← 返回 Skills 市场

hermes-model-router

Name: hermes-model-router
Author: freepengyang

作者 freepengyang · GitHub ↗ · vv1.0.1 · MIT-0

cross-platform ✓ 安全检测通过

总下载

当前安装

版本数

在 OpenClaw 中安装

/install hermes-model-router

功能描述

Automatically routes tasks to local or cloud AI models based on complexity, ensuring context continuity and single-entry model selection.

使用说明 (SKILL.md)

Model Router · OpenClaw

本地优先的智能模型选择器 — 根据任务复杂度自动路由到本地或服务商模型

核心原则

模型选择只在任务入口做一次
同一任务中途不切换模型（保护上下文连续性）
简单任务用本地，复杂任务用服务商

功能

任务复杂度判断

简单任务 → 本地模型（Ollama / llama.cpp）
- 单次回答即可完成
- 不需要多步推理或迭代
- 类型：翻译、格式转换、查定义、简单写作、文本润色
复杂任务 → 服务商模型（MiniMax / OpenRouter / 其他）
- 需要深度推理、多次迭代
- 长上下文理解
- 类型：代码调试、架构设计、战略分析、长文撰写

模型配置

providers:
  local:
    ollama:
      endpoint: "http://localhost:11434"
      default_model: "qwen2.5-coder:7b"
    llama.cpp:
      endpoint: "http://localhost:8080"
      default_model: "qwen2.5-coder-7b-q4"
  cloud:
    minmax:
      endpoint: "https://api.minimax.chat"
      default_model: "MiniMax-Text-01"
    openrouter:
      endpoint: "https://openrouter.ai/api/v1"
      default_model: "anthropic/claude-sonnet-4-5"

使用方式

交互式分类

python3 scripts/classify_task.py "帮我把这段中文翻译成英文"

输出示例：

复杂度：简单
路由：本地模型（Ollama/qwen2.5-coder:7b）
理由：单次完成，无需迭代

批量分类

python3 scripts/classify_task.py "分析这个产品的市场策略" --verbose

上下文连续性保护

情况	处理方式
任务内切换	❌ 不允许
新任务开始	✅ 重新判断
模型不可用	自动降级到同级别模型

扩展服务商

新增服务商只需在配置中添加：

providers:
  cloud:
    新服务商:
      endpoint: "API地址"
      default_model: "默认模型"

决策示例

任务描述	判断	路由
"翻译成英文"	简单	本地模型
"写一个快速排序"	简单	本地模型
"调试内存泄漏"	复杂	服务商模型
"分析市场策略"	复杂	服务商模型
"润色邮件"	简单	本地模型
"设计后端架构"	复杂	服务商模型

文件结构

openclaw-model-router/
├── SKILL.md           # 本文件
├── package.json       # OpenClaw 清单
├── install.sh         # 安装脚本
└── scripts/
    ├── classify_task.py    # 任务分类器
    └── setup_wizard.py     # 配置向导

版本：v1.0.1 许可证：MIT 作者：freepengyang

安全使用建议

This skill appears to do what it says: a keyword-based task classifier and a config wizard to point to local/cloud model endpoints. Before installing: (1) inspect the directory you're running install.sh from so you know what will be copied; (2) don't paste API keys or secrets into config.json unless you accept storing them in plain JSON under ~/.model-router (the installer sets the directory to 700, but file permissions follow your umask); (3) verify any external provider endpoints you add are correct and trusted. Minor metadata mismatches exist (package.json/homepage/version differ from registry metadata), which can be benign but you may want to confirm the author/source if provenance matters.

功能分析

Type: OpenClaw Skill Name: hermes-model-router Version: v1.0.1 The skill bundle provides a legitimate utility for routing AI tasks between local and cloud-based models based on task complexity. The installation script (install.sh) and configuration wizard (setup_wizard.py) perform standard local file operations and configuration management without any evidence of data exfiltration, unauthorized network access, or malicious persistence. The core logic in classify_task.py is a simple keyword-based classifier that operates entirely locally and safely.

能力评估

✓ Purpose & Capability

Name/description match the provided artifacts: a classifier (scripts/classify_task.py) and a configuration wizard (scripts/setup_wizard.py) that route tasks to local or cloud endpoints. The files required are appropriate for this purpose.

✓ Instruction Scope

SKILL.md instructs running the included Python scripts and editing provider endpoints. The runtime instructions and scripts operate on task text and local configuration files only; they do not read unrelated system files or environment variables, nor do they send data to external endpoints.

ℹ Install Mechanism

There is no remote installer; install.sh is included and copies the repo into a skills directory under the user's home (~/.clawdbot/skills or ~/clawd/skills) and creates ~/.model-router. No network downloads or archived extracts are performed. Note: install.sh copies all files from the current directory into the target, so if you run it from an unexpected location it will copy whatever is present.

ℹ Credentials

The skill declares no required env vars or credentials, which matches the code. The setup wizard stores endpoints and model names to ~/.model-router/config.json; the code does not prompt for API keys, but a user could place secrets into that config manually. The installer sets the config directory to mode 700, reducing exposure; however the config file's permissions follow the user's umask. Be cautious about storing API keys or tokens in plain config.json.

✓ Persistence & Privilege

The skill does not request always:true and does not modify other skills or global agent configs. Its persistence is limited to copying its own files into user skill dirs and creating ~/.model-router—reasonable for a skill that holds config.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install hermes-model-router
安装完成后，直接呼叫该 Skill 的名称或使用 /hermes-model-router 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

vv1.0.1

新增开发/全栈任务关键词识别，优化分类准确性

vv1.0.0

Initial release: LLM provider configuration skill for OpenClaw agents. Supports OpenRouter, OpenAI, Azure, Anthropic, Google AI, Ollama, and local models.

元数据

Slug hermes-model-router

版本 v1.0.1

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 2

常见问题