diff --git a/docs/llm-debug-guide.md b/docs/llm-debug-guide.md
new file mode 100644
index 000000000..e91d87d67
--- /dev/null
+++ b/docs/llm-debug-guide.md
@@ -0,0 +1,300 @@
+# LLM Debug 日志指南
+
+## 概述
+
+本文档说明如何使用新增的 LLM 调试日志来排查 API 调用问题，特别是 `404 page not found` 错误。
+
+## 新增的调试日志
+
+### 1. Provider 创建日志 (http_provider.go)
+
+**位置**: `CreateProvider` 函数
+
+**日志级别**:
+- `DebugCF`: 创建 provider 时的详细信息
+- `InfoCF`: Provider 创建成功
+- `ErrorCF`: 配置错误
+
+**记录信息**:
+```go
+// 创建时的调试信息
+logger.DebugCF("llm", "Creating LLM provider", map[string]interface{}{
+    "model": model,
+    "lower_model": lowerModel,
+})
+
+// 创建成功
+logger.InfoCF("llm", "Provider created successfully", map[string]interface{}{
+    "model": model,
+    "api_base": apiBase,
+    "has_api_key": apiKey != "",
+    "api_key_len": len(apiKey),
+})
+```
+
+### 2. HTTP 请求日志
+
+**位置**: `HTTPProvider.Chat` 函数
+
+**发送请求前的日志**:
+```go
+logger.DebugCF("llm", "Sending LLM request", map[string]interface{}{
+    "url": fullURL,
+    "model": model,
+    "api_base": p.apiBase,
+    "has_api_key": p.apiKey != "",
+    "api_key_len": len(p.apiKey),
+    "message_count": len(messages),
+    "tools_count": len(tools),
+    "request_body": string(jsonData),  // 完整请求体
+})
+```
+
+**发送请求时的日志**:
+```go
+logger.DebugCF("llm", "Sending HTTP request", map[string]interface{}{
+    "method": "POST",
+    "url": fullURL,
+    "headers": req.Header,  // 包括 Authorization 等 header
+})
+```
+
+### 3. HTTP 响应日志
+
+**成功接收响应**:
+```go
+logger.DebugCF("llm", "Received LLM response", map[string]interface{}{
+    "status_code": resp.StatusCode,
+    "status": resp.Status,
+    "content_length": len(body),
+    "response_body": string(body),  // 完整响应体
+})
+```
+
+**错误响应 (非 200 状态码)**:
+```go
+logger.ErrorCF("llm", "LLM API returned non-OK status", map[string]interface{}{
+    "status_code": resp.StatusCode,
+    "status": resp.Status,
+    "url": fullURL,
+    "response_body": string(body),
+})
+```
+
+## 如何启用调试日志
+
+### 方法 1: 通过配置文件
+
+修改 `config.yaml`:
+
+```yaml
+logging:
+  level: debug  # 设置为 debug 级别以查看所有日志
+  category_levels:
+    llm: debug    # 只启用 llm 相关的 debug 日志
+    agent: info   # 其他类别保持 info 级别
+```
+
+### 方法 2: 通过环境变量
+
+```bash
+export PICOCLAW_LOG_LEVEL=debug
+```
+
+或者只针对 llm 分类：
+
+```bash
+export PICOCLAW_LOG_CATEGORY_LLM=debug
+```
+
+## 排查 404 错误的步骤
+
+当遇到 `Error processing message: LLM call failed: API error: 404 page not found` 错误时：
+
+### 步骤 1: 检查 Provider 创建日志
+
+查找日志中的 "Provider created successfully" 消息：
+
+```
+[INFO] [llm] Provider created successfully
+  model: gpt-4
+  api_base: https://api.openai.com/v1
+  has_api_key: true
+  api_key_len: 51
+```
+
+**检查点**:
+- ✅ `api_base` 是否正确？（常见错误：末尾多了 `/chat/completions`）
+- ✅ `has_api_key` 是否为 true？
+- ✅ `model` 名称是否正确？
+
+### 步骤 2: 检查请求 URL
+
+查找 "Sending LLM request" 日志：
+
+```
+[DEBUG] [llm] Sending LLM request
+  url: https://api.openai.com/v1/chat/completions
+  model: gpt-4
+  api_base: https://api.openai.com/v1
+  ...
+```
+
+**检查点**:
+- ✅ 完整的 `url` 是否正确？
+- ✅ 路径是否为 `/chat/completions`？
+- ✅ 是否有重复的路径（如 `/v1/v1/chat/completions`）？
+
+### 步骤 3: 检查响应详情
+
+查找 "LLM API returned non-OK status" 错误日志：
+
+```
+[ERROR] [llm] LLM API returned non-OK status
+  status_code: 404
+  status: 404 Not Found
+  url: https://wrong-url.com/v1/chat/completions
+  response_body: 404 page not found
+```
+
+**检查点**:
+- ✅ `status_code` 为 404 表示 URL 路径错误
+- ✅ `response_body` 可能包含更详细的错误信息
+- ✅ 对比 `url` 和正确的 API endpoint
+
+## 常见的 404 错误原因
+
+### 1. API Base 配置错误
+
+**错误示例**:
+```yaml
+providers:
+  openai:
+    api_base: "https://api.openai.com/v1/chat/completions"  # ❌ 错误：包含了完整路径
+```
+
+**正确配置**:
+```yaml
+providers:
+  openai:
+    api_base: "https://api.openai.com/v1"  # ✅ 正确：只包含 base URL
+```
+
+### 2. 自定义代理或中转服务配置错误
+
+**错误示例**:
+```yaml
+providers:
+  openai:
