+    // Action toolbar layout — 6 icon buttons, each 28x28, separated by 8px,
+    // with 4px outer padding. Total = 6*28 + 5*8 + 2*4 = 216px wide.
     CGFloat actionBtnSize = 28;
     CGFloat actionGap = 8;
     CGFloat actionPad = 4;
-    NSInteger actionCount = 5;
+    NSInteger actionCount = 6;
     CGFloat toolbarW = actionPad * 2 + actionBtnSize * actionCount + actionGap * (actionCount - 1);
     CGFloat toolbarH = 40;
     CGFloat x = _selection.origin.x + _selection.size.width - toolbarW;
@@ -631,7 +639,8 @@ static id nativeOverlayKeyMonitor = nil;
     [_clipboardButton  setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 1, btnY, actionBtnSize, actionBtnSize)];
     [_saveButton       setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 2, btnY, actionBtnSize, actionBtnSize)];
     [_saveRemoteButton setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 3, btnY, actionBtnSize, actionBtnSize)];
-    [_uploadButton     setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 4, btnY, actionBtnSize, actionBtnSize)];
+    [_summaryButton    setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 4, btnY, actionBtnSize, actionBtnSize)];
+    [_uploadButton     setFrame:NSMakeRect(baseX + (actionBtnSize + actionGap) * 5, btnY, actionBtnSize, actionBtnSize)];
 
     // Mark toolbar sits to the LEFT of the action toolbar (same row) with an
     // 8px gap between the two groups, so they never overlap on tiny selections.
@@ -799,6 +808,18 @@ static id nativeOverlayKeyMonitor = nil;
     NSString *json = [self annotationsJSON];
     nativeOverlaySaveRemote((int)llround(r.origin.x), (int)llround(r.origin.y), (int)llround(r.size.width), (int)llround(r.size.height), [json UTF8String]);
 }
+
+// `summarizeSelection` uploads the screenshot to S3, sends it to the
+// configured multimodal model, and copies the generated summary.
+- (void)summarizeSelection {
+    if (!_hasSelection) {
+        return;
+    }
+    NSRect r = [self globalRectForSelection:_selection];
+    [self closeOverlayWindow];
+    NSString *json = [self annotationsJSON];
+    nativeOverlaySummarize((int)llround(r.origin.x), (int)llround(r.origin.y), (int)llround(r.size.width), (int)llround(r.size.height), [json UTF8String]);
+}
 @end
 
 // screenContainingCursor returns the NSScreen the cursor is currently on,
diff --git a/operation_status_darwin.go b/operation_status_darwin.go
new file mode 100644
index 0000000..4de9698
--- /dev/null
+++ b/operation_status_darwin.go
@@ -0,0 +1,163 @@
+//go:build darwin
+
+package main
+
+/*
+#cgo CFLAGS: -x objective-c -fobjc-arc -fblocks
+#cgo LDFLAGS: -framework AppKit -framework QuartzCore
+#include <stdlib.h>
+#include <dispatch/dispatch.h>
+#import <AppKit/AppKit.h>
+#import <QuartzCore/QuartzCore.h>
+
+static NSPanel *operationPanel = nil;
+static NSTextField *operationTitleLabel = nil;
+static NSTextField *operationDetailLabel = nil;
+static NSTextField *operationIconLabel = nil;
+static NSProgressIndicator *operationSpinner = nil;
+
+static NSString *statusString(const char *value) {
+    if (value == NULL) {
+        return @"";
+    }
+    NSString *text = [NSString stringWithUTF8String:value];
+    return text ?: @"";
+}
+
+static void ensureOperationPanel(void) {
+    if (operationPanel != nil) {
+        return;
+    }
+
+    NSRect frame = NSMakeRect(0, 0, 360, 118);
+    operationPanel = [[NSPanel alloc]
+        initWithContentRect:frame
+                  styleMask:NSWindowStyleMaskBorderless
+                    backing:NSBackingStoreBuffered
+                      defer:NO];
+    [operationPanel setOpaque:NO];
+    [operationPanel setBackgroundColor:[NSColor clearColor]];
+    [operationPanel setLevel:NSFloatingWindowLevel];
+    [operationPanel setHidesOnDeactivate:NO];
+    [operationPanel setCollectionBehavior:
+        NSWindowCollectionBehaviorCanJoinAllSpaces |
+        NSWindowCollectionBehaviorFullScreenAuxiliary |
+        NSWindowCollectionBehaviorStationary];
+
+    NSView *root = [[NSView alloc] initWithFrame:frame];
+    [root setWantsLayer:YES];
+    [[root layer] setCornerRadius:14];
+    [[root layer] setBackgroundColor:[[NSColor colorWithCalibratedWhite:0.08 alpha:0.92] CGColor]];
+    [[root layer] setShadowColor:[[NSColor blackColor] CGColor]];
+    [[root layer] setShadowOpacity:0.28];
+    [[root layer] setShadowRadius:18];
+    [[root layer] setShadowOffset:CGSizeMake(0, -8)];
+
+    operationSpinner = [[NSProgressIndicator alloc] initWithFrame:NSMakeRect(24, 42, 34, 34)];
