diff --git a/backend/controller/agent_cli.go b/backend/controller/agent_cli.go
index e85f5bd..128181e 100644
--- a/backend/controller/agent_cli.go
+++ b/backend/controller/agent_cli.go
@@ -166,6 +166,7 @@ func (ac *agentConn) startTask(sshCfg remoteexec.SSHConfig, llmCfg model.LLMConf
 	// 构建引擎
 	engine, err := agentcli.NewEngine(agentcli.EngineConfig{
 		LLM: llm.Config{
+			Protocol: resolveProtocol(llmCfg.Provider),
 			Endpoint: resolveEndpoint(llmCfg.Provider, llmCfg.Endpoint),
 			APIKey:   llmCfg.APIKey, Model: llmCfg.Model,
 			Temperature: llmCfg.Temperature, MaxTokens: llmCfg.MaxTokens,
diff --git a/backend/controller/credential_assign.go b/backend/controller/credential_assign.go
index 81432c8..8fed486 100644
--- a/backend/controller/credential_assign.go
+++ b/backend/controller/credential_assign.go
@@ -1,6 +1,7 @@
 package controller
 
 import (
+	"encoding/json"
 	"net/http"
 	"strconv"
 
@@ -43,15 +44,7 @@ func (h *CredentialHandler) AssignCredential(c *gin.Context) {
 		return
 	}
 
-	if cred.SecretType == "ssh" {
-		server.SSHKey = plain
-		server.AuthType = "key"
-		server.Password = ""
-	} else {
-		server.Password = plain
-		server.AuthType = "password"
-		server.SSHKey = ""
-	}
+	applyCredential(cred, plain, &server)
 
 	if err := h.db.Save(&server).Error; err != nil {
 		c.JSON(http.StatusInternalServerError, model.Response{Code: 500, Message: "分配失败: " + err.Error()})
@@ -60,3 +53,30 @@ func (h *CredentialHandler) AssignCredential(c *gin.Context) {
 
 	c.JSON(http.StatusOK, model.Response{Code: 200, Message: "凭据已关联"})
 }
+
+// applyCredential 将解密后的明文 plain 填充到 server 认证字段（统一入口，避免 plain JSON 误用） // Logan.Li
+// ssh 凭证 plain 是 JSON {"private_key":...,"ssh_user":...,"ssh_port":...}，须解析出 private_key（旧实现直接把
+// 整个 JSON 当私钥，导致 ssh.ParsePrivateKey 失败 → ssh-key 登录不通）；其余类型 plain 即密码。
+// 供 AssignCredential / 单台服务器 CRUD / 批量导入 三处复用。
+func applyCredential(cred model.Credential, plain string, server *model.Server) {
+	if cred.SecretType == "ssh" {
+		privateKey := plain
+		var sshc model.SSHSecretContent
+		if json.Unmarshal([]byte(plain), &sshc) == nil && sshc.PrivateKey != "" {
+			privateKey = sshc.PrivateKey
+			if sshc.SSHUser != "" {
+				server.Username = sshc.SSHUser
+			}
+			if sshc.SSHPort > 0 {
+				server.Port = sshc.SSHPort
+			}
+		}
+		server.SSHKey = privateKey
+		server.AuthType = "key"
+		server.Password = ""
+	} else {
+		server.Password = plain
+		server.AuthType = "password"
+		server.SSHKey = ""
+	}
+}
diff --git a/backend/controller/llm.go b/backend/controller/llm.go
index 5426331..3d683a6 100644
--- a/backend/controller/llm.go
+++ b/backend/controller/llm.go
@@ -122,23 +122,6 @@ func (h *LLMHandler) SaveLLMConfig(c *gin.Context) {
 	}})
 }
 
-// OpenAI 兼容请求/响应结构（仅取必要字段） // Logan.Li
-type openaiChatRequest struct {
-	Model       string               `json:"model"`
-	Messages    []model.ChatMessage `json:"messages"`
-	Temperature float64              `json:"temperature"`
-	MaxTokens   int                  `json:"max_tokens,omitempty"`
-}
-type openaiChoice struct {
-	Message model.ChatMessage `json:"message"`
-}
-type openaiChatResponse struct {
-	Choices []openaiChoice `json:"choices"`
-	Usage   struct {
-		TotalTokens int `json:"total_tokens"`
-	} `json:"usage"`
-}
-
 // resolveEndpoint 根据提供商补全默认 endpoint // Logan.Li
 func resolveEndpoint(provider, endpoint string) string {
 	if endpoint != "" {
@@ -147,10 +130,20 @@ func resolveEndpoint(provider, endpoint string) string {
 	switch provider {
 	case "deepseek":
 		return "https://api.deepseek.com/v1"
+	case "glm": // 智谱 GLM 通用 OpenAI 兼容端点（编码套餐改用 https://open.bigmodel.cn/api/coding/paas/v4） // Logan.Li
+		return "https://open.bigmodel.cn/api/paas/v4"
 	case "qwen":
 		return "https://dashscope.aliyuncs.com/compatible-mode/v1"
 	case "moonshot":
 		return "https://api.moonshot.cn/v1"
+	case "kimi": // Kimi 编码套餐 OpenAI 兼容端点（注意：该端点有客户端白名单，通用调用可能被拒） // Logan.Li
+		return "https://api.kimi.com/coding/v1"
+	case "mimo": // 小米 MiMo OpenAI 兼容端点 // Logan.Li
+		return "https://api.mimo.mi.com/v1"
+	case "anthropic": // Anthropic 官方 Messages 协议（anthropic.go 客户端拼 /v1/messages） // Logan.Li
+		return "https://api.anthropic.com"
+	case "glm-coding": // 智谱 GLM 编码套餐 Anthropic 协议端点（模型 GLM-5.2） // Logan.Li
+		return "https://open.bigmodel.cn/api/anthropic"
 	case "ollama":
 		return "http://localhost:11434/v1"
 	default: // openai / custom
@@ -158,6 +151,15 @@ func resolveEndpoint(provider, endpoint string) string {
 	}
 }
 
+// resolveProtocol 根据提供商推断接入协议（决定走 OpenAI 还是 Anthropic 客户端） // Logan.Li
+// 编码套餐（智谱 GLM 编码 / Anthropic 官方）走 Anthropic Messages 协议；其余走 OpenAI 兼容协议。
+func resolveProtocol(provider string) string {
+	if provider == "anthropic" || provider == "glm-coding" {
+		return "anthropic"
+	}
+	return "openai"
+}
+
 // defaultSystemPrompt AIOps 运维助手默认系统提示词 // Logan.Li
 const defaultSystemPrompt = `你是 AIOps 智能运维助手，专注于企业级运维场景。你可以：
 1. 分析服务器、容器、K8s 集群的性能瓶颈与异常
diff --git a/backend/controller/llm_chat.go b/backend/controller/llm_chat.go
index ab03383..1f0ffe2 100644
--- a/backend/controller/llm_chat.go
+++ b/backend/controller/llm_chat.go
@@ -1,22 +1,41 @@
 package controller
 
 import (
-	"bufio"
-	"bytes"
 	"encoding/json"
 	"fmt"
-	"io"
 	"net/http"
 	"strings"
 	"time"
 
 	"ezops/middleware"
 	"ezops/model"
+	"ezops/pkg/agentcli/llm"
 
 	"github.com/gin-gonic/gin"
 )
 
-// Chat AI 对话代理（OpenAI 兼容接口） // Logan.Li
+// buildLLMClient 从数据库配置构造 LLM 客户端（按 provider 协议自动路由 OpenAI/Anthropic） // Logan.Li
+func (h *LLMHandler) buildLLMClient(cfg model.LLMConfig) llm.Client {
+	return llm.NewClient(llm.Config{
+		Protocol:    resolveProtocol(cfg.Provider),
+		Endpoint:    resolveEndpoint(cfg.Provider, cfg.Endpoint),
+		APIKey:      cfg.APIKey,
+		Model:       cfg.Model,
+		Temperature: cfg.Temperature,
+		MaxTokens:   cfg.MaxTokens,
+	})
+}
+
+// toLLMMessages model.ChatMessage[] → llm.Message[]（智能助手纯对话，仅填 Role/Content） // Logan.Li
+func toLLMMessages(msgs []model.ChatMessage) []llm.Message {
+	out := make([]llm.Message, len(msgs))
+	for i, m := range msgs {
+		out[i] = llm.Message{Role: m.Role, Content: m.Content}
+	}
+	return out
+}
+
+// Chat AI 对话代理（统一走 llm.Client，按配置协议自动适配 OpenAI/Anthropic） // Logan.Li
 // POST /api/v1/aiops/chat
 func (h *LLMHandler) Chat(c *gin.Context) {
 	var cfg model.LLMConfig
@@ -38,60 +57,22 @@ func (h *LLMHandler) Chat(c *gin.Context) {
 	}
 	messages := append([]model.ChatMessage{{Role: "system", Content: systemPrompt}}, req.Messages...)
 
