Learn Hermes by operating it, verifying the result, and recovering from failure.
Table of Contents
Hermes Learning Lab is a Chinese interactive curriculum for learners new to Hermes Agent. It starts with the real Desktop and Feishu surfaces, then progresses through CLI usage, model switching, prompt contracts, Memory, Skills, automation, delegation, isolation, and recovery.
The course was last verified on 2026-08-24 against Hermes Agent v0.20.5 (v2026.8.19).
The course adapts the structure of microsoft/AI-For-Beginners: environment setup and diagnostics come first, followed by self-contained labs with explicit inputs, expected results, post-checks, and further reading. Hermes commands and configuration facts are grounded in the official documentation and official repository.
Note
This is an independent community education project, not an official Nous Research or Microsoft product.
| Hands-on learning | Evidence and safety |
|---|---|
| macOS, native Windows, WSL2, and Desktop setup routes | Setup, basic chat, and Doctor establish the first baseline |
| Official Hermes Desktop screenshot with interactive landmarks | Learners submit DESKTOP_OK; the course never reads window content |
| Feishu developer-console and client walkthroughs | Covers WebSocket, mentions, typing, approval cards, and error 200340 |
| One 13-lesson curriculum with three goal-based routes | A three-question diagnostic recommends a route without locking any lesson |
| Pre-check, four operational steps, and a real lab in every lesson | Every lab defines success criteria, evidence, and recovery actions |
| Choices, builders, and interface simulations | Lab verification plus the post-check forms two-part mastery evidence |
| Official references and per-lesson community readings | Community material never replaces official command verification |
| Prompt workbench, evidence rubric, recovery drills, and mastery report | localStorage v4, migration from v1-v3, and optional Supabase sync |
Pre-check
-> Essential context and core operations
-> Observable agent trace
-> Real-environment lab
-> Result checklist
-> Evidence capture and recovery rehearsal
-> Post-check
-> Official and community references
The course shows no estimated lesson durations. A lesson contributes to mastery only after both its real lab and post-check are verified.
Start in the route center. The route changes lesson priority, not course availability.
| Route | Best for | Core lessons |
|---|---|---|
| Quick Start | First working Hermes workflow | Install, Doctor, Prompt contracts, Skills, automation |
| Daily Use | Desktop and Feishu as a reliable personal assistant | Quick Start plus Agent loop, models, approvals, Sessions and Memory |
| Engineering | Auditable and recoverable Agent delivery | Tools, isolation, subagents, backup, evaluation and capstone |
Search and filter all 13 lessons from the sidebar at any time.
| Phase | Lessons | Evidence produced |
|---|---|---|
| I Setup and Foundations | 00 Download and Channels · 01 Environment and Doctor · 02 Agent Loop · 03 Model Switching and Recovery | Desktop/Feishu receipts, Doctor baseline, read-only tool evidence, model rollback record |
| II Reliable Interaction | 04 Prompt Contracts · 05 Tools/Context/Approval · 06 Sessions/Memory/SOUL | Validated prompt, minimal tool scope, redacted Memory |
| III Extensions and Automation | 07 Skill Install/Smoke Test · 08 Gateway/Cron/Hooks/Batch · 09 Parallel Tasks and Review | Skill review, idempotent automation, /agents closure evidence |
| IV Production Engineering | 10 Sandbox/Egress/ACP · 11 Backup/Update/Restore · 12 Capstone | Isolation negative tests, restore rehearsal, evaluation-driven workflow |
See CURRICULUM.md for prerequisites and mastery criteria, and docs/labs/README.md for the real-environment evidence template.
Open the GitHub Pages course, complete the three-question route diagnostic, and start from the recommended lesson. Online mode saves browser progress but never reads Hermes configuration or probes local processes without the separately started Companion.
Note
Before the first deployment, set Pages Source to GitHub Actions as described in GitHub publishing setup.
Requirements: Node.js 22+ and npm 10+.
git clone https://github.com/ChrysFu/hermes-learning-lab.git
cd hermes-learning-lab
npm ci
npm run devOpen the URL printed by Vite, normally http://127.0.0.1:5173/. If the port is busy, Vite reports the replacement port.
