title: From Zero to Hermes Agent in 3 Days — An Honest Beginner's Journey

查看原文 / View Original 编辑 / Edit

From Zero to Hermes Agent in 3 Days — An Honest Beginner’s Journey

从零到 Hermes Agent：三天新手实战之旅

Hermes Agent Challenge Submission: Write About Hermes Agent. This is the honest version nobody tells you. Most tutorials start with a clean Linux machine, a working API key, and someone who already knows what they’re doing. This is not that tutorial. This is the story of building a real project with Hermes Agent starting from a Windows PC with virtualisation disabled in the BIOS, zero Linux environment, and a deadline three days away. I built repo-audit-agent — a tool that uses Hermes Agent to perform first-pass technical reviews of GitHub repositories. Here’s what I actually learned along the way.

Hermes Agent 挑战赛投稿：关于 Hermes Agent 的真实体验。这是没人会告诉你的“大实话”版本。大多数教程都从一台干净的 Linux 机器、一个有效的 API Key 以及一个经验丰富的开发者开始。但这篇教程不是这样。这是一个关于如何从一台 BIOS 禁用了虚拟化功能、没有任何 Linux 环境的 Windows PC 出发，在三天截止日期前用 Hermes Agent 构建真实项目的故事。我开发了 repo-audit-agent——一个利用 Hermes Agent 对 GitHub 仓库进行初步技术审查的工具。以下是我在此过程中学到的真实经验。

Day 1: The Environment Problem

第一天：环境难题

Before I could even think about Hermes Agent, I needed a Linux environment. My Windows machine had WSL2 partially configured but virtualisation was disabled at the BIOS level — a common corporate lockdown. So WSL2 was out. I tried:

  • Google Colab — works, but resets every session. You reinstall everything every time you open it.
  • Replit — changed their UI completely, now pushes everyone toward their own AI agent.
  • CloudShell on AWS — only 1GB of disk space. Hermes Agent needs more.

在考虑 Hermes Agent 之前，我首先需要一个 Linux 环境。我的 Windows 机器虽然配置了部分 WSL2，但 BIOS 禁用了虚拟化功能——这是常见的企业级限制。所以 WSL2 行不通。我尝试了：

  • Google Colab：能用，但每次会话都会重置，每次打开都要重新安装所有东西。
  • Replit：界面大改，现在强制推行他们自己的 AI Agent。
  • AWS CloudShell：只有 1GB 磁盘空间，Hermes Agent 需要更多。

What finally worked: AWS EC2 t2.micro free tier. A real Ubuntu server in the cloud, persistent across sessions, 6.6GB of disk, always available. Once I understood this was the right environment, everything else followed. Lesson learned: Don’t fight your local environment. Cloud VMs exist precisely for this.

最终成功的方案是：AWS EC2 t2.micro 免费层级。这是一个云端的真实 Ubuntu 服务器，会话持久化，拥有 6.6GB 磁盘空间，且随时可用。一旦确定了这是正确的环境，后续一切都顺理成章了。经验教训：不要在本地环境上死磕，云端虚拟机正是为此而生的。

Day 2: Installing Hermes Agent

第二天：安装 Hermes Agent

Once I had an EC2 instance running Ubuntu Server, installation was one command: curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

一旦我的 EC2 实例运行起 Ubuntu Server，安装只需一条命令： curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

The installer is genuinely impressive. It:

  • Detects your OS and installs the right Python version via uv
  • Clones the repository
  • Installs all dependencies
  • Sets up 90 bundled skills
  • Installs Playwright for browser automation
  • Links the hermes command to your PATH

这个安装程序确实令人印象深刻。它能：

  • 检测操作系统并通过 uv 安装正确的 Python 版本
  • 克隆仓库
  • 安装所有依赖项
  • 配置 90 个内置技能
  • 安装用于浏览器自动化的 Playwright
  • 将 hermes 命令链接到你的 PATH 环境变量中

The full output is about 300 lines. When you see ✓ Installation Complete! you know it worked.

完整的输出大约有 300 行。当你看到 ✓ Installation Complete! 时，你就知道安装成功了。

Configuring the LLM Provider

配置大模型提供商

This is where I spent most of Day 2. Hermes Agent supports many providers: Anthropic, OpenRouter, Google AI Studio, Groq, DeepSeek, and many more. The setup wizard is interactive and well-designed. I tried several providers:

  • OpenRouter — ran out of credits immediately.
  • Regolo.ai — Italian provider, interesting, but gave 403 errors I couldn’t resolve.
  • Gemini free tier — worked, but has a hard limit of 5 requests/minute.

这是我第二天花费时间最多的地方。Hermes Agent 支持许多提供商：Anthropic、OpenRouter、Google AI Studio、Groq、DeepSeek 等等。设置向导交互友好且设计精良。我尝试了几个提供商：

  • OpenRouter：额度瞬间用光。
  • Regolo.ai：意大利的提供商，很有趣，但出现了无法解决的 403 错误。
  • Gemini 免费层级：成功运行，但有每分钟 5 次请求的硬性限制。

What I learned about Gemini free tier: It works, but you need to wait between calls. For interactive development this is manageable. For batch processing, you’d want a paid tier or a different provider. The setup wizard command is simply: hermes setup. It walks you through provider selection, API key configuration, and model selection. For Gemini, I selected gemini-2.5-flash — fast, capable, and free.