+    [operationSpinner setStyle:NSProgressIndicatorStyleSpinning];
+    [operationSpinner setIndeterminate:YES];
+
+    operationIconLabel = [NSTextField labelWithString:@"OK"];
+    [operationIconLabel setFrame:NSMakeRect(21, 43, 42, 30)];
+    [operationIconLabel setAlignment:NSTextAlignmentCenter];
+    [operationIconLabel setFont:[NSFont systemFontOfSize:17 weight:NSFontWeightSemibold]];
+    [operationIconLabel setTextColor:[NSColor colorWithCalibratedRed:74.0/255.0 green:222.0/255.0 blue:128.0/255.0 alpha:1.0]];
+    [operationIconLabel setHidden:YES];
+
+    operationTitleLabel = [NSTextField labelWithString:@""];
+    [operationTitleLabel setFrame:NSMakeRect(76, 61, 260, 24)];
+    [operationTitleLabel setFont:[NSFont systemFontOfSize:15 weight:NSFontWeightSemibold]];
+    [operationTitleLabel setTextColor:[NSColor whiteColor]];
+
+    operationDetailLabel = [NSTextField labelWithString:@""];
+    [operationDetailLabel setFrame:NSMakeRect(76, 34, 260, 22)];
+    [operationDetailLabel setFont:[NSFont systemFontOfSize:12 weight:NSFontWeightRegular]];
+    [operationDetailLabel setTextColor:[NSColor colorWithCalibratedWhite:0.82 alpha:1.0]];
+    [operationDetailLabel setLineBreakMode:NSLineBreakByTruncatingTail];
+
+    [root addSubview:operationSpinner];
+    [root addSubview:operationIconLabel];
+    [root addSubview:operationTitleLabel];
+    [root addSubview:operationDetailLabel];
+    [operationPanel setContentView:root];
+}
+
+static void snipShowOperationStatus(const char *titleC, const char *detailC, int state) {
+    NSString *title = statusString(titleC);
+    NSString *detail = statusString(detailC);
+    dispatch_async(dispatch_get_main_queue(), ^{
+        ensureOperationPanel();
+        [operationTitleLabel setStringValue:title];
+        [operationDetailLabel setStringValue:detail];
+
+        if (state == 0) {
+            [operationIconLabel setHidden:YES];
+            [operationSpinner setHidden:NO];
+            [operationSpinner startAnimation:nil];
+        } else {
+            [operationSpinner stopAnimation:nil];
+            [operationSpinner setHidden:YES];
+            [operationIconLabel setHidden:NO];
+            if (state == 2) {
+                [operationIconLabel setStringValue:@"!"];
+                [operationIconLabel setTextColor:[NSColor colorWithCalibratedRed:248.0/255.0 green:113.0/255.0 blue:113.0/255.0 alpha:1.0]];
+            } else {
+                [operationIconLabel setStringValue:@"OK"];
+                [operationIconLabel setTextColor:[NSColor colorWithCalibratedRed:74.0/255.0 green:222.0/255.0 blue:128.0/255.0 alpha:1.0]];
+            }
+        }
+
+        NSScreen *screen = [NSScreen mainScreen];
+        if (screen != nil) {
+            NSRect visible = [screen visibleFrame];
+            NSRect panelFrame = [operationPanel frame];
+            panelFrame.origin.x = NSMidX(visible) - panelFrame.size.width / 2;
+            panelFrame.origin.y = NSMaxY(visible) - panelFrame.size.height - 48;
+            [operationPanel setFrame:panelFrame display:YES];
+        }
+        [operationPanel orderFrontRegardless];
+    });
+}
+
+static void snipHideOperationStatus(void) {
+    dispatch_async(dispatch_get_main_queue(), ^{
+        if (operationPanel != nil) {
+            [operationSpinner stopAnimation:nil];
+            [operationPanel orderOut:nil];
+        }
+    });
+}
+*/
+import "C"
+
+import (
+	"sync/atomic"
+	"time"
+	"unsafe"
+)
+
+const (
+	operationStatusRunning = iota
+	operationStatusSuccess
+	operationStatusError
+)
+
+var operationStatusSeq atomic.Uint64
+
+func showOperationStatus(title, detail string, state int) {
+	operationStatusSeq.Add(1)
+	ctitle := C.CString(title)
+	cdetail := C.CString(detail)
+	defer C.free(unsafe.Pointer(ctitle))
+	defer C.free(unsafe.Pointer(cdetail))
+	C.snipShowOperationStatus(ctitle, cdetail, C.int(state))
+}
+
+func hideOperationStatusAfter(delay time.Duration) {
+	seq := operationStatusSeq.Load()
+	time.AfterFunc(delay, func() {
+		if operationStatusSeq.Load() == seq {
+			C.snipHideOperationStatus()
+		}
+	})
+}
diff --git a/operation_status_other.go b/operation_status_other.go
new file mode 100644
index 0000000..28b4b3f
--- /dev/null
+++ b/operation_status_other.go
@@ -0,0 +1,15 @@
+//go:build !darwin
+
+package main
+
+import "time"
+
+const (
+	operationStatusRunning = iota
+	operationStatusSuccess
+	operationStatusError
+)
+
+func showOperationStatus(_ string, _ string, _ int) {}
+
+func hideOperationStatusAfter(_ time.Duration) {}