使用说明
LLM MCP服务器
MCP（模型上下文协议）服务器，提供对各种LLM API的访问，作为克劳德代码的工具。
特性
- 化学 K2 -最先进的1T参数MoE模型，具有128k上下文窗口
- 双子座2.5 Pro -处理大量上下文窗口以进行全面分析
- 双子座2.5闪光灯 -快速、功能强大的平衡任务模型
- GPT-4.1-nano -超快速、轻量级的完成简单任务
- GPT-4.1分钟 -性能良好的替代平衡模型
- 自定义模型布线 -为每个任务选择合适的模型
完整安装指南
第一步：克隆或下载
cd ~/projects # or your preferred directory
git clone [repository-url] llm-mcp-server
cd llm-mcp-server
步骤2：运行安装脚本
chmod +x setup.sh
./setup.sh
这将：
- 创建Python虚拟环境
- 安装所有依赖项
- 创建一个
.env文件来自.env.example模板 - 使服务器可执行
步骤3：添加您的API密钥
编辑 .env 文件：
nano .env # or use your preferred editor
至少添加一个API密钥（您不需要所有密钥）：
# Anthropic API for Claude models
ANTHROPIC_API_KEY=your_anthropic_api_key_here
# OpenAI API for GPT models
OPENAI_API_KEY=your_openai_api_key_here
# Google API for Gemini models
GOOGLE_API_KEY=your_google_api_key_here
# Moonshot API for Kimi K2 models
MOONSHOT_API_KEY=your_moonshot_api_key_here
# Optional: Log costs to file (JSON lines format)
# LLM_COST_LOG=/path/to/llm_costs.jsonl
步骤4：配置Claude代码
在Claude Code中，使用以下命令添加MCP服务器：
claude mcp add --scope user llm-tools python /full/path/to/llm-mcp-server/server.py
例如：
claude mcp add --scope user llm-tools python /Users/yourname/projects/llm-mcp-server/server.py
步骤5：验证安装
检查是否已配置：
/mcp
您应该看到列出了“llm工具”。
步骤6：重新启动Claude代码
- 完全退出克劳德代码（Mac上的Cmd+Q，Windows上的Alt+F4）
- 重新打开克劳德代码
步骤7：测试工具
在任何克劳德对话中，工具现在都可用。通过提问进行测试：
- “使用analyze_with_gemini检查工具是否正常工作”
- “使用check_costs查看成本跟踪”
可用工具
与戈米尼一起分析
使用Gemini 2.5 Pro的海量上下文窗口分析大型代码库或文档。
Parameters:
- prompt: The analysis query
- files: List of file paths to include in context (optional)
- context: Direct context string (optional)
quick_gpt
使用GPT-4.1-nano快速响应简单任务。
Parameters:
- prompt: The task or question
- temperature: Control randomness (0-1, default 0.3)
平衡llm
使用Gemini 2.5 Flash或GPT-4.1-mini进行平衡任务。
Parameters:
- prompt: The task or question
- model: "gemini-flash" or "gpt-mini" (default: "gemini-flash")
- max_tokens: Maximum response length (default 1000)
路由到测试模型
根据任务自动选择最佳模型。
Parameters:
- prompt: The task description
- task_type: "analysis", "generation", "simple", or "auto"
Kimi聊天
灵活的Kimi/Moonshot聊天界面，支持简单提示、多回合对话和部分预填充。
Parameters (all optional):
- prompt: Simple prompt for single-turn, or latest user message
- messages: Full conversation history array (overrides prompt if provided)
- Each message: {"role": "system/user/assistant", "content": "...", "name": "optional", "partial": true/false}
- system: System message (used only if messages not provided)
- partial_response: Pre-fill the assistant's response to maintain character/format
- model: Model name - Common options:
- 'kimi-k2-0711-preview' (default) - Kimi K2 instruct model for chat, tool use, structured responses
- 'kimi-k2-base' - Base model for creative completions (when available via API)
- 'moonshot-v1-auto' - Automatically selects appropriate context size
- 'moonshot-v1-128k' - 128k context window
- Note: Currently only instruct models are available via API
- temperature: Control randomness (0-1, default 0.6)
- max_tokens: Maximum response length (default 4096)
- available_tools: List of built-in tool names to enable (optional)
- Built-in tools: ["get_current_time", "add_numbers", "list_files"]
- dynamic_tools: Array of custom tool definitions for on-the-fly CLI execution (optional)
- Each tool needs: name, command, and schema properties
- session_id: Persistent conversation management (optional)
- New sessions: Use descriptive keywords like "python-async-debugging" (auto-timestamped)
- Continue existing: Use full ID like "python-async-debugging_20250712_2055"
- Special commands: "@last" (continue most recent), "@list" (show all sessions), "@clear:id" (delete session)
- return_conversation: Return full conversation history with response (default: false)
Usage modes:
1. Simple: kimi_chat(prompt="Hello")
2. With system: kimi_chat(system="You are a pirate", prompt="Tell me about treasure")
3. Multi-turn: kimi_chat(messages=[...])
4. With pre-filling: kimi_chat(prompt="...", partial_response="*thinks carefully* ")
5. Full conversation + pre-fill: kimi_chat(messages=[...], partial_response="...")
6. With built-in tools: kimi_chat(prompt="What time is it?", available_tools=["get_current_time"])
7. With dynamic tools: kimi_chat(prompt="Check port 8080", dynamic_tools=[{...}])
8. Start new session: kimi_chat(prompt="Let's discuss Python", session_id="python-basics")
9. Continue session: kimi_chat(prompt="Tell me more", session_id="python-basics_20250712_2055")
10. Continue last session: kimi_chat(prompt="What were we discussing?", session_id="@last")
11. List sessions: kimi_chat(session_id="@list")
12. Get conversation state: kimi_chat(prompt="Hello", return_conversation=true)
会话管理
Kimi对话可以持久到磁盘上以实现长期连续性：
会话存储：
- 会话存储在
.kimi_sessions/目录（gitignored） - 每个会话都是一个包含消息、元数据和时间戳的JSON文件
- 会话ID使用格式：
keywords_YYYYMMDD_HHMM
会话工作流示例：
# Start a new project discussion
kimi_chat(
prompt="I need help designing a REST API",
session_id="api-design" # Creates: api-design_20250712_2100.json
)
# Continue later (use full ID)
kimi_chat(
prompt="What about authentication?",
session_id="api-design_20250712_2100"
)
# Or continue the most recent session
kimi_chat(
prompt="Should we use JWT or sessions?",
session_id="@last"
)
# Check all active sessions
kimi_chat(session_id="@list")
会话功能：
- 接近上下文限制时自动截断消息
- 从会话ID中提取主题以便于识别
- 创建和上次访问的时间戳
- 会话元数据保存
动态工具生成
这 kimi_chat 该工具支持使用 dynamic_tools 参数：
kimi_chat(
prompt="What's my public IP address?",
dynamic_tools=[{
"name": "get_public_ip",
"command": "curl -s ifconfig.me",
"schema": {
"type": "function",
"function": {
"name": "get_public_ip",
"description": "Get public IP address",
"parameters": {
"type": "object",
"properties": {},
"required": []
}
}
}
}]
)
动态工具示例：
# Network operations
{
"name": "check_port",
"command": "netstat -an | grep :8080",
"schema": {...}
}
# File operations
{
"name": "find_large_files",
"command": "find . -size +100M -type f -exec ls -lh {} \\;",
"schema": {...}
}
# System information
{
"name": "memory_usage",
"command": "free -h",
"schema": {...}
}
# Git operations
{
"name": "recent_commits",
"command": "git log --oneline -5",
"schema": {...}
}
工具执行功能：
- 30秒安全超时
- Stout/stderr捕获
- 错误处理和报告
- 适用于任何shell命令
- 与内置工具无缝结合
完整的动态工具示例：
# Request: "Check if port 8080 is open and count Python files"
kimi_chat(
prompt="Check if port 8080 is open and count Python files in this directory",
dynamic_tools=[
{
"name": "check_port_8080",
"command": "netstat -an | grep :8080",
"schema": {
"type": "function",
"function": {
"name": "check_port_8080",
"description": "Check if port 8080 is listening",
"parameters": {"type": "object", "properties": {}, "required": []}
}
}
},
{
"name": "count_python_files",
"command": "find . -name '*.py' -type f | wc -l",
"schema": {
"type": "function",
"function": {
"name": "count_python_files",
"description": "Count Python files in directory",
"parameters": {"type": "object", "properties": {}, "required": []}
}
}
}
]
)
# Result:
# "Port 8080 is not currently listening, and there are 3,234 Python files in this directory.
#
# **Tool Executions:**
# 🔧 check_port_8080({}) → (no output - port not listening)
# 🔧 count_python_files({}) → 3234"
## Kimi K2 Tool Calling Capabilities
Kimi K2 has native tool-calling capabilities that allow it to autonomously execute functions and orchestrate workflows. The model can:
- **Autonomous Task Decomposition**: Break down complex requests into tool sequences
- **Multiple Tool Calls**: Execute several functions in a single response
- **Parameter Parsing**: Correctly extract and format function arguments
- **Context Preservation**: Maintain conversation state throughout tool interactions
### Tool Calling API Flow
The tool calling process follows OpenAI's tools format:
#### 1. Tool Definition
Tools are defined using JSON schemas:
```json
{
"type": "function",
"function": {
"name": "get_current_time",
"description": "Get the current date and time",
"parameters": {
"type": "object",
"properties": {},
"required": []
}
}
}
2.工具调用请求
当Kimi想要使用工具时，它会回应：
{
"role": "assistant",
"content": "I'll get the current time for you.",
"tool_calls": [
{
"index": 0,
"id": "get_current_time:0",
"type": "function",
"function": {
"name": "get_current_time",
"arguments": "{}"
}
}
]
}
3.工具执行
您的应用程序执行函数并返回结果：
{
"role": "tool",
"tool_call_id": "get_current_time:0",
"content": "2025-07-12 14:30:45"
}
4.最终答复
Kimi将工具结果合成为自然语言：
{
"role": "assistant",
"content": "It is currently 2:30:45 PM on July 12, 2025."