关于 Gemini 免费层级的经验：它能用，但调用之间需要等待。对于交互式开发来说这尚可接受，但如果是批量处理，你还是需要付费层级或更换提供商。设置向导命令很简单：hermes setup。它会引导你完成提供商选择、API Key 配置和模型选择。对于 Gemini，我选择了 gemini-2.5-flash——速度快、能力强且免费。

First Successful Call

第一次成功调用

After configuration, this command produced a real response: hermes chat -q "Say hello in one sentence." -Q Output: Hello! I am Hermes, a CLI AI Agent. That single line, after two days of environment struggles, felt like a genuine milestone.

配置完成后，这条命令产生了真实的响应： hermes chat -q "Say hello in one sentence." -Q 输出：Hello! I am Hermes, a CLI AI Agent. 在经历了整整两天的环境折腾后，看到这一行字，感觉就像是一个真正的里程碑。

Day 3: Building the Actual Tool

第三天：构建实际工具

With Hermes Agent working, I built repo-audit-agent — a Python script that wraps Hermes Agent and produces structured technical audit reports. The core insight is simple: Hermes Agent is not just a chatbot. It uses tools. When you give Hermes Agent a task that requires browsing the web, it actually opens a browser, navigates to the URL, and reads the content. This is what makes it an agent rather than a language model wrapper.

在 Hermes Agent 运行正常后，我构建了 repo-audit-agent——一个封装了 Hermes Agent 并生成结构化技术审计报告的 Python 脚本。核心洞察很简单：Hermes Agent 不仅仅是一个聊天机器人，它会使用工具。当你给它一个需要浏览网页的任务时，它真的会打开浏览器、导航到 URL 并读取内容。这就是它被称为 Agent 而非仅仅是语言模型封装器的原因。

Here’s how I invoke Hermes Agent from Python:

def run_hermes_audit(repo_url: str, max_turns: int = 15) -> str:
    prompt = build_audit_prompt(repo_url, owner, repo_name)
    result = subprocess.run(
        [
            "hermes", "chat", "--query", prompt, "--quiet", 
            "--max-turns", str(max_turns),
        ],
        capture_output=True, text=True, timeout=300,
    )
    return result.stdout

以下是我在 Python 中调用 Hermes Agent 的方式： （代码见上文）

The --max-turns parameter is important: it controls how many tool-calling iterations Hermes Agent can perform. For a repository audit, 15 turns gives it enough room to browse the repo, read the README, analyze the file structure, and generate the report.

--max-turns 参数很重要：它控制 Hermes Agent 可以执行多少次工具调用迭代。对于仓库审计，15 次迭代足以让它浏览仓库、阅读 README、分析文件结构并生成报告。

What Hermes Agent Actually Does During an Audit

Hermes Agent 在审计期间到底做了什么

When I run python3 audit.py https://github.com/NousResearch/hermes-agent, Hermes Agent:

  • Plans — breaks the audit task into sub-steps autonomously
  • Browses — navigates to the GitHub repository using its browser tool
  • Reads — fetches the README, file tree, and visible codebase information
  • Reasons — synthesizes findings into quality observations and risk assessments
  • Generates — produces a structured Markdown report following my template

当我运行 python3 audit.py https://github.com/NousResearch/hermes-agent 时，Hermes Agent 会：

  • 规划：自主将审计任务拆解为子步骤
  • 浏览：使用浏览器工具导航到 GitHub 仓库
  • 读取：获取 README、文件树和可见的代码库信息
  • 推理：将发现综合为质量观察和风险评估
  • 生成：按照我的模板生成结构化的 Markdown 报告

This is genuine agentic behavior. I’m not sending the repository content to a language model — Hermes Agent fetches it itself.

这是真正的 Agent 行为。我并没有把仓库内容发送给语言模型——而是 Hermes Agent 自己去获取的。

The Output

输出结果

Here’s a real excerpt from a report generated on the Hermes Agent repository: (See Markdown table in original text)

这是针对 Hermes Agent 仓库生成的报告中的真实摘录： （见原文 Markdown 表格）

This is a first-pass review, not a replacement for manual code review. But as a starting point for human analysis, it is genuinely useful.

这只是初步审查，不能替代人工代码审查。但作为人类分析的起点，它确实非常有用。

What I Think About Hermes Agent

我对 Hermes Agent 的看法

After three days of building with it, here are my honest observations: What Works Well The installation is excellent. One command, handles everything, works on stock Ubuntu. This is the standard all CLI tools should meet. The provider ecosystem is broad. If one provider has rate limits…

在用它构建了三天之后，这是我的真实观察： 做得好的地方 安装体验极佳。一条命令搞定一切，在原生 Ubuntu 上运行良好。这是所有 CLI 工具都应该达到的标准。提供商生态系统非常广泛。如果某个提供商有限制……

← 返回新闻列表 / Back to news list