+    api_base: "https://my-proxy.com"  # ❌ 缺少 /v1 路径
+```
+
+**正确配置**:
+```yaml
+providers:
+  openai:
+    api_base: "https://my-proxy.com/v1"  # ✅ 包含正确的路径
+```
+
+### 3. vLLM 或本地模型服务配置
+
+**正确示例**:
+```yaml
+providers:
+  vllm:
+    api_base: "http://localhost:8000/v1"  # ✅ vLLM 通常也使用 /v1 路径
+```
+
+## 调试命令
+
+### 查看完整的调试日志
+
+```bash
+# 启动 picoclaw 并查看所有 debug 日志
+PICOCLAW_LOG_LEVEL=debug ./picoclaw
+
+# 只看 llm 相关的日志
+PICOCLAW_LOG_LEVEL=debug ./picoclaw | grep '\[llm\]'
+
+# 保存日志到文件以便分析
+PICOCLAW_LOG_LEVEL=debug ./picoclaw 2>&1 | tee debug.log
+```
+
+### 使用 jq 格式化 JSON 日志（如果日志是 JSON 格式）
+
+```bash
+PICOCLAW_LOG_LEVEL=debug ./picoclaw 2>&1 | jq -r 'select(.category == "llm")'
+```
+
+## 示例：完整的调试流程
+
+假设遇到 404 错误，以下是完整的调试输出示例：
+
+```
+[2026-02-12 10:30:00] [DEBUG] [llm] Creating LLM provider
+  model: gpt-4
+  lower_model: gpt-4
+
+[2026-02-12 10:30:00] [INFO] [llm] Provider created successfully
+  model: gpt-4
+  api_base: https://api.openai.com/wrong-path  # ⚠️ 错误的路径
+  has_api_key: true
+  api_key_len: 51
+
+[2026-02-12 10:30:01] [DEBUG] [llm] Sending LLM request
+  url: https://api.openai.com/wrong-path/chat/completions  # ⚠️ 最终的 URL 错误
+  model: gpt-4
+  message_count: 1
+  request_body: {"model":"gpt-4","messages":[...]}
+
+[2026-02-12 10:30:01] [DEBUG] [llm] Sending HTTP request
+  method: POST
+  url: https://api.openai.com/wrong-path/chat/completions
+  headers: {Content-Type: application/json, Authorization: Bearer sk-...}
+
+[2026-02-12 10:30:02] [DEBUG] [llm] Received LLM response
+  status_code: 404
+  status: 404 Not Found
+  response_body: 404 page not found
+
+[2026-02-12 10:30:02] [ERROR] [llm] LLM API returned non-OK status
+  status_code: 404
+  url: https://api.openai.com/wrong-path/chat/completions
+  response_body: 404 page not found
+
+[2026-02-12 10:30:02] [ERROR] [agent] LLM call failed
+  iteration: 1
+  error: API error: 404 page not found
+```
+
+从这个日志可以清楚地看到：
+1. Provider 创建时使用了错误的 `api_base`
+2. 最终的请求 URL 拼接错误
+3. 服务器返回 404 错误
+
+**解决方案**: 修改配置文件中的 `api_base` 为 `https://api.openai.com/v1`
+
+## 敏感信息保护
+
+注意到在日志中：
+- ✅ API Key 只显示长度和前缀，不会完整输出
+- ✅ Authorization header 会被隐藏
+- ⚠️ 完整的请求体和响应体会被记录（debug 级别）
+
+**生产环境建议**:
+- 使用 `info` 或更高级别，避免泄露敏感信息
+- 只在本地开发或受控环境中使用 `debug` 级别
+
+## 相关文件
+
+- `pkg/providers/http_provider.go` - HTTP Provider 实现和调试日志
+- `pkg/agent/loop.go` - Agent 循环和错误处理
+- `pkg/logger/logger.go` - 日志系统实现
+
+## 获取帮助
+
+如果调试日志无法解决问题，请在 GitHub Issue 中提供：
+1. 完整的配置文件（隐藏 API Key）
+2. 相关的调试日志输出
+3. 使用的模型和 provider
+4. 错误发生的上下文
diff --git a/docs/skills-debugging-guide.md b/docs/skills-debugging-guide.md
new file mode 100644
index 000000000..8a1cd77bd
--- /dev/null
+++ b/docs/skills-debugging-guide.md
@@ -0,0 +1,316 @@
+# PicoClaw Skills 调试指南
+
+## 问题排查和解决方案
+
+### 问题1: Skills 无法被识别和使用
+
+**症状**：
+- Agent 有 21 个 skills，但在处理请求时不使用它们
+- 即使明确指定 skill 名称，也提示 skill 不存在
+
+**根本原因**：
+Skills 文件不在正确的位置。代码从以下位置加载 skills：
+1. `~/.picoclaw/workspace/skills/` (workspace skills - 项目级别)
+2. `~/.picoclaw/skills/` (全局 skills)
+3. 内置 skills 目录
+
+但用户的 skills 实际存放在 `~/.claude/skills/`。
+
+**解决方案**：
+将需要使用的 skills 复制到 workspace：
+
+```bash
+# 复制单个 skill
+cp -r ~/.claude/skills/sentinel-search ~/.picoclaw/workspace/skills/
+
+# 或者批量复制所有 skills
+cp -r ~/.claude/skills/* ~/.picoclaw/workspace/skills/
+```
+
+### 问题2: Anthropic API 404 错误
+
+**症状**：
+```
+Error: LLM call failed: API error: 404 page not found
+url=https://meta.nevis.sina.com.cn/wecode/anthropic/chat/completions
+```
+
+**根本原因**：
+- 原来的代码只有一个通用的 `HTTPProvider`
+- 它使用 OpenAI 风格的 API 端点：`/chat/completions`
+- 但 Anthropic API 使用不同的端点：`/v1/messages`
+- 请求/响应格式也完全不同
+
+**解决方案**：
+创建了专门的 `AnthropicProvider` (pkg/providers/anthropic_provider.go)：
+- ✅ 使用正确的端点：`/v1/messages`
+- ✅ 使用正确的请求头：`x-api-key` 而不是 `Authorization: Bearer`
+- ✅ 转换消息格式（system 单独参数、content blocks 等）
+- ✅ 正确处理 tool_use 和 tool_result
