Hermes 中文社区

HermesCN 是围绕 Hermes Agent @NousResearch 建设的中文社区与协作网络。

我们持续沉淀中文文档、教程、实践案例、Skills 和自动化工作流，为个人提供从学习、实践到职业发展的成长路径。同时连接社区成员与真实业务需求，通过项目协作，为企业提供 Agent 落地的参考与解决方案。

[Hermes 中文社区 @hermescn_org]

欢迎加入 HermesCN 社区，一起学习、探索、实践 Hermes Agent！

[brucexu.eth @brucexu_eth]

https://t.co/c157nT0nM5

[Hermes 中文社区 @hermescn_org]

最近 Hermes Agent 在疯狂做减法。源码几乎砍掉了三分之一。

一些非核心的组件也在插件化。

未来 Hermes 的目标是保持核心功能的稳定，然后通过插件来为 Hermes 提供扩展性。

[Hermes 中文社区 @hermescn_org]

社区成员给 Hermes Agent 写了一个上下文引擎插件，用来处理长对话里细节丢失的问题。

它把一次会话的每条消息存进本地 SQLite，把较早的对话折叠成分层摘要，实时提示词由这些摘要加上最近几条消息组成。

Agent 因此拿到自己的检索工具，需要引用很久以前某一句时，可以把那一页取回来。

插件接的是 Hermes 在配置里开放的上下文引擎插槽，适合长时间跑任务的场景。

https://t.co/cYeoJKcZZV

[witcheer @witcheer]

a Hermes Agent community member built a context engine plugin for long conversations:

it keeps every message of a session in a local SQLite store, folds the older turns into layered summaries, and builds the live prompt from those summaries plus the most recent messages.

the agent gets its own recall tools, so when it needs an exact line from a while ago it pulls that page back.

it plugs into the context engine slot Hermes exposes in config. for anyone who runs Hermes on long jobs and wants the earlier detail within reach.

https://t.co/2hrVqDBLJV

[Hermes 中文社区 @hermescn_org]

有人提议用 Jev 做即时上下文压缩，取代摘要式压缩。

Hermes Agent 的公开压缩评测给出了不同结果。这套做法实际等价于一条程序化规则，不需要模型参与，直接把工具调用从历史里删掉，零成本可做。

问题是会形成循环，每压一次可删的工具调用更少，能省下的 token 更少，最后无处可压，压缩会卡住。

每次压缩还会打断缓存，输入 token 单价可能高出约十倍，压缩后的基线成本也更高。

完整评测已公开，可以自己跑一遍两套方案的对比。

https://t.co/Wv6ji656Qi

[Teknium 🪽 @Teknium]

Okay, everyone wants us to give the unbiased facts.

Jev compaction here has a lot of problems. The biggest problems:

- When I ran it on our public compaction eval (that you can run too), it resulted in a programmatic rule, that you dont need any Jev or other model for - it simply removed all tool calls from the chat history
- This means you could do it for free, first of all, but second of all, it means you run into a viscious cycle.
Every compaction, there are less and less tool calls to remove from the chat history, meaning it compacts less and less tokens, until there is no room for compaction, and you hit a hard stop, and can't compact anymore.
- And, if you are cycling like this, every compaction breaks your cache, so you're paying 10x the price on input tokens, and keeping more input tokens each round means a higher baseline cost after compaction as well.

Here's the full reproducible eval that you can run against Hermes Agent between the two (Hermes' vs Jev's):

https://t.co/0vJV5u10Cn

[Hermes 中文社区 @hermescn_org]

社区成员给 Hermes Agent 做了一个只读的本地网页，把一次安装里积累下来的内容集中摊开。

技能和每个技能所在的位置、历史对话与搜索、定时任务、每个会话花掉多少、记忆、插件和健康状态，都直接从本地的文件和数据库里读出来。

因为它只读不改，你可以一直开着页面，同时让 Agent 照常干活。

https://t.co/Y5FFS0vIqm

[witcheer @witcheer]

a Hermes Agent community member built a read-only web page for everything a Hermes install keeps:

your skills and where each one lives, past conversations with search, the cron schedule, what each session cost, memory, plugins and health, read straight from the files and databases Hermes keeps.