-	// 构造 OpenAI 兼容请求
-	payload := openaiChatRequest{
-		Model:       cfg.Model,
-		Messages:    messages,
-		Temperature: cfg.Temperature,
-		MaxTokens:   cfg.MaxTokens,
-	}
-	body, _ := json.Marshal(payload)
-
-	endpoint := resolveEndpoint(cfg.Provider, cfg.Endpoint)
-	url := endpoint + "/chat/completions"
-
-	httpReq, err := http.NewRequestWithContext(c.Request.Context(), http.MethodPost, url, bytes.NewReader(body))
+	resp, err := h.buildLLMClient(cfg).Complete(c.Request.Context(), &llm.CompletionRequest{
+		Messages: toLLMMessages(messages),
+	})
 	if err != nil {
-		c.JSON(http.StatusOK, model.Response{Code: 500, Message: "构造请求失败: " + err.Error()})
+		c.JSON(http.StatusOK, model.Response{Code: 502, Message: "LLM 调用失败: " + err.Error()})
 		return
 	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	httpReq.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-
-	resp, err := h.client.Do(httpReq)
-	if err != nil {
-		c.JSON(http.StatusOK, model.Response{Code: 502, Message: "LLM 服务连接失败，请检查网络或 Endpoint 配置: " + err.Error()})
-		return
-	}
-	defer resp.Body.Close()
-
-	raw, _ := io.ReadAll(resp.Body)
-	if resp.StatusCode != http.StatusOK {
-		// 提取上游错误信息
-		msg := fmt.Sprintf("LLM 服务返回错误 (HTTP %d)", resp.StatusCode)
-		if upErr := parseOpenAIError(raw); upErr != "" {
-			msg += ": " + upErr
-		}
-		c.JSON(http.StatusOK, model.Response{Code: resp.StatusCode, Message: msg})
-		return
-	}
-
-	var chatResp openaiChatResponse
-	if err := json.Unmarshal(raw, &chatResp); err != nil || len(chatResp.Choices) == 0 {
-		c.JSON(http.StatusOK, model.Response{Code: 500, Message: "解析 LLM 响应失败"})
-		return
-	}
-	replyContent := chatResp.Choices[0].Message.Content
+	replyContent := resp.Message.Content
 
 	// 持久化对话历史（按用户隔离，成对写入 user + assistant） // Logan.Li
 	// 仅在成功时入库；错误气泡为瞬时 UI 不持久化
 	uid := middleware.GetUserID(c)
 	if uid != 0 && len(req.Messages) > 0 {
-		// 取最后一条用户消息（前端已附加当前提问）
 		lastUser := req.Messages[len(req.Messages)-1]
 		now := time.Now()
-		// session_id：前端传入或后端生成（兼容旧版客户端） // Logan.Li
-		sid := req.SessionID
+		sid := req.SessionID // 前端传入或后端生成（兼容旧版客户端）
 		if sid == "" {
 			sid = fmt.Sprintf("sess-%d", now.UnixMilli())
 		}
@@ -104,15 +85,15 @@ func (h *LLMHandler) Chat(c *gin.Context) {
 	c.JSON(http.StatusOK, model.Response{Code: 200, Data: model.ChatResponse{
 		Content:    replyContent,
 		Model:      cfg.Model,
-		TokensUsed: chatResp.Usage.TotalTokens,
+		TokensUsed: resp.Usage.TotalTokens,
 	}})
 }
 
 // ChatStream AI 流式对话代理（SSE） // Logan.Li
 // POST /api/v1/aiops/chat/stream
-// 客户端用 EventSource/fetch-stream 消费，每条 SSE data 为 {"content": "..."}
-// 出错时发送 event: error，data 为 {"message": "..."}
-// 完成后服务端持久化 user + assistant 两条历史，event: done 结束
+// 客户端用 fetch-stream 消费，每条 SSE data 为 {"content": "..."}；出错发 event: error；完成发 event: done。
+// 上游协议差异（OpenAI delta / Anthropic content_block_delta）由 llm.Client.Stream 屏蔽，
+// controller 只负责 SSE 转发 + 客户端断开检测 + 历史持久化。
 func (h *LLMHandler) ChatStream(c *gin.Context) {
 	var cfg model.LLMConfig
 	if err := h.db.First(&cfg).Error; err != nil || !cfg.Enabled {
@@ -136,127 +117,37 @@ func (h *LLMHandler) ChatStream(c *gin.Context) {
 	c.Writer.Header().Set("Connection", "keep-alive")
 	c.Writer.Header().Set("X-Accel-Buffering", "no") // Nginx 关闭缓冲
 	c.Writer.WriteHeader(http.StatusOK)
-	// 注册写入心跳（立即 flush 一个注释行让客户端确认连接建立）
-	c.Writer.Flush()
+	c.Writer.Flush() // 立即 flush 让客户端确认连接建立
 
-	// 组装系统提示词 + 上下文
 	systemPrompt := cfg.SystemPrompt
 	if strings.TrimSpace(systemPrompt) == "" {
 		systemPrompt = defaultSystemPrompt
 	}
 	messages := append([]model.ChatMessage{{Role: "system", Content: systemPrompt}}, req.Messages...)
-	// 构造 OpenAI 兼容流式请求
-	payload := struct {
-		Model       string               `json:"model"`
-		Messages    []model.ChatMessage `json:"messages"`
-		Temperature float64              `json:"temperature"`
-		MaxTokens   int                  `json:"max_tokens,omitempty"`
-		Stream      bool                 `json:"stream"`
-	}{
-		Model:       cfg.Model,
-		Messages:    messages,
-		Temperature: cfg.Temperature,
-		MaxTokens:   cfg.MaxTokens,
-		Stream:      true,
-	}
-	body, _ := json.Marshal(payload)
 