+- ✅ 处理孤立的 tool_result（跳过没有对应 tool_use 的结果）
+
+### 问题3: Tool call 参数传递错误
+
+**症状**：
+```
+Error: messages.5.content.1.tool_use.name: String should have at least 1 character
+Error: messages.5.content.1.tool_use.input: Input should be a valid dictionary
+```
+
+**根本原因**：
+Agent 在构建 assistant 消息时，将 tool call 信息存储在：
+- `tc.Function.Name` - 工具名称
+- `tc.Function.Arguments` - JSON 字符串格式的参数
+
+但 AnthropicProvider 最初只读取：
+- `tc.Name` - 为空
+- `tc.Arguments` - 为空的 map
+
+**解决方案**：
+在 AnthropicProvider 中添加了兼容逻辑：
+
+```go
+// 提取 name
+name := tc.Name
+if name == "" && tc.Function != nil {
+    name = tc.Function.Name
+}
+
+// 提取 arguments
+var input map[string]interface{}
+if len(tc.Arguments) > 0 {
+    input = tc.Arguments
+} else if tc.Function != nil && tc.Function.Arguments != "" {
+    // 从 JSON 字符串解析
+    json.Unmarshal([]byte(tc.Function.Arguments), &input)
+}
+```
+
+### 问题4: 孤立的 tool_result 导致 API 错误
+
+**症状**：
+```
+Error: unexpected tool_use_id found in tool_result blocks: toolu_xxx.
+Each tool_result block must have a corresponding tool_use block in the previous message.
+```
+
+**根本原因**：
+- 会话历史可能被截断或清理
+- tool_result 保留了，但对应的 tool_use 被删除了
+- Anthropic API 严格要求 tool_result 必须紧跟在包含对应 tool_use 的 assistant 消息后
+
+**解决方案**：
+添加了 tool_use ID 跟踪机制：
+
+```go
+// 跟踪当前有效的 tool_use IDs
+validToolUseIDs := make(map[string]bool)
+
+// 在 assistant 消息中记录所有 tool_use IDs
+for _, tc := range msg.ToolCalls {
+    validToolUseIDs[tc.ID] = true
+}
+
+// 检查 tool_result 是否有对应的 tool_use
+if msg.Role == "tool" && !validToolUseIDs[msg.ToolCallID] {
+    // 跳过孤立的 tool_result
+    continue
+}
+```
+
+## Skills 使用流程
+
+### 1. Skills 加载
+
+Skills 从以下位置按优先级加载：
+1. **Workspace skills** (`~/.picoclaw/workspace/skills/`)  - 最高优先级，项目专用
+2. **Global skills** (`~/.picoclaw/skills/`) - 中等优先级，用户全局
+3. **Builtin skills** - 最低优先级，系统内置
+
+### 2. Skills 在系统提示中的展示
+
+Skills 只显示摘要信息：
+```xml
+<skills>
+  <skill>
+    <name>sentinel-search</name>
+    <description>Sentinel 安全平台 - 网络资产、服务和 Web 应用搜索工具</description>
+    <location>/path/to/SKILL.md</location>
+    <source>workspace</source>
+  </skill>
+</skills>
+```
+
+提示 AI："To use a skill, read its SKILL.md file using the read_file tool"
+
+### 3. Skills 使用流程
+
+当用户请求使用某个 skill 时：
+1. AI 使用 `read_file` 工具读取 `SKILL.md`
+2. 理解 skill 的 API 文档和使用方法
+3. 使用 `exec` 工具调用相应的命令/API
+4. 处理结果并返回给用户
+
+## 测试 Skills
+
+### 测试 sentinel-search skill
+
+```bash
+# 1. 确保 skill 在正确位置
+ls ~/.picoclaw/workspace/skills/sentinel-search/SKILL.md
+
+# 2. 明确指定使用 sentinel 平台
+./picoclaw agent -m "使用 sentinel 平台搜索 nginx 服务，查询语句是 fp.nmap.product=nginx"
+
+# 3. 观察日志，应该看到：
+#    - list_dir: 列出 skills 目录
+#    - read_file: 读取 SKILL.md
+#    - exec: 执行 curl 命令调用 API
+```
+
+### 日志分析
+
+成功的 skill 使用日志：
+```
+[INFO] agent: LLM requested tool calls {tools=[list_dir], count=1, iteration=1}
+[INFO] agent: Tool call: list_dir({"path":"/Users/xingyue/.picoclaw/workspace/skills"})
+[INFO] agent: LLM requested tool calls {tools=[read_file], count=1, iteration=2}
+[INFO] agent: Tool call: read_file({"path":"...sentinel-search/SKILL.md"})
+[INFO] agent: LLM requested tool calls {tools=[exec], count=1, iteration=3}
+[INFO] agent: Tool call: exec({"command":"curl -X POST http://..."})
+```
+
+## 调试技巧
+
+### 1. 启用 Debug 日志
+
+```bash
+# 启用所有 debug 日志
+export PICOCLAW_LOG_LEVEL=debug
+
+# 只启用 llm 分类的 debug 日志
+export PICOCLAW_LOG_CATEGORY_LLM=debug
+```
+
+### 2. 检查 Skills 是否加载
+
+```bash
+# 查看 agent 初始化日志
+./picoclaw agent -m "test" 2>&1 | grep "Agent initialized"
+# 应该看到: skills_total=22, skills_available=22
+```
+
+### 3. 验证 API 连接
+
+```bash
+# 手动测试 Sentinel API
+curl -X POST http://172.16.10.239:31223/api/search \
+  -H "Content-Type: application/json" \
+  -d '{"query":"fp.nmap.product=nginx","page":1,"page_size":5}'
+```
+
+### 4. 检查网络连接
+
+如果 skill 涉及网络请求，确保：