- Select a lesson and complete the ungraded pre-check.
- Follow the agent trace and four core operational steps.
- Run the real lab in a dedicated directory, test Profile, or Sandbox.
- Check every item under result verification against the actual outcome.
- Save the requested redacted evidence and confirm the recovery action is executable.
- Record the lab verification, then pass the post-check.
- Continue only after the progress rail confirms both forms of evidence.
The Practice tab contains three browser-only tools:
- Build a Prompt contract from task, allowed scope, expected result, constraints, and a stopping rule.
- Paste or import a redacted TXT/JSON receipt. The rubric checks version, Doctor, fixed receipt, runtime state, and recovery evidence; suspected secrets force a zero score.
- Rehearse failures such as an offline Companion, a silent Feishu bot, or an over-broad Skill write.
The browser does not execute the generated Prompt or upload the imported text. Evidence reports store the score and matched rule names, not the raw receipt.
The Report tab combines post-checks and verified labs into five mastery domains and can export a Markdown learning report. Local progress is always available and is migrated from earlier storage versions.
Optional cross-device sync uses Supabase Auth with GitHub OAuth or email magic links. Apply the included RLS migration and set the two public repository variables described in cloud sync setup. Without those variables, the deployed app clearly remains in local-only mode.
Lesson 00 includes Desktop and Feishu simulations for sessions, models, mentions, and tool approval. They never connect to a real Hermes runtime. Real verification requires the learner to run the prompt in Desktop or Feishu and explicitly submit a redacted receipt.
Only an explicit click invokes the local Vite probe. It checks whether the Hermes command exists, whether the Desktop process is running, and whether Gateway reports a usable state. It returns booleans only, never paths, configuration, secrets, logs, sessions, or messages.
Important
GitHub Pages explicitly shows static learning mode and hides the local-probe button. Simulations, receipt verification, resource links, and browser progress still work.
npm run lint
npm run build
npm run test:e2e
npm run previewUse npm run build:pages to verify the Pages base path. CI reruns the same Playwright flows under /hermes-learning-lab/.
The resources/ folder contains indexes, original summaries, access states, and source links only. It does not copy third-party PDFs, article bodies, or video transcripts.
| Material | Best for | Preview |
|---|---|---|
| Hermes Agent documentation | Current setup, Desktop, Feishu, Skills, Memory, Cron, security | Open docs |
| Official Hermes Feishu guide | App permissions, WebSocket, events, approval cards, troubleshooting | Open Feishu guide |
| NousResearch/hermes-agent | Current commands, releases, source, implementation boundaries | Open repository |
| Hermes Agent Orange Book 2.0 | Chinese overview of UI, memory, multi-agent workflows, and security | Preview Chinese PDF |
| hermes-agent-zh | Chinese setup, providers, automation, messaging, and FAQ | Open tutorial |
| awesome-hermes-agent | Skills, plugins, tools, surfaces, and advanced guides | Open index |
| Full beginner tutorial | Setup, Memory, Subagents, Cron, Skills, backup, firewall | Watch on YouTube |
| Ten beginner tips | Models, SOUL, Memory, Doctor, UI, migration, Skills | Watch on YouTube |
| Desktop release tutorial | Real Windows, macOS, and Linux Desktop UI | Watch on YouTube |
Browse the resource index, official references, community materials, and lesson map.
Note
Community videos are for observing interfaces and workflows. Installation commands, configuration keys, permissions, and security behavior are verified against the Hermes documentation and current release. The index excludes sources that require sign-in, are unavailable, or cannot be publicly verified.
The 2026-08-24 similar-project report compares Hermes Agent, Hermes Desktop, Hermes Workspace, Microsoft AI courses, Hugging Face Agents Course, and LangChain Academy using primary GitHub sources. It records adopted patterns, rejected patterns, repository snapshots, and exact source links.