-	endpoint := resolveEndpoint(cfg.Provider, cfg.Endpoint)
-	url := endpoint + "/chat/completions"
-
-	httpReq, err := http.NewRequestWithContext(c.Request.Context(), http.MethodPost, url, bytes.NewReader(body))
-	if err != nil {
-		writeSSEError(c, 500, "构造请求失败: "+err.Error())
-		return
-	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	httpReq.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-	httpReq.Header.Set("Accept", "text/event-stream")
-
-	// 流式请求不能用全局 120s 超时客户端，构造无超时 client 让 ctx 控制 // Logan.Li
-	streamClient := &http.Client{Timeout: 0}
-	resp, err := streamClient.Do(httpReq)
-	if err != nil {
-		writeSSEError(c, 502, "LLM 服务连接失败，请检查网络或 Endpoint 配置: "+err.Error())
-		return
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		raw, _ := io.ReadAll(resp.Body)
-		msg := fmt.Sprintf("LLM 服务返回错误 (HTTP %d)", resp.StatusCode)
-		if e := parseOpenAIError(raw); e != "" {
-			msg += ": " + e
-		}
-		writeSSEError(c, resp.StatusCode, msg)
-		return
-	}
-
-	// 逐行解析上游 SSE，转发增量内容给前端 // Logan.Li
+	// 流式补全：逐 token 回调转发 SSE；ctx 取消（客户端断开）由 llm.Client 内部终止上游读取
 	var contentBuilder strings.Builder
-	reader := bufio.NewReader(resp.Body)
-	clientGone := false
-	for {
-		// 客户端断开时停止读取（c.Request.Context().Done() 在客户端断开时触发） // Logan.Li
-		select {
-		case <-c.Request.Context().Done():
-			clientGone = true
-		default:
-		}
-		if clientGone {
-			break
-		}
-
-		line, readErr := reader.ReadString('\n')
-		line = strings.TrimSpace(line)
-		if line != "" {
-			// 仅处理 data: 开头的行
-			if strings.HasPrefix(line, "data:") {
-				data := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
-				if data == "[DONE]" {
-					// 上游结束，跳出循环准备持久化 + 发送 [DONE]
-					break
-				}
-				// 解析 {"choices":[{"delta":{"content":"..."}}]}
-				var chunk struct {
-					Choices []struct {
-						Delta struct {
-							Content string `json:"content"`
-						} `json:"delta"`
-					} `json:"choices"`
-				}
-				if json.Unmarshal([]byte(data), &chunk) == nil && len(chunk.Choices) > 0 {
-					delta := chunk.Choices[0].Delta.Content
-					if delta != "" {
-						contentBuilder.WriteString(delta)
-						// 转发给前端：data: {"content":"..."}
-						writeSSEChunk(c, delta)
-					}
-				}
-			}
-		}
-		if readErr != nil {
-			if readErr != io.EOF {
-				// 读取出错（非正常结束）
-				if contentBuilder.Len() == 0 {
-					writeSSEError(c, 502, "LLM 流式响应读取失败: "+readErr.Error())
-					return
-				}
-				// 已有部分内容则正常结束
-			}
-			break
-		}
-	}
-
+	_, streamErr := h.buildLLMClient(cfg).Stream(c.Request.Context(), &llm.CompletionRequest{
+		Messages: toLLMMessages(messages),
+	}, func(delta string) {
+		contentBuilder.WriteString(delta)
+		writeSSEChunk(c, delta)
+	})
 	fullContent := contentBuilder.String()
 
-	// 持久化对话历史（仅在成功且非空时入库；错误/中断不入库避免污染历史） // Logan.Li
+	// 出错且无任何内容 → 发 error 事件
+	if streamErr != nil && fullContent == "" {
+		writeSSEError(c, 502, "LLM 流式响应失败: "+streamErr.Error())
+		return
+	}
+
+	// 持久化 + 结束事件（仅在非空且客户端未断开时入库；错误/中断不入库避免污染历史） // Logan.Li
+	clientGone := c.Request.Context().Err() != nil
 	if fullContent != "" && !clientGone {
 		uid := middleware.GetUserID(c)
 		if uid != 0 {
 			lastUser := req.Messages[len(req.Messages)-1]
 			now := time.Now()
-			// session_id：前端传入或后端生成（兼容旧版客户端） // Logan.Li
 			sid := req.SessionID
 			if sid == "" {
 				sid = fmt.Sprintf("sess-%d", now.UnixMilli())
@@ -266,7 +157,6 @@ func (h *LLMHandler) ChatStream(c *gin.Context) {
 				{UserID: uid, SessionID: sid, Role: "assistant", Content: fullContent, Model: cfg.Model, CreatedAt: now.Add(time.Millisecond)},
 			})
 		}
-		// 发送结束标志
 		c.Writer.WriteString("event: done\ndata: {\"model\":\"" + cfg.Model + "\"}\n\n")
 		c.Writer.Flush()
 	} else if !clientGone {
@@ -298,19 +188,6 @@ func writeSSEError(c *gin.Context, code int, message string) {
 	c.Writer.Flush()
 }
 
-// parseOpenAIError 从上游错误响应中提取 message 字段 // Logan.Li
-func parseOpenAIError(raw []byte) string {
-	var m struct {
-		Error struct {
-			Message string `json:"message"`
-		} `json:"error"`
-	}
-	if json.Unmarshal(raw, &m) == nil && m.Error.Message != "" {
-		return m.Error.Message
-	}
-	return ""
-}
-
 // TestLLMConnection 测试 LLM 连接（发送一条 ping 消息） // Logan.Li
 // POST /api/v1/aiops/llm-config/test
 func (h *LLMHandler) TestLLMConnection(c *gin.Context) {
@@ -319,31 +196,12 @@ func (h *LLMHandler) TestLLMConnection(c *gin.Context) {
 		c.JSON(http.StatusOK, model.Response{Code: 400, Message: "请先保存配置"})
 		return
 	}
-	payload := openaiChatRequest{
-		Model:       cfg.Model,
-		Messages:    []model.ChatMessage{{Role: "user", Content: "ping"}},
-		Temperature: 0,
-		MaxTokens:   16,
-	}
-	body, _ := json.Marshal(payload)
-	endpoint := resolveEndpoint(cfg.Provider, cfg.Endpoint)
-	httpReq, _ := http.NewRequestWithContext(c.Request.Context(), http.MethodPost, endpoint+"/chat/completions", bytes.NewReader(body))
-	httpReq.Header.Set("Content-Type", "application/json")
-	httpReq.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-	resp, err := h.client.Do(httpReq)
+	_, err := h.buildLLMClient(cfg).Complete(c.Request.Context(), &llm.CompletionRequest{
+		Messages: []llm.Message{{Role: "user", Content: "ping"}},
+	})
 	if err != nil {
 		c.JSON(http.StatusOK, model.Response{Code: 502, Message: "连接失败: " + err.Error()})
 		return
 	}
-	defer resp.Body.Close()
-	if resp.StatusCode != http.StatusOK {
-		raw, _ := io.ReadAll(resp.Body)
-		msg := fmt.Sprintf("HTTP %d", resp.StatusCode)
-		if e := parseOpenAIError(raw); e != "" {
-			msg += ": " + e
-		}
-		c.JSON(http.StatusOK, model.Response{Code: resp.StatusCode, Message: msg})
-		return
-	}
 	c.JSON(http.StatusOK, model.Response{Code: 200, Message: "连接成功"})
 }
diff --git a/backend/controller/server.go b/backend/controller/server.go
index 6157628..47a0756 100644
--- a/backend/controller/server.go
+++ b/backend/controller/server.go
@@ -223,6 +223,21 @@ func (h *ServerHandler) CreateServer(c *gin.Context) {
 		server.AuthType = "password"
 	}
 
+	// 关联 SSH 凭证：解密后覆盖认证字段（同 AssignCredential，避免单台 CRUD 忽略 credential_id） // Logan.Li
+	if req.CredentialID != nil && *req.CredentialID > 0 {
+		var cred model.Credential
+		if err := h.db.First(&cred, *req.CredentialID).Error; err != nil {
+			c.JSON(http.StatusBadRequest, model.Response{Code: 400, Message: "凭据不存在"})
+			return
+		}
+		plain, err := h.crypt.Decrypt(cred.Content)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, model.Response{Code: 500, Message: "凭据解密失败"})
+			return
+		}
+		applyCredential(cred, plain, &server)
+	}
+
 	if err := h.db.Create(&server).Error; err != nil {
 		c.JSON(http.StatusInternalServerError, model.Response{
 			Code:    500,
@@ -297,6 +312,25 @@ func (h *ServerHandler) UpdateServer(c *gin.Context) {
 	if req.AuthType != "" {
 		updates["auth_type"] = req.AuthType
 	}
+	// 关联 SSH 凭证：解密后覆盖认证字段（applyCredential 填充已 load 的 server，再同步到 updates） // Logan.Li
+	if req.CredentialID != nil && *req.CredentialID > 0 {
+		var cred model.Credential
+		if err := h.db.First(&cred, *req.CredentialID).Error; err != nil {
+			c.JSON(http.StatusBadRequest, model.Response{Code: 400, Message: "凭据不存在"})
+			return
+		}
+		plain, err := h.crypt.Decrypt(cred.Content)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, model.Response{Code: 500, Message: "凭据解密失败"})
+			return
+		}
+		applyCredential(cred, plain, &server)
+		updates["ssh_key"] = server.SSHKey
+		updates["password"] = server.Password
+		updates["auth_type"] = server.AuthType
+		updates["username"] = server.Username
+		updates["port"] = server.Port
+	}
 	if req.OS != "" {
 		updates["os"] = req.OS
 	}
diff --git a/backend/controller/server_batch.go b/backend/controller/server_batch.go
index 88fa65f..c06da7f 100644
--- a/backend/controller/server_batch.go
+++ b/backend/controller/server_batch.go
@@ -30,26 +30,24 @@ func (h *ServerHandler) BatchCreateServers(c *gin.Context) {
 		authType = "password"
 	}
 