+- ✅ VPN 已连接（如果需要）
+- ✅ 防火墙允许访问
+- ✅ 目标服务正常运行
+
+## 创建自定义 Skills
+
+### Skill 目录结构
+
+```
+~/.picoclaw/workspace/skills/my-skill/
+├── SKILL.md          # Skill 文档（必需）
+└── .claude/          # 可选的元数据
+    └── config.json
+```
+
+### SKILL.md 格式
+
+```markdown
+---
+name: my-skill
+description: 简短描述（会显示在 skills 列表中）
+version: 1.0.0
+author: your-name
+tags:
+  - tag1
+  - tag2
+triggers:
+  - 触发词1
+  - 触发词2
+---
+
+# Skill 名称
+
+详细的使用说明和 API 文档...
+
+## 使用示例
+
+\```bash
+# 示例命令
+curl ...
+\```
+```
+
+### Best Practices
+
+1. **清晰的文档** - 提供完整的 API 文档和示例
+2. **错误处理** - 说明常见错误和解决方法
+3. **网络要求** - 明确说明网络依赖和访问限制
+4. **参数说明** - 详细描述所有参数和选项
+
+## 相关文件
+
+- `pkg/skills/loader.go` - Skills 加载器
+- `pkg/agent/context.go` - Skills 在上下文中的集成
+- `pkg/providers/anthropic_provider.go` - Anthropic API 适配器
+- `docs/llm-debug-guide.md` - LLM 调试指南
+
+## 未来改进
+
+### 1. Skills 自动同步
+
+考虑添加配置选项，自动从 `~/.claude/skills/` 同步到 workspace：
+
+```yaml
+skills:
+  auto_sync: true
+  source: ~/.claude/skills
+```
+
+### 2. Skills 作为工具
+
+考虑将常用 skills 直接注册为工具，而不需要每次都 read_file：
+
+```go
+// 将 skill 转换为 tool definition
+func (s *Skill) AsToolDefinition() providers.ToolDefinition {
+    // ...
+}
+```
+
+### 3. Skills 模板
+
+提供 skill 创建模板：
+
+```bash
+picoclaw skill new my-skill --template api-wrapper
+```
+
+## 总结
+
+经过调试，现在 PicoClaw 的 skills 系统可以正常工作：
+
+1. ✅ Skills 从正确的位置加载
+2. ✅ Anthropic API 正确处理消息和工具调用
+3. ✅ Tool calls 的 name 和 arguments 正确提取
+4. ✅ 孤立的 tool_result 被过滤
+5. ✅ 详细的调试日志帮助排查问题
+
+Skills 使用流程清晰简单：
+1. 将 skill 放到 workspace/skills/
+2. 明确告诉 AI 使用哪个 skill
+3. AI 读取 SKILL.md 并执行相应操作
diff --git a/pkg/providers/anthropic_provider.go b/pkg/providers/anthropic_provider.go
new file mode 100644
index 000000000..33f690bdb
--- /dev/null
+++ b/pkg/providers/anthropic_provider.go
@@ -0,0 +1,342 @@
+// PicoClaw - Ultra-lightweight personal AI agent
+// Inspired by and based on nanobot: https://github.com/HKUDS/nanobot
+// License: MIT
+//
+// Copyright (c) 2026 PicoClaw contributors
+
+package providers
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/sipeed/picoclaw/pkg/logger"
+)
+
+// AnthropicProvider implements the Anthropic Messages API
+type AnthropicProvider struct {
+	apiKey     string
+	apiBase    string
+	httpClient *http.Client
+}
+
+func NewAnthropicProvider(apiKey, apiBase string) *AnthropicProvider {
+	if apiBase == "" {
+		apiBase = "https://api.anthropic.com/v1"
+	}
+	return &AnthropicProvider{
+		apiKey:  apiKey,
+		apiBase: apiBase,
+		httpClient: &http.Client{
+			Timeout: 0,
+		},
+	}
+}
+
+func (p *AnthropicProvider) Chat(ctx context.Context, messages []Message, tools []ToolDefinition, model string, options map[string]interface{}) (*LLMResponse, error) {
+	if p.apiBase == "" {
+		return nil, fmt.Errorf("API base not configured")
+	}
+
+	// Convert messages to Anthropic format
+	anthropicMessages := make([]map[string]interface{}, 0, len(messages))
+	var systemPrompt string
+
+	// Track tool_use IDs from the previous assistant message
+	validToolUseIDs := make(map[string]bool)
+
+	for i, msg := range messages {
+		if msg.Role == "system" {
+			// Anthropic uses separate system parameter
+			systemPrompt = msg.Content
+			continue
+		}
+
+		anthMsg := map[string]interface{}{
+			"role":    msg.Role,
+			"content": msg.Content,
+		}
+
+		// Handle tool results - only include if there's a corresponding tool_use
+		if msg.Role == "tool" || msg.ToolCallID != "" {
+			// Check if this tool_call_id was in the previous assistant message
+			if !validToolUseIDs[msg.ToolCallID] {
+				logger.WarnCF("llm", "Skipping tool result without corresponding tool_use",
+					map[string]interface{}{
+						"tool_call_id": msg.ToolCallID,
+						"message_idx":  i,
+					})
+				continue // Skip orphaned tool results
+			}
+
+			anthMsg["role"] = "user"
+			anthMsg["content"] = []map[string]interface{}{