The React workspace loads curriculum and resource data from src/data.js. learningEngine.js owns route recommendation, search, Prompt contracts, evidence scoring, mastery, migration, and merge rules. Browser localStorage is the default store; cloudSync.js dynamically loads Supabase only when deployment variables exist. The loopback Companion remains a separate, explicit read-only process. The editable diagram source is learning-system.drawio.
hermes-learning-lab/
├── .github/ # CI/Pages and contribution templates
├── src/ # React learning engine, curriculum data, styles
├── supabase/migrations/ # Optional progress table and RLS policies
├── resources/ # Electronic material indexes and lesson map
├── tests/e2e/ # Desktop, static-mode, and mobile Playwright checks
├── docs/labs/README.md # Real-environment lab guide
├── docs/TROUBLESHOOTING.md # Setup, UI, tooling, and recovery guide
├── docs/GITHUB-SETUP.md # Pages, Topics, and Discussions setup
├── docs/CLOUD-SYNC.md # Optional GitHub/email progress sync setup
├── docs/research/ # Dated primary-source comparison reports
├── docs/METRICS.md # Privacy-first measurement plan
├── docs/adr/ # Architecture decisions
├── assets/readme/ # README architecture asset and editable source
├── public/ui-reference/ # Attributed official UI reference
├── CURRICULUM.md # Curriculum and two-part mastery model
├── RESEARCH.md # Source research and evidence limits
├── CONTRIBUTING.md # Contribution and evidence rules
├── LICENSE # Dual code/content license
└── preview.png # Current product screenshot
npm run lint
npm run test:content
npm run build
npm run test:e2e
npm audit- Browser simulations never invoke a shell, modify
~/.hermes, install extensions, or connect messaging platforms. - The local probe runs only after an explicit click and returns three boolean status values.
- Real labs should use dedicated directories, isolated Profiles, Sandboxes, and one-time approvals.
- Receipts and reports must exclude tokens, app secrets, user IDs, chat history, and personal paths; suspected secrets fail evidence scoring.
- Supabase sync stores progress and redacted scores only. Prompt drafts, raw evidence, Companion status, and Hermes data stay local.
- Community commands must be checked against the current official documentation.
- The Pages build is exercised under
/hermes-learning-lab/to catch blank deployments and asset 404s.
See ADR 0001, ADR 0002, and ADR 0003 for the implementation boundaries.