-	// 如果指定了凭据，提前解密 // Logan.Li
-	var decryptedPwd, decryptedKey string
+	// 如果指定了凭据，提前解密（plain 在循环内由 applyCredential 解析填充） // Logan.Li
+	var (
+		hasCred   bool
+		cred      model.Credential
+		credPlain string
+	)
 	if req.CredentialID != nil && *req.CredentialID > 0 {
-		var cred model.Credential
 		if err := h.db.First(&cred, *req.CredentialID).Error; err != nil {
 			c.JSON(http.StatusBadRequest, model.Response{Code: 400, Message: "凭据不存在"})
 			return
 		}
-		plain, err := h.crypt.Decrypt(cred.Content)
+		p, err := h.crypt.Decrypt(cred.Content)
 		if err != nil {
 			c.JSON(http.StatusInternalServerError, model.Response{Code: 500, Message: "凭据解密失败"})
 			return
 		}
-		if cred.SecretType == "ssh" {
-			decryptedKey = plain
-			authType = "key"
-		} else {
-			decryptedPwd = plain
-			authType = "password"
-		}
+		credPlain = p
+		hasCred = true
 	}
 
 	result := model.BatchCreateResult{Errors: []string{}}
@@ -84,9 +82,7 @@ func (h *ServerHandler) BatchCreateServers(c *gin.Context) {
 			IP:       ip,
 			Port:     port,
 			Username: username,
-			AuthType: authType,
-			Password: decryptedPwd,
-			SSHKey:   decryptedKey,
+			AuthType: authType, // 无凭证时用请求 authType（默认 password）
 			OS:       item.OS,
 			CPU:      item.CPU,
 			Memory:   item.Memory * 1024, // GB → MB
@@ -95,6 +91,9 @@ func (h *ServerHandler) BatchCreateServers(c *gin.Context) {
 			RegionID: req.RegionID,
 			Status:   "stopped",
 		}
+		if hasCred {
+			applyCredential(cred, credPlain, &server) // 有凭证则覆盖认证字段（ssh→私钥/ssh_user/ssh_port） // Logan.Li
+		}
 		if err := h.db.Create(&server).Error; err != nil {
 			result.Failed++
 			result.Errors = append(result.Errors, fmt.Sprintf("%s(%s): %s", hostname, ip, err.Error()))
diff --git a/backend/model/llm.go b/backend/model/llm.go
index 5f7494c..d3b4f13 100644
--- a/backend/model/llm.go
+++ b/backend/model/llm.go
@@ -3,11 +3,11 @@ package model
 import "time"
 