it only reads them, so you can leave it open while the agent is working.

https://t.co/6AbOCeRhWr

[Hermes 中文社区 @hermescn_org]

Hermes 接下来的方向是核心更精简，内置功能少打包一些，更多交给插件。

第一步是把内置的记忆 provider 全部移出主仓库，改由各自的作者在自己的仓库维护，之后进插件市场上架。对用户来说，以后换记忆方案更像装插件，而不是等官方更新。

官方也说明，这不代表 Hermes 要放弃个人助理的定位，或者变成纯编码 agent。

https://t.co/G4PWE93nnc

[Teknium 🪽 @Teknium]

We are going to lean into making Hermes more like Pi, and less like OpenClaw

[Hermes 中文社区 @hermescn_org]

Hermes Agent 团队正在收集多人共用一个 Agent 的实际做法，原文列了三种常见方式：同一台机器上每人一个 profile；用一个 gateway 同时服务多个人；用 Hermes Business 的团队余额，给每位成员各托管一个 Agent。也欢迎说说别的做法，比如你自己跑的是哪种、一共带几个人。

https://t.co/36EIvYbc1k

[witcheer @witcheer]

I'm trying to get a read on how people run Hermes Agent for more than one person today, at home or at work.

- one profile per person on the same box
- one gateway serving several people
- a hosted agent each on a team balance with Hermes Business
- something else?

tell me what you run and for how many people.

[Hermes 中文社区 @hermescn_org]

Hermes Desktop 的多标签会话：多个标签页各开一个独立会话，共用一个输入框——光标停在哪个会话上，输入就发给哪个，不用再为切换会话换输入框。

witcheer 想要的桌面端就是这样：几个 agent 同时开工，自己盯哪个就管哪个。

https://t.co/dCdXV5ZcUi

[witcheer @witcheer]

several sessions open side by side and one place to type: the session under your cursor is the one that hears you.

this is the Hermes Desktop I want to work in, a few agents busy at once and me steering whichever one I am looking at.

[Hermes 中文社区 @hermescn_org]

Hermes Agent 有了插件目录：入口在内置的 Plugins 页面，首批上架 4 个官方插件和 96 个社区插件，覆盖桌面端模组、新聊天平台、浏览工具等。

每个社区插件都由 Nous 团队审核后才出现，之后持续补充。想提交自己的插件，清单见原帖里的文档链接。

https://t.co/0ifFnFOXmM

[witcheer @witcheer]

this is the one a lot of you have been asking for: a plugin catalog inside Hermes Agent!

one place to find plugins, and a Nous review on every entry before they are listed.

want your agent on another chat platform, a new pane in Hermes Desktop, a browser tool, an integration for your niche? open the Plugins page, pick it, add it to your agent.

and if you have been building on Hermes Agent, this is your moment: a plugin you publish today can be a feature everyone installs tomorrow. the checklist is here:

https://t.co/WFEw6tRGp7

[Hermes 中文社区 @hermescn_org]

让 Hermes Agent 在每次会话开始前先自己熟悉一遍项目：这段提示词会替你写出 AGENTS.md。

在仓库里打开一个会话，把提示词粘进去。它会先读 README、构建与测试文件、目录结构和最近的提交，再问仓库没能告诉它的问题，最多三个，然后起草 AGENTS.md；这个文件会在该目录的每次会话开始时被 Hermes 加载。

草稿会整份先给你看，你点头它才保存；拿不到你的答复就什么都不写。

提示词在原帖首条回复里，可以直接复制使用。

https://t.co/WSIyVMP7jS

[witcheer @witcheer]

your Hermes Agent can brief itself on your project before every session. this prompt writes the file that does it.

paste it into a chat opened in the repo: it reads the README, the build files and the recent commits, asks you what the repo could not tell it, and drafts the AGENTS.md that Hermes loads at the start of every session in that folder.

copy / paste prompt in the first comment.

