diff --git a/internal/plugins/remotedevice/binary_test.go b/internal/plugins/remotedevice/binary_test.go
index e70c927..db18ca2 100644
--- a/internal/plugins/remotedevice/binary_test.go
+++ b/internal/plugins/remotedevice/binary_test.go
@@ -295,3 +295,76 @@ func TestPushDataOfflineDevice(t *testing.T) {
 		t.Fatalf("expected not online error, got %v", err)
 	}
 }
+
+// ===== screensee：截屏回传 + 视觉描述回调 =====
+
+func TestScreenseeEndToEnd(t *testing.T) {
+	reg := NewRegistry()
+	token := "test-token-see"
+	reg.SetAcceptToken(func(provided string) bool { return provided == token })
+
+	dev := &devicectlDevice{reg: reg}
+	var gotDataURL string
+	dev.SetSeeHandler(func(dataURL string, provider string) string {
+		gotDataURL = dataURL
+		return "屏幕上显示的是测试画面"
+	})
+
+	srv := httptest.NewServer(http.HandlerFunc(reg.ServeWS))
+	defer srv.Close()
+
+	cli := dialTestWS(t, srv.URL, token)
+	defer cli.close()
+
+	// 设备 hello + bind（bind 需 token 才能被授权流程识别，这里直接手动授权）
+	cli.sendText([]byte(`{"op":"hello","device":{"device_id":"see-dev","name":"屏幕机","kind":"computer","caps":["cmd"]}}`))
+	if _, _, err := cli.readMsg(); err != nil {
+		t.Fatalf("read hello_ack: %v", err)
+	}
+	reg.SetAuthorized("see-dev", true)
+
+	// 设备侧循环收命令并回执（模拟 GUI screensee 实现）
+	go func() {
+		for {
+			op, payload, err := cli.readMsg()
+			if err != nil {
+				return
+			}
+			if op != 0x1 {
+				continue
+			}
+			var msg map[string]interface{}
+			if json.Unmarshal(payload, &msg) != nil {
+				continue
+			}
+			if msg["op"] == "cmd" && msg["command"] == "screensee" {
+				reqID, _ := msg["req_id"].(string)
+				fakeJPEG := []byte{0xFF, 0xD8, 0xFF, 0xE0, 0x00, 0x10} // JPEG magic
+				b64 := base64.StdEncoding.EncodeToString(fakeJPEG)
+				cli.sendText(mustJSON(map[string]interface{}{
+					"op": "cmd_result", "req_id": reqID, "device_id": "see-dev",
+					"status": "ok", "output": "data:image/jpeg;base64," + b64,
+				}))
+			}
+		}
+	}()
+
+	// agent 调用 screensee
+	res, err := dev.Execute("screensee", map[string]interface{}{"device_id": "see-dev"})
+	if err != nil {
+		t.Fatalf("screensee: %v", err)
+	}
+	m := res.(map[string]interface{})
+	if m["description"] != "屏幕上显示的是测试画面" {
+		t.Fatalf("unexpected description: %v", m["description"])
+	}
+	if !strings.HasPrefix(gotDataURL, "data:image/jpeg;base64,") {
+		t.Fatalf("handler received bad dataURL: %s", gotDataURL)
+	}
+
+	// 未授权设备应拒绝
+	reg.SetAuthorized("see-dev", false)
+	if _, err := dev.Execute("screensee", map[string]interface{}{"device_id": "see-dev"}); err == nil {
+		t.Fatal("expected unauthorized error")
+	}
+}
diff --git a/internal/plugins/remotedevice/device.go b/internal/plugins/remotedevice/device.go
index 4547925..3a4f0a4 100644
--- a/internal/plugins/remotedevice/device.go
+++ b/internal/plugins/remotedevice/device.go
@@ -10,11 +10,15 @@ import (
 )
 
 // devicectlDevice 把设备网关暴露为 IOManager 的一个 Device：
-// Tools() 提供 devicedetect / device_ctl_status / device_ctl_cmdrun / device_ctl_cmdresult，
+// Tools() 提供 devicedetect / device_ctl_status / device_ctl_cmdrun / device_ctl_cmdresult / screensee，
 // Execute() 检查授权并路由到 WS 在线设备。
 type devicectlDevice struct {
 	reg     *Registry
 	persist func() // 授权变更后持久化
+
+	// screensee 回调：设备截屏回传后由 agent 核心消费（视觉描述）。
+	// 由插件 Start 注入；nil 时退化为仅返回 base64 数据。
+	seeHandler func(dataURL string, provider string) string
 }
 