 // LLMConfig 大语言模型配置（单行配置，类似 LDAPConfig） // Logan.Li
-// 支持 OpenAI 兼容接口（OpenAI / DeepSeek / 通义千问 / Moonshot / Ollama / 自定义）
+// 支持 OpenAI 兼容接口（OpenAI / DeepSeek / 智谱GLM / 通义千问 / Moonshot / Kimi / 小米MiMo / Anthropic(需代理) / Ollama / 自定义）
 type LLMConfig struct {
 	ID           uint      `gorm:"primarykey" json:"id"`
-	Provider     string    `gorm:"size:50;not null;default:openai" json:"provider"`      // openai/deepseek/qwen/moonshot/ollama/custom
-	Model        string    `gorm:"size:100;not null;default:gpt-4o-mini" json:"model"`    // 模型名，如 gpt-4o-mini / deepseek-chat
+	Provider     string    `gorm:"size:50;not null;default:openai" json:"provider"`      // openai/deepseek/glm/qwen/moonshot/kimi/mimo/anthropic/ollama/custom
+	Model        string    `gorm:"size:100;not null;default:gpt-4o-mini" json:"model"`    // 模型名，如 gpt-4o-mini / deepseek-v4-flash
 	APIKey       string    `gorm:"size:500" json:"api_key"`                               // API Key（GET 时会被脱敏）
 	Endpoint     string    `gorm:"size:500" json:"endpoint"`                              // 基础地址，如 https://api.openai.com/v1
 	Temperature  float64   `gorm:"default:0.7" json:"temperature"`                        // 采样温度 0-2
diff --git a/backend/model/server.go b/backend/model/server.go
index acf57be..8e30286 100644
--- a/backend/model/server.go
+++ b/backend/model/server.go
@@ -137,6 +137,7 @@ type CreateServerRequest struct {
 	Password    string  `json:"password"`
 	SSHKey      string  `json:"ssh_key"`
 	AuthType    string  `json:"auth_type" binding:"omitempty,oneof=password key"`
+	CredentialID *uint  `json:"credential_id,omitempty"` // 关联 SSH 凭证，非空则后端解密覆盖认证字段
 	OS          string  `json:"os"`
 	CPU         int     `json:"cpu"`
 	Memory      int64   `json:"memory"`
@@ -158,6 +159,7 @@ type UpdateServerRequest struct {
 	Password    string  `json:"password"`
 	SSHKey      string  `json:"ssh_key"`
 	AuthType    string  `json:"auth_type" binding:"omitempty,oneof=password key"`
+	CredentialID *uint  `json:"credential_id,omitempty"` // 关联 SSH 凭证，非空则后端解密覆盖认证字段
 	OS          string  `json:"os"`
 	CPU         int     `json:"cpu"`
 	Memory      int64   `json:"memory"`
diff --git a/backend/pkg/agentcli/llm/anthropic.go b/backend/pkg/agentcli/llm/anthropic.go
new file mode 100644
index 0000000..396578f
--- /dev/null
+++ b/backend/pkg/agentcli/llm/anthropic.go
@@ -0,0 +1,337 @@
+// anthropic.go - Anthropic Messages 协议 LLM 客户端实现 // Logan.Li
+// 支持官方 Claude（api.anthropic.com）及智谱 GLM 编码套餐（open.bigmodel.cn/api/anthropic）等
+// 所有 Anthropic /v1/messages 兼容端点。与 openai.go 实现同一 Client 接口，由 client.go 按 Protocol 路由。
+// 鉴权统一用 Authorization: Bearer（智谱编码套餐实测通过；官方 Claude 较新版本亦支持）+ anthropic-version。
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// AnthropicClient Anthropic Messages 协议客户端 // Logan.Li
+type AnthropicClient struct {
+	cfg    Config
+	client *http.Client
+}
+
+// NewAnthropicClient 创建 Anthropic 兼容客户端 // Logan.Li
+func NewAnthropicClient(cfg Config) Client {
+	timeout := cfg.Timeout
+	if timeout <= 0 {
+		timeout = 180 * time.Second
+	}
+	return &AnthropicClient{cfg: cfg, client: &http.Client{Timeout: timeout}}
+}
+
+// Name 返回客户端名称 // Logan.Li
+func (c *AnthropicClient) Name() string { return "anthropic(" + c.cfg.Model + ")" }
+
+// ---- 请求体结构（Anthropic Messages 格式） // Logan.Li ----
+
+// anthropicMessage Anthropic 消息：content 可为字符串或 block 数组，用 RawMessage 原样嵌入。
+type anthropicMessage struct {
+	Role    string          `json:"role"`
+	Content json.RawMessage `json:"content"`
+}
+
+// anthropicRequest Anthropic /v1/messages 请求体 // Logan.Li
+type anthropicRequest struct {
+	Model       string             `json:"model"`
+	MaxTokens   int                `json:"max_tokens"`
+	System      string             `json:"system,omitempty"` // system 抽到顶层（不在 messages 内）
+	Messages    []anthropicMessage `json:"messages"`
+	Tools       []anthropicTool    `json:"tools,omitempty"`
+	Temperature float64            `json:"temperature,omitempty"`
+	Stream      bool               `json:"stream,omitempty"`
+}
+
+// anthropicTool Anthropic 工具定义（input_schema，对应 OpenAI 的 parameters） // Logan.Li
+type anthropicTool struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	InputSchema map[string]interface{} `json:"input_schema"`
+}
+
+// 响应 content block：text 或 tool_use（互斥字段） // Logan.Li
+type anthropicBlock struct {
+	Type  string          `json:"type"` // text | tool_use
+	Text  string          `json:"text,omitempty"`
+	ID    string          `json:"id,omitempty"`    // tool_use
+	Name  string          `json:"name,omitempty"`  // tool_use
+	Input json.RawMessage `json:"input,omitempty"` // tool_use
+}
+
+// 请求侧 content block（assistant 输出 tool_use / user 输出 tool_result） // Logan.Li
+type textBlock struct {
+	Type string `json:"type"` // 固定 "text"
+	Text string `json:"text"`
+}
+type toolUseBlock struct {
+	Type  string          `json:"type"` // 固定 "tool_use"
+	ID    string          `json:"id"`
+	Name  string          `json:"name"`
+	Input json.RawMessage `json:"input"`
+}
+type toolResultBlock struct {
+	Type      string `json:"type"` // 固定 "tool_result"
+	ToolUseID string `json:"tool_use_id"`
+	Content   string `json:"content"`
+}
+
+// anthropicResponse Anthropic /v1/messages 响应体 // Logan.Li
+type anthropicResponse struct {
+	ID      string           `json:"id"`
+	Model   string           `json:"model"`
+	Content []anthropicBlock `json:"content"`
+	Usage   struct {
+		InputTokens  int `json:"input_tokens"`
+		OutputTokens int `json:"output_tokens"`
+	} `json:"usage"`
+	Error *struct {
+		Type    string `json:"type"`
+		Message string `json:"message"`
+	} `json:"error,omitempty"`
+}
+
+// buildMessages 把 OpenAI 风格 Message[] 转成 Anthropic messages + 顶层 system // Logan.Li
+// - system role 抽到顶层 system 字段
+// - tool role（工具结果）→ user 消息的 tool_result 块
+// - assistant 带 tool_calls → content 含 tool_use 块
+func (c *AnthropicClient) buildMessages(req *CompletionRequest) (string, []anthropicMessage) {
+	var system string
+	var msgs []anthropicMessage
+	for _, m := range req.Messages {
+		switch m.Role {
+		case "system":
+			if system == "" {
+				system = m.Content
+			} else {
+				system += "\n" + m.Content
+			}
+		case "tool":
+			raw, _ := json.Marshal([]toolResultBlock{{
+				Type: "tool_result", ToolUseID: m.ToolCallID, Content: m.Content,
+			}})
+			msgs = append(msgs, anthropicMessage{Role: "user", Content: raw})
+		case "assistant":
+			// assistant 消息可能含文本 + tool_use 计划，组装成 content block 数组
+			blocks := make([]interface{}, 0, 1+len(m.ToolCalls))
+			if m.Content != "" {
+				blocks = append(blocks, textBlock{Type: "text", Text: m.Content})
+			}
+			for _, tc := range m.ToolCalls {
+				args := json.RawMessage("{}")
+				if strings.TrimSpace(tc.Function.Arguments) != "" {
+					args = json.RawMessage(tc.Function.Arguments)
+				}
+				blocks = append(blocks, toolUseBlock{
+					Type: "tool_use", ID: tc.ID, Name: tc.Function.Name, Input: args,
+				})
+			}
+			raw, _ := json.Marshal(blocks)
+			msgs = append(msgs, anthropicMessage{Role: "assistant", Content: raw})
+		default: // user
+			raw, _ := json.Marshal(m.Content)
+			msgs = append(msgs, anthropicMessage{Role: "user", Content: raw})
+		}
+	}
+	return system, msgs
+}
+
+// buildTools OpenAI Tool[](parameters) → Anthropic [](input_schema) // Logan.Li
+func buildTools(tools []Tool) []anthropicTool {
+	out := make([]anthropicTool, 0, len(tools))
+	for _, t := range tools {
+		out = append(out, anthropicTool{
+			Name: t.Function.Name, Description: t.Function.Description, InputSchema: t.Function.Parameters,
+		})
+	}
+	return out
+}
+
+// buildRequest 构造 Anthropic 请求体（Complete 与 Stream 共用） // Logan.Li
+func (c *AnthropicClient) buildRequest(req *CompletionRequest, stream bool) ([]byte, error) {
+	system, msgs := c.buildMessages(req)
+	payload := anthropicRequest{
+		Model: c.cfg.Model, MaxTokens: c.cfg.MaxTokens, System: system,
+		Messages: msgs, Tools: buildTools(req.Tools),
+		Temperature: c.cfg.Temperature, Stream: stream,
+	}
+	if payload.MaxTokens <= 0 {
+		payload.MaxTokens = 4096
+	}
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return nil, fmt.Errorf("序列化请求失败: %w", err)
+	}
+	return body, nil
+}
+
+// newHTTPRequest 构造 POST /v1/messages 请求（Bearer + anthropic-version） // Logan.Li
+func (c *AnthropicClient) newHTTPRequest(ctx context.Context, body []byte) (*http.Request, error) {
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, c.cfg.Endpoint+"/v1/messages", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("构造请求失败: %w", err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+c.cfg.APIKey)
+	httpReq.Header.Set("anthropic-version", "2023-06-01")
+	return httpReq, nil
+}
+
+// Complete 一次性补全（Anthropic /v1/messages） // Logan.Li
+// 响应 content blocks：text 拼为 Content，tool_use 转为 ToolCalls；usage 映射。
+func (c *AnthropicClient) Complete(ctx context.Context, req *CompletionRequest) (*CompletionResponse, error) {
+	body, err := c.buildRequest(req, false)
+	if err != nil {
+		return nil, err
+	}
+	httpReq, err := c.newHTTPRequest(ctx, body)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := c.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("LLM 连接失败: %w", err)
+	}
+	defer resp.Body.Close()
+	raw, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("LLM 返回错误 (HTTP %d): %s", resp.StatusCode, extractErrorMessage(raw))
+	}
+	var parsed anthropicResponse
+	if err := json.Unmarshal(raw, &parsed); err != nil {
+		return nil, fmt.Errorf("解析 LLM 响应失败: %w", err)
+	}
+	if parsed.Error != nil && parsed.Error.Message != "" {
+		return nil, fmt.Errorf("LLM 错误: %s", parsed.Error.Message)
+	}
+
+	var textBuf strings.Builder
+	var toolCalls []ToolCall
+	for _, b := range parsed.Content {
+		switch b.Type {
+		case "text":
+			textBuf.WriteString(b.Text)
+		case "tool_use":
+			args := string(b.Input)
+			if strings.TrimSpace(args) == "" {
+				args = "{}"
+			}
+			toolCalls = append(toolCalls, ToolCall{
+				ID: b.ID, Type: "function",
+				Function: FunctionCall{Name: b.Name, Arguments: args},
+			})
+		}
+	}
+	return &CompletionResponse{
+		ID:    parsed.ID,
+		Model: parsed.Model,
+		Message: Message{
+			Role:      "assistant",
+			Content:   textBuf.String(),
+			ToolCalls: toolCalls,
+		},
+		Usage: Usage{
+			PromptTokens:     parsed.Usage.InputTokens,
+			CompletionTokens: parsed.Usage.OutputTokens,
+			TotalTokens:      parsed.Usage.InputTokens + parsed.Usage.OutputTokens,
+		},
+	}, nil
+}
+
+// Stream 流式补全（Anthropic SSE） // Logan.Li
+// 逐 token 回调 text_delta；usage 从 message_start(input) + message_delta(output) 取。
+// 注：流式主要用于纯对话（智能助手），tool_use 流式增量在此不组装（Agent CLI 走 Complete）。
+func (c *AnthropicClient) Stream(ctx context.Context, req *CompletionRequest, onDelta func(string)) (*CompletionResponse, error) {
+	body, err := c.buildRequest(req, true)
+	if err != nil {
+		return nil, err
+	}
+	httpReq, err := c.newHTTPRequest(ctx, body)
+	if err != nil {
+		return nil, err
+	}
+	httpReq.Header.Set("Accept", "text/event-stream")
+
+	streamClient := &http.Client{Timeout: 0} // 流式由 ctx 控制超时
+	resp, err := streamClient.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("LLM 连接失败: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		raw, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("LLM 返回错误 (HTTP %d): %s", resp.StatusCode, extractErrorMessage(raw))
+	}
+
+	var contentBuilder strings.Builder
+	var usage Usage
+	var eventType string
+	reader := bufio.NewReader(resp.Body)
+	for {
+		line, readErr := reader.ReadString('\n')
+		trimmed := strings.TrimSpace(line)
+		switch {
+		case strings.HasPrefix(trimmed, "event:"):
+			eventType = strings.TrimSpace(strings.TrimPrefix(trimmed, "event:"))
+		case strings.HasPrefix(trimmed, "data:"):
+			data := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
+			if data != "" && data != "[DONE]" {
+				switch eventType {
+				case "content_block_delta":
+					var d struct {
+						Delta struct {
+							Type string `json:"type"`
+							Text string `json:"text"`
+						} `json:"delta"`
+					}
+					if json.Unmarshal([]byte(data), &d) == nil && d.Delta.Type == "text_delta" && d.Delta.Text != "" {
+						contentBuilder.WriteString(d.Delta.Text)
+						if onDelta != nil {
+							onDelta(d.Delta.Text)
+						}
+					}
+				case "message_start":
+					var d struct {
+						Message struct {
+							Usage struct {
+								InputTokens int `json:"input_tokens"`
+							} `json:"usage"`
+						} `json:"message"`
+					}
+					if json.Unmarshal([]byte(data), &d) == nil {
+						usage.PromptTokens = d.Message.Usage.InputTokens
+					}
+				case "message_delta":
+					var d struct {
+						Usage struct {
+							OutputTokens int `json:"output_tokens"`
+						} `json:"usage"`
+					}
+					if json.Unmarshal([]byte(data), &d) == nil {
+						usage.CompletionTokens = d.Usage.OutputTokens
+						usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+					}
+				}
+			}
+			eventType = "" // 消费完 data 重置，防默认事件串用
+		}
+		if readErr != nil {
+			break
+		}
+	}
+	return &CompletionResponse{
+		Model:  c.cfg.Model,
+		Message: Message{Role: "assistant", Content: contentBuilder.String()},
+		Usage:  usage,
+	}, nil
+}
diff --git a/backend/pkg/agentcli/llm/client.go b/backend/pkg/agentcli/llm/client.go
new file mode 100644
index 0000000..0d853ad
--- /dev/null
+++ b/backend/pkg/agentcli/llm/client.go
@@ -0,0 +1,14 @@
+// client.go - LLM 客户端工厂（按 Config.Protocol 路由到具体实现） // Logan.Li
+// 上层（controller/agentcli）统一调 NewClient，无需感知协议差异。
+// 新增协议只需在此分支 + 实现对应 Client。
+package llm
+
+// NewClient 按 Config.Protocol 选择 LLM 客户端实现 // Logan.Li
+// "anthropic" → Anthropic Messages 协议（Claude 官方 / 智谱 GLM 编码套餐等）；
+// 其余（含空值）→ OpenAI 兼容协议。
+func NewClient(cfg Config) Client {
+	if cfg.Protocol == "anthropic" {
+		return NewAnthropicClient(cfg)
+	}
+	return NewOpenAIClient(cfg)
+}
diff --git a/backend/pkg/agentcli/llm/openai.go b/backend/pkg/agentcli/llm/openai.go
index 4ca038d..ed47b14 100644
--- a/backend/pkg/agentcli/llm/openai.go
+++ b/backend/pkg/agentcli/llm/openai.go
@@ -4,17 +4,20 @@
 package llm
 
