Hermes Agent Guide: seventeen markdown volumes and one PDF script, no agent code

There is no Hermes code in this repository, only a book

The first thing to settle is what the repository actually contains. The root holds seventeen markdown files numbered 00 through 16, a README, a LICENSE, `_COMMAND_MAP.md`, `_config.yml`, `generate_pdf.py` and a cover image named `封面.png`. Not one of those files is part of the Hermes Agent framework. The framework itself belongs to Nous Research, and the guide's own introduction gives the reason for a separate book: Hermes documentation is scattered across repositories and platforms, with no systematic Chinese guide in one place.

The author writes as 鲲鹏Talk, describing himself as an AI trend researcher and long-time practitioner of the Hermes project, and the book is signed April 2026. The last push to the default branch was 2026-09-14, and the repository has no GitHub releases at all, so there is no tag to compare against and nothing to install. Python shows up as the primary language only because of the PDF script at the end of the pipeline.

That makes this a documentation repository in the plain sense. Reading it tells you how someone who has used the framework for a while thinks its pieces fit together, and it tells you nothing about the current state of the code, which you have to get from the Hermes project itself.

The catalogue starts at zero, so the volume count and the file count differ

The headline claim is a book of 16 volumes and more than 300,000 characters of Chinese prose. The catalogue table underneath lists seventeen rows, because the numbering starts at 00 with a navigation volume called 目录与导读 and then runs through 16. So sixteen volumes of subject matter plus a table of contents file, which is worth knowing if you cite volume numbers.

The titles map out the argument of the book. Volume 01 covers why the book exists, five advantages of Hermes, and a glossary of more than 20 terms. Volume 02 is an industry view with a 2023 to 2026 history and a comparison of mainstream frameworks. Volume 03 traces the Nous Research background and the move from OpenClaw to Hermes. Volume 04 takes the core architecture apart layer by layer and describes a self-learning loop design.

The rest follows a familiar order: install and deploy in 05, first run and model access in 06, the memory system in 07, skills in 08, the built-in tools in 09, multi-platform access in 10, MCP and automation in 11, advanced patterns and case studies in 12, OpenClaw comparison and migration in 13, monetization paths in 14, community and resources in 15, and a trend forecast with roadmap reading in 16.

Four routes, four different hour counts, and gaps between them

Instead of one table of contents the book offers four reading routes, and the choice matters more than the order. The quickstart route is budgeted at 3 to 4 hours and takes five files with minute estimates attached: the preface at 30 minutes, the birth and evolution volume at 20, install and deploy at 60, basic usage at 60, and volume 12 as optional reading at 40. Its stated philosophy is to get it running first and understand it later.

The deep technical route is 15 to 20 hours across nine files and is the one aimed at people who will build on the framework. It adds the industry overview, the architecture volume, the three-layer memory system, the skills guide, the 47-tool reference, MCP and automation, and the OpenClaw comparison. Volume 12 is not on it, which is a hint that the advanced patterns and case studies sit outside the architecture track.

The monetization route runs 8 to 10 hours over eight files for entrepreneurs and independent developers, and the migration route runs 6 to 8 hours over eight files for existing OpenClaw users, built around a smooth transition rather than a clean break. Between them they skip volume 02, 10, 11 and 12, so no single route covers the whole book.

Five layers, 47 tools and three kinds of memory carry the technical half

Three volumes hold the weight of the technical case. Volume 04 walks a five-layer architecture and the self-learning loop that the author credits as an upgrade over OpenClaw. Volume 09 covers the 47 built-in tools in seven categories, each with its use case and a configuration example. Volume 07 explains the three-layer memory system, separating session memory, persistent memory and skill-level memory, with the configuration for each.

Volume 08 covers the skill lifecycle, automatic skill creation and a Skills Hub ecosystem, and volume 11 covers the MCP protocol with a stated library of more than 6,000 MCP services, plus Cron scheduled tasks and multi-agent orchestration. Volume 10 is about reaching the agent from other places, naming Feishu, WeChat, Discord and Slack among more than 15 platforms.

Read together they describe a framework whose selling points are memory, skills and automation rather than raw tool count. Volume 01 compresses that into a list of five advantages and a 20-plus term glossary, which is the fastest way to find out whether the vocabulary in the other volumes is worth your time.

Install is covered three ways, and picking a model is volume six

Volume 05 handles deployment with three routes, described as local, Docker and VPS, pitched at readers who get stuck on installation. Volume 06 covers the first run, conversation mode and model access configuration, which is where you decide which model the agent talks to. Volume 12 goes further out with sandboxed secure execution, voice interaction and what it calls complete enterprise cases.

The audience table is unusually specific about who each volume serves. A beginner gets a walkthrough that runs a first agent with no code. A developer gets the five-layer architecture, all 47 tools and a deep reading of the MCP protocol. An entrepreneur gets nine monetization paths with deployment cost analysis and income estimates. An investment or industry analyst gets a 12-dimension comparison and a 2026 to 2027 trend call. An enterprise decision maker gets integration options, security and compliance, and total cost of ownership.

Those are promises about coverage, and the honest test is whether the numbers survive a spot check. The 47 tools, the 6,000 MCP services and the 15-plus platforms are the kind of counts that go stale without a version stamp, and this repository carries none.

The only executable code here turns markdown into a PDF

The pipeline is two commands long:

pip install markdown weasyprint
python generate_pdf.py

The script writes one file, `Hermes-Agent白皮书-养马从入门到精通-鲲鹏Talk.pdf`, with the author name and the subtitle baked into the filename. Weasyprint is a real layout engine with a system font stack behind it, so a machine without the fonts the book expects will produce a PDF that looks wrong rather than one that fails loudly. `_config.yml` sits next to the script, which suggests the same markdown was also meant to be published as a site.

None of this touches Hermes. The PDF is a convenience for readers who want the book offline, and `generate_pdf.py` is the reason the repository is reported as Python when its real content is prose.

Non-commercial on the prose, MIT on the samples, and a monetization volume

The licensing is split, and the split is the one thing here a reader should read twice. The book content is CC BY-NC-SA 4.0, attribution, non-commercial and share alike. The example code is MIT, free for commercial or non-commercial projects. The repository's own license field reads NOASSERTION rather than naming either, so tooling that inspects the API will not tell you what applies to a given file.

The tension is not subtle. Volume 14 is 九大变现路径, nine monetization paths broken down with real cases and income estimates, sold as the route for entrepreneurs and independent developers. The methods in that volume are yours to use, since ideas are not the protected expression, but the text explaining them is not, so a course, a paid report or a commercial book built by copying the prose runs into the non-commercial clause. The MIT half only covers the sample code.

Everything else about the project reads as documentation rather than software, which is why the absence of releases, tags and a changelog matters less here than it would for a tool, and why the audience claims in the README deserve a check against the actual framework before you act on them.

Editorial conclusion

Read the Hermes Agent Guide if you want one author map of the framework before you commit to it, especially the architecture, memory, skills and OpenClaw migration volumes, and treat it as orientation rather than reference. Do not rely on it for facts about the current Hermes release, because the repository ships no agent code, no releases and no version pinning, and it is third-party work rather than Nous Research documentation. Check the license before you reuse anything: the prose is CC BY-NC-SA 4.0, which blocks commercial reuse of the text itself.

Frequently asked questions

Official sources

1. Issues
2. jwangkun/hermes-agent-guide on GitHub
3. README