+				{
+					"type":         "tool_result",
+					"tool_use_id":  msg.ToolCallID,
+					"content":      msg.Content,
+				},
+			}
+		}
+
+		// Handle assistant messages with tool calls
+		if msg.Role == "assistant" && len(msg.ToolCalls) > 0 {
+			// Clear previous valid IDs and record new ones
+			validToolUseIDs = make(map[string]bool)
+
+			contentBlocks := make([]map[string]interface{}, 0, len(msg.ToolCalls)+1)
+
+			// Add text content if present
+			if msg.Content != "" {
+				contentBlocks = append(contentBlocks, map[string]interface{}{
+					"type": "text",
+					"text": msg.Content,
+				})
+			}
+
+			// Add tool use blocks
+			for _, tc := range msg.ToolCalls {
+				// Extract name from either tc.Name or tc.Function.Name
+				name := tc.Name
+				if name == "" && tc.Function != nil {
+					name = tc.Function.Name
+				}
+
+				// Extract arguments - they might be in tc.Arguments (map) or tc.Function.Arguments (JSON string)
+				var input map[string]interface{}
+				if len(tc.Arguments) > 0 {
+					input = tc.Arguments
+				} else if tc.Function != nil && tc.Function.Arguments != "" {
+					// Parse JSON string to map
+					if err := json.Unmarshal([]byte(tc.Function.Arguments), &input); err != nil {
+						logger.ErrorCF("llm", "Failed to parse tool arguments",
+							map[string]interface{}{
+								"id":    tc.ID,
+								"name":  name,
+								"error": err.Error(),
+								"raw":   tc.Function.Arguments,
+							})
+						continue
+					}
+				} else {
+					input = make(map[string]interface{})
+				}
+
+				logger.DebugCF("llm", "Converting tool call to Anthropic format",
+					map[string]interface{}{
+						"id":    tc.ID,
+						"name":  name,
+						"type":  tc.Type,
+						"input": input,
+					})
+
+				if name == "" {
+					logger.ErrorCF("llm", "Tool call has no name",
+						map[string]interface{}{
+							"id":            tc.ID,
+							"type":          tc.Type,
+							"has_function":  tc.Function != nil,
+							"tc_name":       tc.Name,
+						})
+					continue // Skip this tool call
+				}
+
+				// Record this as a valid tool_use ID
+				validToolUseIDs[tc.ID] = true
+
+				contentBlocks = append(contentBlocks, map[string]interface{}{
+					"type":  "tool_use",
+					"id":    tc.ID,
+					"name":  name,
+					"input": input,
+				})
+			}
+
+			anthMsg["content"] = contentBlocks
+		} else if msg.Role == "assistant" {
+			// Assistant message without tool calls - clear valid IDs
+			validToolUseIDs = make(map[string]bool)
+		}
+
+		anthropicMessages = append(anthropicMessages, anthMsg)
+	}
+
+	// Build request body
+	requestBody := map[string]interface{}{
+		"model":    model,
+		"messages": anthropicMessages,
+	}
+
+	if systemPrompt != "" {
+		requestBody["system"] = systemPrompt
+	}
+
+	// Add tools if present
+	if len(tools) > 0 {
+		anthropicTools := make([]map[string]interface{}, 0, len(tools))
+		for _, tool := range tools {
+			anthropicTools = append(anthropicTools, map[string]interface{}{
+				"name":         tool.Function.Name,
+				"description":  tool.Function.Description,
+				"input_schema": tool.Function.Parameters,
+			})
+		}
+		requestBody["tools"] = anthropicTools
+	}
+
+	// Add max_tokens (required by Anthropic)
+	if maxTokens, ok := options["max_tokens"].(int); ok {
+		requestBody["max_tokens"] = maxTokens
+	} else {
+		requestBody["max_tokens"] = 8192 // Default
+	}
+
+	// Add temperature if specified
+	if temperature, ok := options["temperature"].(float64); ok {
+		requestBody["temperature"] = temperature
+	}
+
+	jsonData, err := json.Marshal(requestBody)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	fullURL := p.apiBase + "/messages"
+
+	// Debug log: Log request details
+	logger.DebugCF("llm", "Sending Anthropic API request",
+		map[string]interface{}{
+			"url":           fullURL,