[Hermes 中文社区 @hermescn_org]

Hermes 技巧 #74：provider_routing

在 OpenRouter 上，同一个模型名由多家托管方提供服务，每次请求由 OpenRouter 选一家。

如果某家正在限流，可以在 config.yaml 用 provider_routing 把它写进 ignore，该模型的请求就不再发往它，即使它最便宜；其他托管方照常服务。

Nous Portal 会替你为每个模型选托管方，无需配置。

文档与提示词见原帖首条回复。

https://t.co/QGUcY6GVJR

[witcheer @witcheer]

Hermes Wingtips #74: `provider_routing`

on OpenRouter, one model name is served by several hosts, and OpenRouter picks the host for each request.

but what if one of those hosts is rate-limiting you today?

`provider_routing` in config.yaml lets you choose.

list a host under `ignore` and your requests for that model stop going there, even when it is the cheapest one. the other hosts keep serving as before.

on Nous Portal you can skip this step: the Portal picks the host for every model on your behalf, so there is nothing to set.

[Hermes 中文社区 @hermescn_org]

清理百万行 Python：1393 个子代理跑了 19 小时，代码库缩小 34.4%。这次跑法没有只属于 Teknium 环境的特殊设置。

/goal 让目标一直保持有效，代理不会中途停下等指令。

干活的是 delegate_task 一个工具，每个子任务在自己的 git worktree 里，不会互相覆盖。

标准来自代理自己写的 skill，靠一次次纠正慢慢积累。

这些都在今天可直接安装的 Hermes Agent 里。

https://t.co/auTd37YwaK

[witcheer @witcheer]

one thing to keep from that: nothing in that run is special to Teknium's setup.

`/goal` kept the objective alive when the agent would otherwise stop and wait.

the workers came from one tool, `delegate_task`, each in its own git worktree so they never wrote over each other.

and the standards they followed came from a skill his agent had written for itself, one correction at a time.

all of it is in the Hermes Agent you install today.

go read the blog it is super interesting :)

[Hermes 中文社区 @hermescn_org]

四个 Hermes Autopilot 提示词汇总：
https://t.co/YaumDhbF11

Autopilot 是粘进新 Hermes 会话的提示词，只在 agent 里可用，能读记忆与历史会话。

用途：记录工作方式到记忆；把任务转成可复用 skill；用 hermes doctor 体检给出三处改进；按上周会话写周一回顾并自行排期。

都在真实 Hermes Agent 上跑过。

https://t.co/s5kP1BOseM

[witcheer @witcheer]

I put every Hermes Autopilot prompt on one page.

https://t.co/ulCgaImY0r

an Autopilot is a prompt you paste into a fresh Hermes Agent chat, and it only works inside an agent: it reads your memory, your past sessions or your own config, things a chat window cannot see

four so far:
- one interviews you and saves how you work to memory
- one turns the task you just finished into a reusable skill
- one audits your own setup with hermes doctor and hands you the three changes worth making
- one writes your Monday review from last week's sessions and schedules itself

each one ran on a real Hermes Agent box before it was posted. copy the text, paste it, answer what it asks.

[Hermes 中文社区 @hermescn_org]

Hermes Business 问答：

数据：每团队在 Nous 基础设施独立租户；Enterprise 可自部署。

技能库：成员的 skill 进团队库，所有 agent 都能用。

余额：团队余额与个人分开，邮箱邀请，每人一个托管 agent 和消费上限。

渠道：Discord、Telegram、Slack、WhatsApp 走 Team Gateway，另有终端、桌面端、Cloud。

模型：Nous 推理 API 全部模型加自有 key。

https://t.co/85hycjdcPH

[witcheer @witcheer]

here are the answers to the questions you asked most under yesterday's Hermes Business announcement:

(1) where your data lives

in an isolated tenant per team on Nous infrastructure. Enterprise deployments keep it on infrastructure you control.

(2) your own machines

Hermes Enterprise is the same stack on-prem or in your private cloud, with SSO, SLAs and white-glove onboarding. Business is the self-serve version on Nous infrastructure.

(3) the team skill library

a skill any member builds is published to the team library and becomes available to everyone's agent.

(4) one balance, one hosted agent per person

the team has its own balance, separate from your personal one. you invite colleagues by email from the Team page, each member gets their own hosted Hermes Agent and draws from the team balance under a spend cap you set.

(5) roles and channels

owners and admins manage roles, caps and API keys; members sign in and use their agent, on Discord, Telegram, Slack and WhatsApp through the Team Gateway, plus the terminal, the Desktop app and the Cloud.