Read CONTRIBUTING.md before submitting changes. See GitHub setup for Pages, Topics, and Discussion routing, and course metrics for the privacy boundary.
- AI-For-Beginners: setup, lab, quiz, and further-reading structure.
- Hermes Agent: runtime, commands, and UI facts.
- Hermes Agent v0.20.5 /
v2026.8.19: current verification baseline. - Similar-project research: primary-source comparison and adoption decisions.
- Hermes 4.3 model card: model, prompt, serving, and sampling references.
- RESEARCH.md: research findings, adoption decisions, and version risks.
Project code is available under the MIT License. Original lessons, labs, research summaries, and project-created diagrams use CC BY-NC-SA 4.0, which permits attributed non-commercial sharing and translation under the same license. Third-party screenshots, trademarks, model cards, articles, videos, and PDFs are excluded from this grant.
Hermes Learning Lab 是一个面向零基础学习者的 Hermes Agent 中文交互课程。课程从 Desktop 和飞书真实界面开始，再逐步进入 CLI、模型切换、Prompt 契约、Memory、Skills、自动化、并行任务、安全隔离与恢复。
课程最后核验于 2026-08-24，版本基线为 Hermes Agent v0.20.5（v2026.8.19）。
课程结构参考 microsoft/AI-For-Beginners：先准备环境和诊断基础，再通过独立实验、明确输入、预期结果、课后检查与延伸资料形成学习闭环。Hermes 的命令和配置事实以官方文档及官方仓库为准。
Note
这是独立社区教学项目，不是 Nous Research 或 Microsoft 的官方产品。
| 操作学习 | 结果与安全 |
|---|---|
| macOS、Windows 原生、WSL2 与 Desktop 安装路径 | 安装后先完成 Setup、普通聊天和 Doctor 基线 |
| Hermes Desktop 官方实景图与可点击地标 | 用户主动粘贴 DESKTOP_OK 回执，课程不读取窗口内容 |
| 飞书控制台与客户端双路径指南 | 覆盖 WebSocket、@提及、Typing、审批卡和 200340 排错 |
| 同一套 13 课、三条目标路线 | 三问诊断推荐路线，但不锁定任何课程 |
| 每课有课前诊断、四步操作和真实实验 | 每项实验都给出成功标准、证据与恢复动作 |
| 单选、多选、配置构建器和模拟界面 | 实验验收与课后检查组成双证据掌握度 |
| 官方资料与社区延伸阅读按课匹配 | 社区内容不替代官方命令核验 |
| Prompt 工作台、证据评分、恢复演练和掌握报告 | localStorage v4、兼容迁移 v1-v3，并可选 Supabase 同步 |
每课都执行同一套可验证流程：
课前诊断
-> 必要说明与核心操作
-> Agent 行动轨迹
-> 真实环境实验
-> 逐项结果核验
-> 保存证据与演练恢复
-> 课后检查
-> 官方资料与社区延伸阅读
课程不会显示学习时长估计。完成课后题不等于完成真实操作；只有实验验收和课后检查都通过，本课才计入掌握度。
首次进入路线中心，路线只改变优先顺序，不限制课程访问。
| 路线 | 适合人群 | 核心内容 |
|---|---|---|
| 快速上手 | 第一次建立可工作的 Hermes | 安装、Doctor、Prompt 契约、Skills、自动化 |
| 日常使用 | 通过 Desktop 与飞书构建稳定个人助手 | 快速上手 + Agent Loop、模型、审批、Session 与 Memory |
| 工程进阶 | 交付可审计、可恢复的 Agent 工作流 | 工具、隔离、子任务、备份、评测与毕业项目 |
左侧课程导航支持搜索，并可按任一条路线筛选全部 13 课。
| 阶段 | 课程 | 交付结果 |
|---|---|---|
| I 启动与基础 | 00 下载与多端接入 · 01 环境与诊断 · 02 Agent Loop · 03 模型切换与回退 | Desktop/飞书回执、Doctor 基线、只读工具证据、模型回切记录 |
| II 可靠交互 | 04 Prompt 契约 · 05 工具/Context/审批 · 06 Session/Memory/SOUL | 可校验 Prompt、最小工具范围、脱敏 Memory |
| III 扩展与自动化 | 07 Skill 安装/冒烟测试 · 08 Gateway/Cron/Hooks/Batch · 09 并行任务与收口 | Skill 审查记录、幂等自动化、/agents 收口证据 |