+			"model":         model,
+			"api_base":      p.apiBase,
+			"has_api_key":   p.apiKey != "",
+			"api_key_len":   len(p.apiKey),
+			"message_count": len(anthropicMessages),
+			"tools_count":   len(tools),
+			"request_body":  string(jsonData),
+		})
+
+	req, err := http.NewRequestWithContext(ctx, "POST", fullURL, bytes.NewReader(jsonData))
+	if err != nil {
+		logger.ErrorCF("llm", "Failed to create HTTP request",
+			map[string]interface{}{
+				"url":   fullURL,
+				"error": err.Error(),
+			})
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	// Set Anthropic-specific headers
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("x-api-key", p.apiKey)
+	req.Header.Set("anthropic-version", "2023-06-01")
+
+	logger.DebugCF("llm", "Sending HTTP request",
+		map[string]interface{}{
+			"method": "POST",
+			"url":    fullURL,
+		})
+
+	resp, err := p.httpClient.Do(req)
+	if err != nil {
+		logger.ErrorCF("llm", "HTTP request failed",
+			map[string]interface{}{
+				"url":   fullURL,
+				"error": err.Error(),
+			})
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		logger.ErrorCF("llm", "Failed to read response body",
+			map[string]interface{}{
+				"status_code": resp.StatusCode,
+				"error":       err.Error(),
+			})
+		return nil, fmt.Errorf("failed to read response: %w", err)
+	}
+
+	// Debug log: Log response details
+	logger.DebugCF("llm", "Received Anthropic API response",
+		map[string]interface{}{
+			"status_code":    resp.StatusCode,
+			"status":         resp.Status,
+			"content_length": len(body),
+			"response_body":  string(body),
+		})
+
+	if resp.StatusCode != http.StatusOK {
+		logger.ErrorCF("llm", "Anthropic API returned non-OK status",
+			map[string]interface{}{
+				"status_code":   resp.StatusCode,
+				"status":        resp.Status,
+				"url":           fullURL,
+				"response_body": string(body),
+			})
+		return nil, fmt.Errorf("API error: %s", string(body))
+	}
+
+	return p.parseResponse(body)
+}
+
+func (p *AnthropicProvider) parseResponse(body []byte) (*LLMResponse, error) {
+	var apiResponse struct {
+		ID      string `json:"id"`
+		Type    string `json:"type"`
+		Role    string `json:"role"`
+		Content []struct {
+			Type  string `json:"type"`
+			Text  string `json:"text,omitempty"`
+			ID    string `json:"id,omitempty"`
+			Name  string `json:"name,omitempty"`
+			Input map[string]interface{} `json:"input,omitempty"`
+		} `json:"content"`
+		StopReason string `json:"stop_reason"`
+		Usage      struct {
+			InputTokens  int `json:"input_tokens"`
+			OutputTokens int `json:"output_tokens"`
+		} `json:"usage"`
+	}
+
+	if err := json.Unmarshal(body, &apiResponse); err != nil {
+		return nil, fmt.Errorf("failed to unmarshal response: %w", err)
+	}
+
+	// Extract text content and tool calls
+	var textContent string
+	toolCalls := make([]ToolCall, 0)
+
+	for _, content := range apiResponse.Content {
+		switch content.Type {
+		case "text":
+			textContent += content.Text
+		case "tool_use":
+			toolCalls = append(toolCalls, ToolCall{
+				ID:        content.ID,
+				Name:      content.Name,
+				Arguments: content.Input,
+			})
+		}
+	}
+
+	return &LLMResponse{
+		Content:      textContent,
+		ToolCalls:    toolCalls,
+		FinishReason: apiResponse.StopReason,
+		Usage: &UsageInfo{
+			PromptTokens:     apiResponse.Usage.InputTokens,
+			CompletionTokens: apiResponse.Usage.OutputTokens,
+			TotalTokens:      apiResponse.Usage.InputTokens + apiResponse.Usage.OutputTokens,
+		},
+	}, nil
+}
+
+func (p *AnthropicProvider) GetDefaultModel() string {
+	return "claude-sonnet-4-5"
+}
diff --git a/pkg/providers/http_provider.go b/pkg/providers/http_provider.go
index 12909df9e..ff811e242 100644
--- a/pkg/providers/http_provider.go
+++ b/pkg/providers/http_provider.go
@@ -16,6 +16,7 @@ import (
 	"strings"
 