}
示例：多个工具调用
请求：
"What time is it, and then add 10 plus 5?"
Kimi的回应：
{
"tool_calls": [
{
"id": "get_current_time:0",
"function": {"name": "get_current_time", "arguments": "{}"}
},
{
"id": "add_numbers:1",
"function": {"name": "add_numbers", "arguments": "{\"a\": 10, \"b\": 5}"}
}
]
}
工具结果：
[
{"role": "tool", "tool_call_id": "get_current_time:0", "content": "2025-07-12 14:45:30"},
{"role": "tool", "tool_call_id": "add_numbers:1", "content": "15"}
]
最终回复：
"It's currently **2:45:30 PM** on July 12, 2025.
And **10 + 5 = 15**."
关键实施说明
- 工具调用ID：每个工具调用都有一个唯一的ID（例如。，
"function_name:index") - 参数格式：函数参数是JSON字符串，即使是简单的参数也是如此
- 多种工具：Kimi可以在一个响应中调用多个工具来处理复杂的请求
- 错误处理：工具执行错误应作为工具消息返回，供Kimi处理
- 流媒体：工具调用可用于流式和非流式API调用
check_costs
检查此会话中所有LLM使用的累积成本。
Parameters: none
Returns: Detailed cost breakdown by model
Prompts中的文件插值
所有工具都支持使用简单语法进行自动文件插值：
单个文件
# Include entire file
analyze_with_gemini(
prompt="Review this code: {file:server.py}"
)
# Include specific line range
quick_gpt(
prompt="Explain this function: {file:utils/helper.py:10-25}"
)
多个文件
# Use glob patterns
analyze_with_gemini(
prompt="Review all test files: {files:tests/*.py}"
)
# Mix multiple files
kimi_chat(
prompt="Compare {file:old_version.py} with {file:new_version.py}"
)
复杂示例
# Combine with regular text
analyze_with_gemini(
prompt="""Based on the README: {file:README.md}
Please explain how this implementation works:
{file:core/engine.py:50-150}
And how it relates to these tests:
{files:tests/test_engine*.py}"""
)
文件内容会自动读取并插入到您的提示中，使引用代码变得容易，而无需手动复制粘贴。
使用示例
在Claude Code中，这些工具可供使用：
# Analyze entire codebase
result = analyze_with_gemini(
prompt="Find all API endpoints and their authentication methods",
files=["src/**/*.py"]
)
# Quick formatting
formatted = quick_gpt(
prompt="Format this JSON: {a:1,b:2}"
)
# Smart routing
response = route_to_best_model(
prompt="Explain the authentication flow",
task_type="analysis"
)
# Kimi/Moonshot examples
# Simple query
answer = kimi_chat(prompt="What is quantum computing?")
# Role-playing with system message
story = kimi_chat(
system="You are a wise wizard who speaks in riddles",
prompt="Tell me about dragons"
)
# Multi-turn conversation
chat = kimi_chat(
messages=[
{"role": "system", "content": "You are a helpful coding assistant"},
{"role": "user", "content": "How do I sort a list in Python?"},
{"role": "assistant", "content": "You can use the sorted() function or list.sort() method."},
{"role": "user", "content": "What's the difference?"}
]
)
# With partial pre-filling for character consistency
response = kimi_chat(
system="You are Dr. Kelsier from Arknights",
prompt="What do you think about the current situation?",
partial_response="*adjusts her monocle with a slight frown* The situation, you ask? "
)
# Using different Moonshot models
auto_response = kimi_chat(
prompt="Explain quantum computing",
model="moonshot-v1-auto" # Automatically selects context size
)
成本跟踪
每个请求都会自动跟踪：
- 令牌使用情况（估计）
- 每次请求的成本
- 每种型号的累计成本
- 所有型号的总成本
每个回复都包括成本信息：
[Response content]
---
💰 Cost: $0.000125 | Total: $0.0045
使用 check_costs() 工具随时查看详细明细。
故障排除
工具没有出现？
- 确保您在中使用了完整的绝对路径
claude mcp add命令 - 检查
/mcp将服务器显示为“正在运行” - 确保您完全重新启动了Claude Code
- 验证中是否至少设置了一个API密钥
.env
正在获取API错误？
- 检查您的API密钥是否有效
- 确保您的帐户上有信用/配额
- 谷歌API关键需求Gemini API在谷歌云控制台中启用
- OpenAI密钥需要适当的模型访问
- Anthropic密钥需要来自Anthropics控制台
Python错误？
- 确保已安装Python 3.8+
- 尝试使用虚拟环境：
source venv/bin/activate - 检查已安装的所有依赖项：
pip list
服务器未启动？
- 检查server.py是否具有执行权限：
chmod +x server.py - 验证MCP配置中的路径是否正确
- 检查克劳德代码日志中的错误消息
Slash命令
您可以使用以下工具构建斜线命令：
# .claude/commands/analyze-all.md
Analyze entire codebase with Gemini
Use Gemini's 1M context to analyze all files:
- Security vulnerabilities
- Architecture patterns
- Code quality issues