 func (d *devicectlDevice) Name() string                                 { return "devicectl" }
@@ -89,6 +93,21 @@ func (d *devicectlDevice) Tools() []agentIO.ToolDef {
 				},
 			},
 		},
+		{
+			Name: "screensee",
+			Description: "查看一台已授权设备的屏幕当前画面（截屏回传）。" +
+				"与 screensue（向用户屏幕显示内容）配对：screensue 是给用户看，screensee 是你看。" +
+				"返回屏幕截图的自动视觉描述；如需读取屏上文字可接着用 ocr_image。" +
+				"需要 device_id（来自 devicedetect）。设备必须已授权且在线。",
+			Parameters: map[string]interface{}{
+				"type": "object",
+				"properties": map[string]interface{}{
+					"device_id": map[string]interface{}{"type": "string", "description": "目标设备 ID"},
+					"provider":  map[string]interface{}{"type": "string", "description": "可选：用于视觉描述的 LLM 源名称，不填则使用默认模型"},
+				},
+				"required": []interface{}{"device_id"},
+			},
+		},
 		{
 			Name: "deviceinfo",
 			Description: "探查一台设备接入网关时声明的详细信息与支持能力。" +
@@ -115,6 +134,8 @@ func (d *devicectlDevice) Execute(tool string, args map[string]interface{}) (int
 		return d.cmdrun(args)
 	case "device_ctl_cmdresult":
 		return d.cmdresult(args)
+	case "screensee":
+		return d.screensee(args)
 	case "deviceinfo":
 		return d.info(args)
 	default:
@@ -283,3 +304,55 @@ func (d *devicectlDevice) info(args map[string]interface{}) (interface{}, error)
 	}
 	return out, nil
 }
+
+// SetSeeHandler 注入 screensee 的视觉描述回调（agent 核心提供）。
+func (d *devicectlDevice) SetSeeHandler(fn func(dataURL string, provider string) string) {
+	d.seeHandler = fn
+}
+
+// screensee 实现 screensee：向设备下发 homeagent-screensee 截屏命令，
+// 等待回传 jpeg base64，交给 seeHandler（agent 核心）做视觉描述。
+func (d *devicectlDevice) screensee(args map[string]interface{}) (interface{}, error) {
+	id, _ := args["device_id"].(string)
+	provider, _ := args["provider"].(string)
+	if id == "" {
+		return nil, fmt.Errorf("device_id required")
+	}
+	m, ok := d.reg.Get(id)
+	if !ok {
+		return nil, fmt.Errorf("device %s 不存在", id)
+	}
+	if !m.Authorized {
+		return nil, fmt.Errorf("device %s 未授权，无法查看屏幕（请先在设备管理页授权）", id)
+	}
+	if !m.Online {
+		return nil, fmt.Errorf("device %s 不在线", id)
+	}
+	reqID := newReqID()
+	if err := d.reg.PushCmd(id, reqID, "screensee", "homeagent"); err != nil {
+		return nil, fmt.Errorf("下发截屏命令失败: %w", err)
+	}
+	res, err := d.reg.AwaitResult(reqID, 30*time.Second)
+	if err != nil {
+		d.reg.SaveResult(reqID, map[string]interface{}{"accepted": true, "error": err.Error(), "pending": true})
+		return nil, fmt.Errorf("设备未在超时内回传屏幕画面: %w", err)
+	}
+	if res["status"] != "ok" {
+		errMsg, _ := res["error"].(string)
+		if errMsg == "" {
+			errMsg = fmt.Sprintf("status=%v", res["status"])
+		}
+		return nil, fmt.Errorf("设备截屏失败: %s", errMsg)
+	}
+	output, _ := res["output"].(string)
+	// 设备端回传 data URL（data:image/jpeg;base64,...）或裸 base64
+	if !strings.HasPrefix(output, "data:") {
+		output = "data:image/jpeg;base64," + output
+	}
+	d.reg.SaveResult(reqID, res)
+	if d.seeHandler == nil {
+		return map[string]interface{}{"image_data_url": output, "note": "无视觉描述处理器，仅返回原始图像数据"}, nil
+	}
+	desc := d.seeHandler(output, provider)
+	return map[string]interface{}{"description": desc}, nil
+}
diff --git a/internal/plugins/remotedevice/plugin.go b/internal/plugins/remotedevice/plugin.go
index 1295bdd..b23e74c 100644
--- a/internal/plugins/remotedevice/plugin.go
+++ b/internal/plugins/remotedevice/plugin.go
@@ -109,6 +109,8 @@ func (p *Plugin) Start(s *sdk.PluginSDK) error {
 