| IV 工程化与进阶 | 10 Sandbox/Egress/ACP · 11 备份/更新/恢复 · 12 毕业项目 | 隔离负向测试、恢复演练、评测驱动工作流 |
完整先修关系和掌握标准见 CURRICULUM.md，真实环境记录模板见 docs/labs/README.md。
打开 GitHub Pages 课程，完成三问路线诊断，再从推荐课程开始。在线模式保存浏览器进度；只有另行启动并配对 Companion 后，才允许用户主动发起本机只读检测。
Note
仓库首次部署前，需要按 GitHub 发布设置 把 Pages Source 设为 GitHub Actions。
环境要求：Node.js 22+、npm 10+。
git clone https://github.com/ChrysFu/hermes-learning-lab.git
cd hermes-learning-lab
npm ci
npm run dev打开终端显示的地址；默认是 http://127.0.0.1:5173/。如果该端口已占用，Vite 会显示实际使用的新端口。
- 从左侧选择课程，先完成不计分的课前诊断。
- 按顺序查看 Agent 轨迹和四个核心操作步骤。
- 在专用练习目录、测试 Profile 或 Sandbox 中执行真实实验。
- 对照“操作结果检验”逐项核对成功标准。
- 保存课程指定的脱敏证据，并确认失败恢复动作可执行。
- 点击“确认结果已核验”，再通过课后检查。
- 当右侧显示“本课已取得实验与课后检查双证据”后进入下一课。
第 00 课包含 Hermes Desktop 和飞书端模拟。模拟练习用于熟悉会话、模型、@提及和工具审批，不会连接真实 Hermes。真实界面验收需要学习者在 Desktop 或飞书中主动执行校验 Prompt，再把脱敏回执粘贴回课程。
GitHub Pages 不能直接读取你的电脑。需要在自己的 macOS 或 Windows 终端启动一个只绑定 127.0.0.1:43127 的伴随服务，网页通过一次性配对码连接；每次检测前服务都会在终端询问是否允许。
npm ci
npm run local:bridge保持终端运行，复制终端显示的配对码，在课程的“本机只读检测”区域输入并点击“配对”。配对后点击“检测本机状态”，回到终端输入 y 允许本次读取。服务只返回 Hermes 是否安装、Desktop/Gateway 是否运行、版本和 Doctor 摘要，不返回路径、配置、密钥、日志、会话或消息。
在课程页面下载 macOS 启动脚本 或 Windows PowerShell 脚本，先阅读源码，再由用户手动执行。脚本不会静默安装 Node.js 或 Hermes；电脑需要 Node.js 22+。服务未连接时，静态课程模拟和 DESKTOP_OK / FEISHU_OK 回执验收仍可继续。
Important
不要把配对码、服务状态截图或本机伴随服务目录提交到 Git。完成练习后可点击“解除配对”，或在运行服务的终端按 Ctrl+C 停止。
npm run lint
npm run build
npm run test:e2e
npm run previewPages 子路径验证使用 npm run build:pages，CI 会在 /hermes-learning-lab/ 下重复运行同一组 Playwright 流程。
“练习台”包含三个完全在浏览器运行的工具：
- 根据任务、允许范围、验收结果和约束生成带停止条件的 Prompt 契约。
- 粘贴或导入脱敏 TXT/JSON 结果，检查版本、Doctor、固定回执、运行状态和恢复记录；发现疑似密钥时评分直接归零。
- 演练 Companion 离线、飞书不回复和 Skill 越界写入等故障的最短恢复路径。
网页不会执行生成的 Prompt，也不会上传导入的原始文本。证据报告只保存分数和命中的规则名称。
“报告”页面把课后检查和真实实验验收合并为五类掌握度，并支持导出 Markdown 学习报告。本地进度始终可用，也会自动迁移旧版本记录。
可选跨设备同步使用 Supabase Auth，同时支持 GitHub OAuth 和邮箱魔法链接。按云同步配置应用 RLS 迁移并配置两个公开仓库变量后即可启用；变量不存在时，线上应用会明确保持本地模式，不影响任何课程。
仓库的 resources/ 目录只保存分类索引、原创摘要、访问状态和来源链接，不复制第三方 PDF、文章正文或视频字幕。
| 资料 | 适合学习 | 在线预览 |
|---|---|---|
| Hermes Agent 官方文档 | 当前安装、Desktop、飞书、Skills、Memory、Cron、安全与排错 | 打开文档 |
| Hermes 飞书官方接入指南 | 应用权限、WebSocket、事件订阅、审批卡与排错 | 打开飞书指南 |
| NousResearch/hermes-agent | 当前命令、Release、源码与实现边界 | 打开仓库 |
| Hermes Agent 橙皮书 2.0 | 中文整体理解、界面、记忆、多 Agent 与安全 | 预览中文版 PDF |
| hermes-agent-zh | 中文安装、Provider、自动化、IM 和 FAQ | 打开教程 |
| awesome-hermes-agent | Skills、插件、工具、界面与进阶指南 | 打开索引 |
| Hermes Agent 保姆级教学 | 安装、Memory、Subagent、Cron、Skills、备份和防火墙演示 | 观看 YouTube |
| Hermes Agent 新手使用十大技巧 | 模型、SOUL、Memory、Doctor、UI、迁移和 Skills | 观看 YouTube |