 import (
+	"bufio"
 	"bytes"
 	"context"
 	"encoding/json"
 	"fmt"
 	"io"
 	"net/http"
+	"strings"
 	"time"
 )
 
 // Config LLM 客户端配置（endpoint 已由上层解析为完整 base url） // Logan.Li
 type Config struct {
+	Protocol    string        // "openai" | "anthropic"，由 client.go 工厂决定具体实现
 	Endpoint    string        // 完整 base url，如 https://api.deepseek.com/v1
 	APIKey      string        // Bearer token
 	Model       string        // 模型名，如 deepseek-chat / gpt-4o-mini
@@ -29,8 +32,9 @@ type OpenAIClient struct {
 	client *http.Client
 }
 
-// NewClient 创建 OpenAI 兼容客户端 // Logan.Li
-func NewClient(cfg Config) Client {
+// NewOpenAIClient 创建 OpenAI 兼容客户端 // Logan.Li
+// 顶层工厂在 client.go（按 Protocol 路由到 OpenAI / Anthropic 实现）。
+func NewOpenAIClient(cfg Config) Client {
 	timeout := cfg.Timeout
 	if timeout <= 0 {
 		timeout = 180 * time.Second
@@ -52,6 +56,7 @@ type openaiRequest struct {
 	ToolChoice  string    `json:"tool_choice,omitempty"` // "auto" 让模型自主决定
 	Temperature float64   `json:"temperature,omitempty"`
 	MaxTokens   int       `json:"max_tokens,omitempty"`
+	Stream      bool      `json:"stream,omitempty"` // 流式补全（Stream 方法置 true）
 }
 
 // openaiResponse OpenAI chat/completions 响应体 // Logan.Li
@@ -131,6 +136,88 @@ func (c *OpenAIClient) Complete(ctx context.Context, req *CompletionRequest) (*C
 	}, nil
 }
 
+// Stream 流式补全（OpenAI 兼容 SSE），逐 token 回调 onDelta，返回聚合响应 // Logan.Li
+// 解析 data: {"choices":[{"delta":{"content":"..."}}]} 增量；[DONE] 结束；末块 usage 若有则取。
+func (c *OpenAIClient) Stream(ctx context.Context, req *CompletionRequest, onDelta func(string)) (*CompletionResponse, error) {
+	payload := openaiRequest{
+		Model:       c.cfg.Model,
+		Messages:    req.Messages,
+		Tools:       req.Tools,
+		ToolChoice:  "auto",
+		Temperature: c.cfg.Temperature,
+		MaxTokens:   c.cfg.MaxTokens,
+		Stream:      true,
+	}
+	if payload.MaxTokens <= 0 {
+		payload.MaxTokens = 4096
+	}
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return nil, fmt.Errorf("序列化请求失败: %w", err)
+	}
+	url := c.cfg.Endpoint + "/chat/completions"
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("构造请求失败: %w", err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+c.cfg.APIKey)
+	httpReq.Header.Set("Accept", "text/event-stream")
+
+	streamClient := &http.Client{Timeout: 0} // 流式由 ctx 控制超时
+	resp, err := streamClient.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("LLM 连接失败: %w", err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		raw, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("LLM 返回错误 (HTTP %d): %s", resp.StatusCode, extractErrorMessage(raw))
+	}
+
+	var contentBuilder strings.Builder
+	var usage Usage
+	reader := bufio.NewReader(resp.Body)
+	for {
+		line, readErr := reader.ReadString('\n')
+		line = strings.TrimSpace(line)
+		if line != "" && strings.HasPrefix(line, "data:") {
+			data := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
+			if data == "[DONE]" {
+				break
+			}
+			var chunk struct {
+				Choices []struct {
+					Delta struct {
+						Content string `json:"content"`
+					} `json:"delta"`
+				} `json:"choices"`
+				Usage *Usage `json:"usage"`
+			}
+			if json.Unmarshal([]byte(data), &chunk) == nil {
+				if len(chunk.Choices) > 0 && chunk.Choices[0].Delta.Content != "" {
+					delta := chunk.Choices[0].Delta.Content
+					contentBuilder.WriteString(delta)
+					if onDelta != nil {
+						onDelta(delta)
+					}
+				}
+				if chunk.Usage != nil {
+					usage = *chunk.Usage
+				}
+			}
+		}
+		if readErr != nil {
+			break
+		}
+	}
+	return &CompletionResponse{
+		Model:  c.cfg.Model,
+		Message: Message{Role: "assistant", Content: contentBuilder.String()},
+		Usage:  usage,
+	}, nil
+}
+
 // extractErrorMessage 从上游错误响应提取 message 字段 // Logan.Li
 func extractErrorMessage(raw []byte) string {
 	var m struct {
diff --git a/backend/pkg/agentcli/llm/types.go b/backend/pkg/agentcli/llm/types.go
index 4aec90a..a9f90d8 100644
--- a/backend/pkg/agentcli/llm/types.go
+++ b/backend/pkg/agentcli/llm/types.go
@@ -65,9 +65,13 @@ type Usage struct {
 }
 