 	// ---- devicectl Device（agent 工具） ----------------
 	p.dev = &devicectlDevice{reg: p.registry, persist: p.persistAuthorized}
+	// screensee 视觉描述回调：截屏回传后用视觉模型描述屏幕内容
+	p.dev.SetSeeHandler(p.describeScreen)
 	if err := s.RegisterChannel("devicectl", p.dev); err != nil {
 		log.Printf("[remotedevice] register devicectl channel: %v", err)
 	}
@@ -250,6 +252,41 @@ func (p *Plugin) handleDeviceAuth(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, map[string]interface{}{"device_id": req.DeviceID, "authorized": req.Authorize})
 }
 
+// describeScreen 用视觉模型描述设备屏幕截图（screensee 回调）。
+// provider 为空时使用默认 LLM 源；模型不支持视觉时返回友好错误。
+func (p *Plugin) describeScreen(dataURL string, provider string) string {
+	if p.sdk == nil || p.sdk.LLM() == nil {
+		return "LLM 不可用，无法描述屏幕内容"
+	}
+	llm := p.sdk.LLM()
+	req := &sdk.LLMCompletionRequest{
+		MaxTokens: 2048,
+		Messages: []sdk.LLMMessage{{
+			Role: "user",
+			Blocks: []sdk.LLMContentBlock{
+				{Type: "text", Text: "这是用户设备的屏幕截图。请详细描述屏幕上显示的内容：正在运行的窗口/应用、可见的文字内容、界面状态等。如果是代码编辑器或终端，尽量转述关键文字信息。"},
+				{Type: "image_url", ImageURL: dataURL},
+			},
+		}},
+	}
+	// 指定源：临时切换（低频操作，用完恢复原源）
+	if provider != "" {
+		prev := llm.CurrentSource()
+		if err := llm.SetSource(provider); err != nil {
+			log.Printf("[remotedevice] screensee set source %s: %v", provider, err)
+		} else if prev != "" {
+			defer func() { _ = llm.SetSource(prev) }()
+		}
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
+	defer cancel()
+	resp, err := llm.Chat(ctx, req)
+	if err != nil {
+		return fmt.Sprintf("屏幕截图视觉描述失败: %v（当前模型可能不支持图像输入）", err)
+	}
+	return resp.Content
+}
+
 func (p *Plugin) Stop() error {
 	if p.server != nil {
 		ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
diff --git a/internal/sdk/llm.go b/internal/sdk/llm.go
index 6e1baf5..e585248 100644
--- a/internal/sdk/llm.go
+++ b/internal/sdk/llm.go
@@ -23,6 +23,16 @@ type LLMMessage struct {
 	ReasoningContent string        `json:"reasoning_content,omitempty"`
 	ToolCallID       string        `json:"tool_call_id,omitempty"`
 	ToolCalls        []LLMToolCall `json:"tool_calls,omitempty"`
+	// Blocks 多模态内容块（与 Content 二选一；非空时优先）。
+	// 支持 text 与 image_url 两类，用于视觉模型看图（如 screensee 截屏描述）。
+	Blocks []LLMContentBlock `json:"blocks,omitempty"`
+}
+
+// LLMContentBlock 是多模态消息中的单个内容块。
+type LLMContentBlock struct {
+	Type     string `json:"type"` // "text" | "image_url"
+	Text     string `json:"text,omitempty"`
+	ImageURL string `json:"image_url,omitempty"` // data URL 或 http(s) URL
 }
 
 // LLMToolCall 是中立的工具调用请求。
diff --git a/internal/sdk/llm_impl.go b/internal/sdk/llm_impl.go
index a923898..dafbb4f 100644
--- a/internal/sdk/llm_impl.go
+++ b/internal/sdk/llm_impl.go
@@ -71,6 +71,18 @@ func (l *llmImpl) Chat(ctx context.Context, req *LLMCompletionRequest) (*LLMComp
 				ReasoningContent: m.ReasoningContent,
 				ToolCallID:       m.ToolCallID,
 			}
+			// 多模态 Blocks：text/image_url → agentAPI.ContentBlock
+			for _, b := range m.Blocks {
+				switch b.Type {
+				case "text":
+					msg.Blocks = append(msg.Blocks, agentAPI.ContentBlock{Type: "text", Text: b.Text})
+				case "image_url":
+					msg.Blocks = append(msg.Blocks, agentAPI.ContentBlock{
+						Type:     "image_url",
+						ImageURL: &agentAPI.ImageURL{URL: b.ImageURL, Detail: "high"},
+					})
+				}
+			}
 			if len(m.ToolCalls) > 0 {
 				msg.ToolCalls = make([]agentAPI.ToolCall, len(m.ToolCalls))
 				for j, tc := range m.ToolCalls {