(6) models

every model on the Nous inference API, plus your own provider keys.

[Hermes 中文社区 @hermescn_org]

Hermes 技巧 #73：gateway.message_timestamps

Agent 通过 Telegram 或 Discord 读你的消息，看不到发送时间。在 config.yaml 的 gateway 下加 message_timestamps 即可。

开启 enabled:true 后，你发的每条消息前会带上发送时间，只给模型看：聊天界面和磁盘记录都不变，agent 能分清今天早上和上周。

只作用于你发的消息，回复和系统提示不变。

https://t.co/yenPCIp1Wu

[witcheer @witcheer]

Hermes Wingtips #73: `gateway[.]message_timestamps`

your agent reads your messages through Telegram or Discord, but it does not see when you sent them.

`message_timestamps` under `gateway` in config.yaml changes that.

with `enabled:true`, every message you send carries its send time in front, for the model only. the chat you read does not change and the transcripts on disk stay clean. the agent can tell this morning from last week, notice a long gap, and answer with the right day in mind.

your messages only. the agent's replies and the system prompt are untouched.

[Hermes 中文社区 @hermescn_org]

MCP Command Center 在 Hermes Desktop 更新了：

每个 profile 独立 MCP 服务器与登录，Capabilities 有作用域选择。

远程 agent 的 OAuth 回调本机 Desktop 承接后转交 gateway。

token 登录后缓存复用，刷新失败即挂起该服务器，每天提醒，附 Sign in／Disable。

目录经 Nous Research 审核、以 PR 合并，安装时可勾选要暴露的工具。

https://t.co/WHzs6FDLhm

[witcheer @witcheer]

Tonbi recorded a tour of the MCP Command Center in Hermes Desktop!

a few things about MCP in Hermes Agent that come up a lot in the community, and how the current release handles them:

(1) per profile

each profile keeps its own MCP servers and its own logins, and the Capabilities tab has a scope selector so you see the set for the profile you are working in.

(2) signing in from your laptop to a remote Hermes

when your agent runs on another machine, the Desktop app hosts the OAuth callback on your machine and hands the authorisation to the gateway.

(3) tokens

once you sign in, the token is cached and reused until it can no longer be refreshed. when that happens the server is parked and the app tells you once a day, with a Sign in and a Disable button.

(4) the catalog

every entry was reviewed by Nous Research and merged by PR. at install you get a checklist of the server's tools and untick the ones you do not want your agent to see.

[Hermes 中文社区 @hermescn_org]

每周一早上由 Agent 自己写一份上周回顾：这段提示词粘进一个新会话即可。

它会检索最近七天的会话，写出十行以内的回顾——做完了什么、还开着什么、下一件值得做的事，用平实语言，不加小标题；只依据会话记录和记忆，记录单薄就直说，不凭猜测补空。

写完先把草稿给你看，再问哪天几点、投递到哪里，并列出要创建的定时任务：运行时间、提示词、投递目标。确认之后它才创建，并告诉你任务名和删除方法。

完整提示词在该帖首条回复里。

https://t.co/dz4HFFlCZh

[witcheer @witcheer]

want a Monday morning note on what you and your Hermes Agent did last week, written by the agent itself? paste this prompt into a fresh chat.

it searches your sessions from the last seven days, writes a ten-line review, and shows you the scheduled job: the day, the hour, the prompt it will run and where it lands.

copy / paste prompt in the first comment.

[Hermes 中文社区 @hermescn_org]

150 多条回复的反馈：最想要的是移动端 App（iOS/Android）；其次是 token——agent 读文件、调工具、维护记忆，每轮工作量比聊天窗口大，希望看清花在哪里；再就是更顺手的第一个小时：哪个界面做什么、怎么跑多个 agent。

已可用的还有：限流时的 fallback provider、hermes backup 与 hermes import，以及 /handoff 在 Telegram、Discord、CLI 之间移动会话。

https://t.co/GfQHoS4YKI

[witcheer @witcheer]

I read the 150+ replies under this question, here is what you said:

- a mobile app, iOS and Android, is the number one thing

- tokens came up a lot. an agent that reads your files, runs tools and keeps memory does more work per turn than a chat window, and you want to see where that work goes

- an easier first hour: which surface for what, how to run several agents

- a pairing guide for local models, which model for which card. it is now done on Hermes Desktop!