 	"github.com/sipeed/picoclaw/pkg/config"
+	"github.com/sipeed/picoclaw/pkg/logger"
 )
 
 type HTTPProvider struct {
@@ -24,6 +25,14 @@ type HTTPProvider struct {
 	httpClient *http.Client
 }
 
+// Helper function to get minimum of two integers
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
 func NewHTTPProvider(apiKey, apiBase string) *HTTPProvider {
 	return &HTTPProvider{
 		apiKey:  apiKey,
@@ -67,8 +76,28 @@ func (p *HTTPProvider) Chat(ctx context.Context, messages []Message, tools []Too
 		return nil, fmt.Errorf("failed to marshal request: %w", err)
 	}
 
-	req, err := http.NewRequestWithContext(ctx, "POST", p.apiBase+"/chat/completions", bytes.NewReader(jsonData))
+	fullURL := p.apiBase + "/chat/completions"
+
+	// Debug log: Log request details
+	logger.DebugCF("llm", "Sending LLM request",
+		map[string]interface{}{
+			"url":          fullURL,
+			"model":        model,
+			"api_base":     p.apiBase,
+			"has_api_key":  p.apiKey != "",
+			"api_key_len":  len(p.apiKey),
+			"message_count": len(messages),
+			"tools_count":  len(tools),
+			"request_body": string(jsonData),
+		})
+
+	req, err := http.NewRequestWithContext(ctx, "POST", fullURL, bytes.NewReader(jsonData))
 	if err != nil {
+		logger.ErrorCF("llm", "Failed to create HTTP request",
+			map[string]interface{}{
+				"url":   fullURL,
+				"error": err.Error(),
+			})
 		return nil, fmt.Errorf("failed to create request: %w", err)
 	}
 