 // Client LLM 客户端接口 // Logan.Li
+// 实现按协议分两种：OpenAI 兼容（openai.go）与 Anthropic 兼容（anthropic.go），
+// 由 client.go 工厂按 Config.Protocol 路由。上层（controller/agentcli）只面向本接口。
 type Client interface {
 	// Complete 一次性补全（返回思考文本 + 工具调用计划）。
 	Complete(ctx context.Context, req *CompletionRequest) (*CompletionResponse, error)
+	// Stream 流式补全，逐 token 回调 onDelta；返回最终聚合响应（含完整 content + usage）。
+	Stream(ctx context.Context, req *CompletionRequest, onDelta func(string)) (*CompletionResponse, error)
 	// Name 返回客户端名称（用于日志/审计）。
 	Name() string
 }
diff --git a/frontend/src/components/aiops/AIOpsAssistant.tsx b/frontend/src/components/aiops/AIOpsAssistant.tsx
index 1f44a33..dd6d380 100644
--- a/frontend/src/components/aiops/AIOpsAssistant.tsx
+++ b/frontend/src/components/aiops/AIOpsAssistant.tsx
@@ -27,16 +27,6 @@ import {
 } from 'lucide-react';
 import type { JsonData } from '@/types/common';
 
-// LLM 提供商列表（OpenAI 兼容接口） // Logan.Li
-const LLM_PROVIDERS: { id: string; name: string; defaultEndpoint: string; models: string[] }[] = [
-  { id: 'openai', name: 'OpenAI', defaultEndpoint: 'https://api.openai.com/v1', models: ['gpt-4o-mini', 'gpt-4o', 'gpt-4-turbo'] },
-  { id: 'deepseek', name: 'DeepSeek', defaultEndpoint: 'https://api.deepseek.com/v1', models: ['deepseek-chat', 'deepseek-reasoner'] },
-  { id: 'qwen', name: '通义千问 (DashScope)', defaultEndpoint: 'https://dashscope.aliyuncs.com/compatible-mode/v1', models: ['qwen-plus', 'qwen-turbo', 'qwen-max'] },
-  { id: 'moonshot', name: 'Moonshot 月之暗面', defaultEndpoint: 'https://api.moonshot.cn/v1', models: ['moonshot-v1-8k', 'moonshot-v1-32k'] },
-  { id: 'ollama', name: 'Ollama 本地', defaultEndpoint: 'http://localhost:11434/v1', models: ['qwen2.5:7b', 'llama3.2', 'deepseek-r1:7b'] },
-  { id: 'custom', name: '自定义 (OpenAI 兼容)', defaultEndpoint: '', models: [] },
-];
-
 const defaultConfig: LLMConfigData = {
   provider: 'openai',
   model: 'gpt-4o-mini',
@@ -323,15 +313,6 @@ export function AIOpsAssistant() {
               {config.model}
             </Badge>
           ) : null}
-          <Button
-            variant="outline"
-            size="sm"
-            className="border-slate-700 text-slate-300 hover:text-purple-400"
-            onClick={() => navigate('/ai-llm-config')}
-          >
-            <Settings className="h-4 w-4 mr-1" />
-            模型配置
-          </Button>
         </div>
       </div>
 
diff --git a/frontend/src/components/aiops/LLMConfig.tsx b/frontend/src/components/aiops/LLMConfig.tsx
index 62499a9..cfd942c 100644
--- a/frontend/src/components/aiops/LLMConfig.tsx
+++ b/frontend/src/components/aiops/LLMConfig.tsx
@@ -13,19 +13,27 @@ import { toast } from 'sonner';
 import { Loader2, Save, Wifi, KeyRound, Brain } from 'lucide-react';
 import { aiApi, type LLMConfigData } from '@/api';
 
-// OpenAI 兼容提供商列表 // Logan.Li
+// 提供商列表（后端按 provider 自动路由协议） // Logan.Li
+// Anthropic 协议（glm-coding / anthropic）：端点不含 /v1，后端拼 /v1/messages。
+// OpenAI 协议（其余）：端点含 /v1，后端拼 /chat/completions。
 const PROVIDERS = [
-  { id: 'openai', name: 'OpenAI', endpoint: 'https://api.openai.com/v1', models: 'gpt-4o-mini, gpt-4o' },
-  { id: 'deepseek', name: 'DeepSeek', endpoint: 'https://api.deepseek.com/v1', models: 'deepseek-chat, deepseek-reasoner' },
+  { id: 'openai', name: 'OpenAI', endpoint: 'https://api.openai.com/v1', models: 'gpt-4o-mini, gpt-4o, gpt-4.1' },
+  { id: 'deepseek', name: 'DeepSeek', endpoint: 'https://api.deepseek.com/v1', models: 'deepseek-v4-flash, deepseek-v4-pro' },
+  { id: 'glm', name: '智谱 GLM (通用)', endpoint: 'https://open.bigmodel.cn/api/paas/v4', models: 'glm-4.6, glm-4.5, glm-4-flash' },
+  { id: 'glm-coding', name: '智谱 GLM (编码套餐)', endpoint: 'https://open.bigmodel.cn/api/anthropic', models: 'GLM-5.2' },
   { id: 'qwen', name: '通义千问 (DashScope)', endpoint: 'https://dashscope.aliyuncs.com/compatible-mode/v1', models: 'qwen-plus, qwen-turbo, qwen-max' },
-  { id: 'moonshot', name: 'Moonshot 月之暗面', endpoint: 'https://api.moonshot.cn/v1', models: 'moonshot-v1-8k, moonshot-v1-32k' },
+  { id: 'moonshot', name: 'Moonshot (Kimi)', endpoint: 'https://api.moonshot.cn/v1', models: 'kimi-latest, moonshot-v1-8k, moonshot-v1-32k' },
+  { id: 'kimi', name: 'Kimi 编码套餐', endpoint: 'https://api.kimi.com/coding/v1', models: 'kimi-for-coding' },
+  { id: 'mimo', name: '小米 MiMo', endpoint: 'https://api.mimo.mi.com/v1', models: 'mimo-v2-flash, mimo-v2' },
+  { id: 'anthropic', name: 'Anthropic Claude', endpoint: 'https://api.anthropic.com', models: 'claude-sonnet-4-6, claude-opus-4-8, claude-haiku-4-5' },
   { id: 'ollama', name: 'Ollama 本地', endpoint: 'http://localhost:11434/v1', models: 'qwen2.5:7b, llama3.2' },
   { id: 'custom', name: '自定义 (OpenAI 兼容)', endpoint: '', models: '' },
 ];
 
+// 生成参数（温度/max_tokens/系统提示词）已按 Claude Code 模式精简：UI 不再暴露，走内置默认。 // Logan.Li
 const defaultCfg: LLMConfigData = {
-  provider: 'deepseek', model: 'deepseek-chat', api_key: '',
-  endpoint: 'https://api.deepseek.com/v1', temperature: 0.3, max_tokens: 4096,
+  provider: 'deepseek', model: 'deepseek-v4-flash', api_key: '',
+  endpoint: 'https://api.deepseek.com/v1', temperature: 0.7, max_tokens: 8192,
   system_prompt: '', enabled: false,
 };
 