- Chrome profiles in browser use: pick which Chrome account the built-in browser opens with. it is also now integrated in Hermes Agent :)

- and a few asks are things that are already live: fallback providers for rate limits, hermes backup and hermes import to move an install, hermes sessions import to bring conversations in, /handoff to move a session between Telegram, Discord and CLI. I am sending every linked documentation in the first comment

your feedback is clear and the team has the full list. thank you @HermesWatcher for asking it.

Hermes Agent community for the win🪽

[Hermes 中文社区 @hermescn_org]

Hermes Agent 没有规定该用哪个界面，各入口背后是同一个 agent：

仓库里看工具调用用终端；其余时间用 Hermes Desktop，多个对话同开，改设置不用碰 yaml，另有内置浏览器和 Bot Mode。

手机上用 Discord 追加任务、读结果、确认某一步，终端里 /handoff discord 可把进行中的对话移过去。

底层用 Hermes Cloud，合上笔记本后 gateway 和定时任务继续跑。

https://t.co/542o1uo4b3

[witcheer @witcheer]

a few of you asked which Hermes Agent surface to use. there is no rule, it is one agent behind all of them.

here is what I do:

terminal when I am in a repo and want to watch the tool calls.

Hermes Desktop the rest of the day: several chats open, settings without touching yaml, the built-in browser, bots in Bot Mode.

Discord from the phone: nudge a task, read the result, say yes to a step. /handoff discord from the terminal moves a running conversation over.

Hermes Cloud underneath, so the gateway and the cron jobs keep running when the laptop is closed. Desktop talks to it as its backend.

and when I want a second agent that does not share memory with the first, hermes profile create.

https://t.co/5cIOWqT8tR

[Hermes 中文社区 @hermescn_org]

Hermes Business 已在 Nous Portal 上线：一个团队共用一个余额，从 Team 页面用邮箱邀请同事，为每位成员设置消费上限，每个人在常用渠道上都有自己的 agent。

成员做的技能会进入团队技能库，团队内其他 agent 也能直接用。

现在即可自助开通，跑在 Nous 的基础设施上；需要放在自己机器上的公司可以用 Hermes Enterprise，同一套栈部署在本地或自有云，支持 SSO 和 SLA。

https://t.co/GgHMZ3S5AY

[witcheer @witcheer]

Hermes Business is live on Nous Portal, and it answers the question I got most from teams this month:

"how do we run this for several people without giving everyone their own account?"

one team, one balance. you invite colleagues by email from the Team page, set a spend cap per member, and every one of them gets their agent on the channels they use. any skill one person builds lands in the team library and every other agent on the team can use it.

self-serve on Nous infrastructure today. and for the companies that need it on their own machines, Hermes Enterprise runs the same stack on-prem or in your cloud, with SSO and SLAs.

[Hermes 中文社区 @hermescn_org]

Microduck 开源到守护进程：模拟器跑实机同一套 robotd、tofd、robotctl，机体在 MuJoCo，实机到手前就能先训练。

全是模拟数据（RTX 5090）：控制环 50/50 Hz、302 tick 无遗漏；单条意图走 8 秒仍站立；测试 57 组 1381 全过；首训 1024 环境 0.65 秒/迭代 1.9 GB 显存。

下一步：做一个连通 robotd 的 Hermes Agent 工具，再训练行走策略。

https://t.co/m8dOHvjrLc

[witcheer @witcheer]

the Microduck is open source down to the daemons. the simulator runs the same robotd, tofd and robotctl binaries the real one boots, against a MuJoCo body. so I can start to train it before receiving it IRL :)

what ran on my RTX 5090 box today, all sim:

- standing: control loop 50.0 of 50.0 Hz, 302 ticks, 0 missed, trunk at 0.116 m
- walking: 8 s forward on a single intent, still standing after
- their whole test suite: 57 suites, 1381 passed, 0 failed
- first training probe on the GPU: 1024 environments, 0.65 s per iteration, 1.9 GB of VRAM

what I want to build next: a Hermes Agent tool that talks to robotd, then our own walking policy trained.