| Hermes Agent 桌面版教程 | Windows、macOS、Linux Desktop 真实界面 | 观看 YouTube |
更多访问状态与课程映射：
Note
社区视频用于观察界面与工作流；安装命令、配置键、权限和安全行为均以 Hermes 官方文档及当前 Release 为准。资料索引不收录需要登录、已失效或无法公开验证的入口。
2026-08-24 同类项目调研报告基于 GitHub 一手资料，对比 Hermes Agent、Hermes Desktop、Hermes Workspace、Microsoft AI 课程、Hugging Face Agents Course 与 LangChain Academy，并记录采用项、拒绝项、仓库快照和精确来源链接。
React 工作区从 src/data.js 读取课程和资料。learningEngine.js 负责路线推荐、搜索、Prompt 契约、证据评分、掌握度、迁移和合并规则。浏览器 localStorage 是默认存储；只有部署变量齐全时，cloudSync.js 才动态加载 Supabase。本机 Companion 仍是独立、显式确认的 loopback 只读进程。可编辑图源见 learning-system.drawio。
hermes-learning-lab/
├── .github/ # CI/Pages、PR 模板与 Discussion 入口
├── src/
│ ├── App.jsx # 学习引擎、实验验收、进度迁移与视图
│ ├── learningEngine.js # 路线、搜索、Prompt、评分与掌握度规则
│ ├── learningViews.jsx # 路线中心、练习台、报告与同步界面
│ ├── cloudSync.js # 可选 Supabase Auth 和进度仓库
│ ├── data.js # 13 课、资料映射、安装与界面数据
│ └── styles.css # 三栏工作台与响应式样式
├── resources/
│ ├── README.md # 电子资料总索引
│ ├── OFFICIAL.md # 官方资料与命令核验入口
│ ├── COMMUNITY.md # 社区电子书、文章与视频
│ └── LESSON-MAP.md # 逐课资料映射
├── docs/
│ ├── labs/README.md # 真实环境实验手册
│ ├── TROUBLESHOOTING.md # 安装、界面、工具和恢复排错
│ ├── GITHUB-SETUP.md # Pages、Topics 与 Discussions 设置
│ ├── CLOUD-SYNC.md # GitHub/邮箱进度同步配置
│ ├── research/ # 按日期保存的一手资料调研
│ ├── METRICS.md # 隐私优先的课程衡量方案
│ └── adr/ # 架构决策记录
├── tests/e2e/ # Playwright 桌面、静态与移动端验收
├── supabase/migrations/ # 可选进度表和 RLS 策略
├── assets/readme/ # README 架构图与可编辑图源
├── public/ui-reference/ # 有来源说明的官方界面参考图
├── CURRICULUM.md # 课程体系和双证据掌握标准
├── RESEARCH.md # 资料研究、采用边界和版本说明
├── CONTRIBUTING.md # 贡献流程与证据标准
├── LICENSE # 代码/课程双许可证
└── preview.png # 当前产品界面截图
npm run lint
npm run test:content
npm run build
npm run test:e2e
npm audit- 浏览器模拟不会调用 Shell、修改
~/.hermes、安装扩展或连接消息平台。 - 本机探针只在明确点击后执行，并只返回 Hermes、Desktop、Gateway 三类布尔状态。
- 真实实验建议使用专用目录、隔离 Profile、Sandbox 和单次审批。
- 回执和实验记录不得包含 Token、App Secret、用户 ID、聊天历史或个人路径。
- 证据评分检测到疑似敏感字段会归零；云端只同步进度和脱敏分数，不同步 Prompt、原始证据或 Hermes 数据。
- 社区资料中的命令必须回到当前官方文档复核。
- GitHub Pages 构建会验证
/hermes-learning-lab/子路径，避免部署后出现空白页或资源 404。
设计取舍见 ADR 0001、ADR 0002 和 ADR 0003。
提交修改前请阅读 CONTRIBUTING.md。Pages、Topics 与 Discussion 设置见 docs/GITHUB-SETUP.md，衡量指标和隐私边界见 docs/METRICS.md。
- AI-For-Beginners：课程 Setup、实验、测验和延伸阅读结构参考。
- Hermes Agent：运行时、命令和界面事实源。
- Hermes Agent v0.20.5 /
v2026.8.19：当前课程核验基线。 - 同类项目调研：一手资料对比和采用决策。
- Hermes 4.3 model card：模型、Prompt、Serving 与采样资料。
- RESEARCH.md：来源核验、采用决策和版本风险。
项目代码采用 MIT License；原创课程文字、实验、研究摘要和项目自制图表采用 CC BY-NC-SA 4.0，允许署名、非商业转载与翻译，并要求衍生内容使用相同许可。第三方截图、商标、模型卡、文章、视频和 PDF 不在本项目再授权范围内。