houtini-lm

概述

Houtini LM 是一个专门的 MCP 服务器，旨在通过将有界任务从 Claude Code 卸载到本地或云 LLM 来节省代币。它可与多种端点协同工作，包括 LM Studio、Ollama、vLLM、DeepSeek、Groq、Cerebras 和 OpenRouter。该服务器能够智能地将任务路由到合适的模型，跟踪性能指标，并提供模型发现功能。它专注于卸载明确定义的任务，如代码审查、测试生成和文档编写，同时将复杂的推理任务保留给 Claude。

试试问 AI

装完之后，这里有 5 个你可以让 AI 做的事：

• chatmessage: string, system?: string, temperature?: number, max_tokens?: number, json_schema?: object, model?: string

  Send a task to an LLM and get an answer. Offloads bounded tasks like generating boilerplate or explanations.

  试试:Generate unit tests for this function试试:Explain this Python code in plain English试试:Draft a commit message for these changes
• custom_promptinstruction: string, system?: string, context?: string, temperature?: number, max_tokens?: number, json_schema?: object, model?: string

  Send a three-part prompt (system, context, instruction) to an LLM for consistent high-quality output.

  试试:Review this code for security issues试试:Summarize this research paper试试:Convert this API spec to OpenAPI format
• code_taskcode: string, task: string, language?: string, max_tokens?: number, model?: string

  Built-in code analysis with pre-configured system prompts for finding bugs, explaining code, or writing tests.

  试试:Find bugs in this TypeScript function试试:Explain this Python class试试:Generate unit tests for this Rust code
• code_task_filestask: string, files: string[], max_tokens?: number, model?: string

  Review multiple related files or large files directly from disk without context window limits.

  试试:Review these React components for consistency试试:Analyze this entire codebase for architectural issues试试:Refactor these helper functions

可对比工具

安装

# 用于 Claude Code
claude mcp add houtini-lm -- npx -y @houtini/lm

# 用于 Claude Desktop
添加到 claude_desktop_config.json:
{
  "mcpServers": {
    "houtini-lm": {
      "command": "npx",
      "args": ["-y", "@houtini/lm"],
      "env": {
        "HOUTINI_LM_ENDPOINT_URL": "http://localhost:1234"
      }
    }
  }
}

FAQ

哪些类型的任务会被卸载到本地模型？

有界且定义明确的任务，如生成测试存根、代码审查、格式转换、生成模拟数据和编写类型定义。复杂的推理、工具访问和多文件更改仍保留给 Claude。

模型路由如何工作？

Houtini-LM 查询您的 LLM 服务器以获取可用模型，在 HuggingFace 上查找它们的元数据，并在本地数据库中维护模型功能。然后它会根据任务类型对加载的模型进行评分，自动为每个任务选择最佳模型。

houtini-lm 对比