@@ -76,20 +105,57 @@ func (p *HTTPProvider) Chat(ctx context.Context, messages []Message, tools []Too
 	if p.apiKey != "" {
 		authHeader := "Bearer " + p.apiKey
 		req.Header.Set("Authorization", authHeader)
+		logger.DebugCF("llm", "Authorization header set",
+			map[string]interface{}{
+				"key_prefix": p.apiKey[:min(10, len(p.apiKey))] + "...",
+			})
 	}
 
+	logger.DebugCF("llm", "Sending HTTP request",
+		map[string]interface{}{
+			"method":  "POST",
+			"url":     fullURL,
+			"headers": req.Header,
+		})
+
 	resp, err := p.httpClient.Do(req)
 	if err != nil {
+		logger.ErrorCF("llm", "HTTP request failed",
+			map[string]interface{}{
+				"url":   fullURL,
+				"error": err.Error(),
+			})
 		return nil, fmt.Errorf("failed to send request: %w", err)
 	}
 	defer resp.Body.Close()
 
 	body, err := io.ReadAll(resp.Body)
 	if err != nil {
+		logger.ErrorCF("llm", "Failed to read response body",
+			map[string]interface{}{
+				"status_code": resp.StatusCode,
+				"error":       err.Error(),
+			})
 		return nil, fmt.Errorf("failed to read response: %w", err)
 	}
 
+	// Debug log: Log response details
+	logger.DebugCF("llm", "Received LLM response",
+		map[string]interface{}{
+			"status_code":   resp.StatusCode,
+			"status":        resp.Status,
+			"content_length": len(body),
+			"response_body": string(body),
+		})
+
 	if resp.StatusCode != http.StatusOK {
+		logger.ErrorCF("llm", "LLM API returned non-OK status",
+			map[string]interface{}{
+				"status_code":   resp.StatusCode,
+				"status":        resp.Status,
+				"url":           fullURL,
+				"response_body": string(body),
+			})
 		return nil, fmt.Errorf("API error: %s", string(body))
 	}
 
@@ -177,6 +243,12 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 
 	lowerModel := strings.ToLower(model)
 
+	logger.DebugCF("llm", "Creating LLM provider",
+		map[string]interface{}{
+			"model":       model,
+			"lower_model": lowerModel,
+		})
+
 	switch {
 	case strings.HasPrefix(model, "openrouter/") || strings.HasPrefix(model, "anthropic/") || strings.HasPrefix(model, "openai/") || strings.HasPrefix(model, "meta-llama/") || strings.HasPrefix(model, "deepseek/") || strings.HasPrefix(model, "google/"):
 		apiKey = cfg.Providers.OpenRouter.APIKey
@@ -187,12 +259,23 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 		}
 
 	case (strings.Contains(lowerModel, "claude") || strings.HasPrefix(model, "anthropic/")) && cfg.Providers.Anthropic.APIKey != "":
+		// Use dedicated Anthropic provider
 		apiKey = cfg.Providers.Anthropic.APIKey
 		apiBase = cfg.Providers.Anthropic.APIBase
 		if apiBase == "" {
 			apiBase = "https://api.anthropic.com/v1"
 		}
 
+		logger.InfoCF("llm", "Anthropic provider created successfully",
+			map[string]interface{}{
+				"model":       model,
+				"api_base":    apiBase,
+				"has_api_key": apiKey != "",
+				"api_key_len": len(apiKey),
+			})
+
+		return NewAnthropicProvider(apiKey, apiBase), nil
+
 	case (strings.Contains(lowerModel, "gpt") || strings.HasPrefix(model, "openai/")) && cfg.Providers.OpenAI.APIKey != "":
 		apiKey = cfg.Providers.OpenAI.APIKey
 		apiBase = cfg.Providers.OpenAI.APIBase
@@ -239,12 +322,28 @@ func CreateProvider(cfg *config.Config) (LLMProvider, error) {
 	}
 
 	if apiKey == "" && !strings.HasPrefix(model, "bedrock/") {
+		logger.ErrorCF("llm", "No API key configured",
+			map[string]interface{}{
+				"model": model,
+			})
 		return nil, fmt.Errorf("no API key configured for provider (model: %s)", model)
 	}
 
 	if apiBase == "" {
+		logger.ErrorCF("llm", "No API base configured",
+			map[string]interface{}{
+				"model": model,
+			})
 		return nil, fmt.Errorf("no API base configured for provider (model: %s)", model)
 	}
 
+	logger.InfoCF("llm", "Provider created successfully",
+		map[string]interface{}{
+			"model":       model,
+			"api_base":    apiBase,
+			"has_api_key": apiKey != "",
+			"api_key_len": len(apiKey),
+		})
+
 	return NewHTTPProvider(apiKey, apiBase), nil
 }
\ No newline at end of file