@@ -105,36 +113,11 @@ export default function LLMConfigPage() {
         </CardContent>
       </Card>
 
-      <Card className="bg-slate-900/60 border-slate-800">
-        <CardHeader><CardTitle className="text-slate-200 text-base">生成参数</CardTitle></CardHeader>
-        <CardContent className="space-y-4">
-          <div className="grid grid-cols-2 gap-4">
-            <div className="space-y-1.5">
-              <Label className="text-slate-300">温度（Temperature）</Label>
-              <Input type="number" step="0.1" min="0" max="2" value={cfg.temperature}
-                onChange={(e) => set('temperature', Number(e.target.value))}
-                className="bg-slate-800 border-slate-700 text-slate-200" />
-            </div>
-            <div className="space-y-1.5">
-              <Label className="text-slate-300">最大输出 Token</Label>
-              <Input type="number" min="256" value={cfg.max_tokens}
-                onChange={(e) => set('max_tokens', Number(e.target.value))}
-                className="bg-slate-800 border-slate-700 text-slate-200" />
-            </div>
-          </div>
-          <div className="space-y-1.5">
-            <Label className="text-slate-300">系统提示词（可选，留空使用内置运维助手角色）</Label>
-            <textarea value={cfg.system_prompt} onChange={(e) => set('system_prompt', e.target.value)} rows={3}
-              className="w-full bg-slate-800 border border-slate-700 rounded-md px-3 py-2 text-sm text-slate-200 resize-none" />
-          </div>
-          <div className="flex items-center justify-between">
-            <Label className="text-slate-300">启用 LLM 服务</Label>
-            <Switch checked={cfg.enabled} onCheckedChange={(v) => set('enabled', v)} />
-          </div>
-        </CardContent>
-      </Card>
-
       <div className="flex items-center gap-2">
+        <div className="flex items-center gap-2 text-sm text-slate-300">
+          <Switch checked={cfg.enabled} onCheckedChange={(v) => set('enabled', v)} />
+          <span>启用 LLM 服务</span>
+        </div>
         <Button onClick={save} disabled={saving} className="bg-blue-600 hover:bg-blue-700">
           {saving ? <Loader2 className="w-4 h-4 mr-1 animate-spin" /> : <Save className="w-4 h-4 mr-1" />} 保存配置
         </Button>
diff --git a/frontend/src/components/aiops/PredictiveCapacityTab.tsx b/frontend/src/components/aiops/PredictiveCapacityTab.tsx
index 1098a15..1fc8288 100644
--- a/frontend/src/components/aiops/PredictiveCapacityTab.tsx
+++ b/frontend/src/components/aiops/PredictiveCapacityTab.tsx
@@ -1,7 +1,11 @@
-import { HardDrive, Database, Zap, Loader2 } from 'lucide-react';
+import { useState } from 'react';
+import { HardDrive, Database, Zap, Loader2, Sparkles } from 'lucide-react';
 import { Badge } from '@/components/ui/badge';
+import { Button } from '@/components/ui/button';
 import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
-import type { CapacityForecast } from './predictUtils';
+import { toast } from 'sonner';
+import { aiApi } from '@/api';
+import { buildCapacityInsightPrompt, type CapacityForecast } from './predictUtils';
 
 export function PredictiveCapacityTab({
   forecasts,
@@ -10,6 +14,25 @@ export function PredictiveCapacityTab({
   forecasts: CapacityForecast[];
   loading: boolean;
 }) {
+  const [insight, setInsight] = useState('');
+  const [analyzing, setAnalyzing] = useState(false);
+
+  // AI 深度解读：把容量预测硬数据喂给 LLM，生成综合风险评估与运维建议 // Logan.Li
+  // 复用 aiApi.chat（双协议后自动支持 GLM-5.2 等）；失败仅 toast，下方逐条模板建议仍可读。
+  const analyze = async () => {
+    const prompt = buildCapacityInsightPrompt(forecasts);
+    if (!prompt) return;
+    setAnalyzing(true);
+    try {
+      const res = await aiApi.chat([{ role: 'user', content: prompt }]);
+      setInsight(res?.content || '');
+    } catch {
+      toast.error('AI 解读失败，请检查大模型配置或稍后重试');
+    } finally {
+      setAnalyzing(false);
+    }
+  };
+
   return (
     <Card className="bg-slate-800 border-slate-700">
       <CardHeader className="pb-3">
@@ -18,7 +41,22 @@ export function PredictiveCapacityTab({
             <HardDrive className="h-5 w-5 text-amber-400" />
             容量规划预警
           </CardTitle>
-          <span className="text-xs text-slate-500">基于日均增长率推算阈值触及时间</span>
+          {!loading && forecasts.length > 0 && (
+            <Button
+              variant="outline"
+              size="sm"
+              onClick={analyze}
+              disabled={analyzing}
+              className="border-purple-500/30 text-purple-300 hover:text-purple-200 h-7"
+            >
+              {analyzing ? (
+                <Loader2 className="h-3.5 w-3.5 mr-1 animate-spin" />
+              ) : (
+                <Sparkles className="h-3.5 w-3.5 mr-1" />
+              )}
+              AI 深度解读
+            </Button>
+          )}
         </div>
       </CardHeader>
       <CardContent>
@@ -30,6 +68,16 @@ export function PredictiveCapacityTab({
           <div className="text-center py-12 text-slate-500">暂无容量数据</div>
         ) : (
           <div className="space-y-4">
+            {/* AI 深度解读结果（点击按钮后展示） */}
+            {insight && (
+              <div className="p-4 rounded-lg border border-purple-500/30 bg-purple-500/5">
+                <div className="flex items-center gap-2 mb-2">
+                  <Sparkles className="h-4 w-4 text-purple-400" />
+                  <span className="text-sm font-medium text-purple-300">AI 深度解读</span>
+                </div>
+                <p className="text-sm text-slate-300 whitespace-pre-wrap break-words">{insight}</p>
+              </div>
+            )}
             {forecasts.map((forecast, idx) => (
               <div key={idx} className="p-4 rounded-lg border border-slate-700 bg-slate-800/50">
                 <div className="flex items-center justify-between mb-3">
diff --git a/frontend/src/components/aiops/predictUtils.ts b/frontend/src/components/aiops/predictUtils.ts
index 31bb85f..deb98a2 100644
--- a/frontend/src/components/aiops/predictUtils.ts
+++ b/frontend/src/components/aiops/predictUtils.ts
@@ -319,3 +319,26 @@ export function trendColorClass(trend: string): string {
       return 'text-slate-400';
   }
 }
+
+/**
+ * 构建容量预测「AI 深度解读」的 prompt —— 把硬数据序列化为结构化上下文喂给 LLM。 // Logan.Li
+ * 纯函数，供 PredictiveCapacityTab 调 aiApi.chat 使用。无数据返回空串（调用方据此禁用按钮）。
+ */
+export function buildCapacityInsightPrompt(forecasts: CapacityForecast[]): string {
+  if (forecasts.length === 0) return '';
+  const lines = forecasts.map((f) => {
+    const risk = f.daysUntilThreshold >= 999 ? '增长平缓，暂无触顶风险' : `${f.daysUntilThreshold} 天后触及阈值`;
+    return `- ${f.resource}：当前 ${f.currentUsage}% → 7天后预测 ${f.predictedUsage}%（阈值 ${f.threshold}%），${risk}`;
+  });
+  return [
+    '你是 AIOps 容量规划专家。基于以下服务器资源容量预测数据，给出综合风险评估与可执行的运维建议：',
+    '',
+    '【容量预测数据】',
+    ...lines,
+    '',
+    '要求：',
+    '1. 指出最紧急的风险资源（触及阈值最快者优先）',
+    '2. 给出针对性处置建议（扩容 / 清理 / 优化，具体到操作命令或步骤）',
+    '3. 简洁中文，Markdown 无序列表，不超过 200 字',
+  ].join('\n');
+}