← Frameworks

Hermes AgentHermes Agent

Nous Research

PythonPython 开源Open source MITMIT 通用harnessGeneral-purpose harness 通用助手General-purpose assistant 常驻Always-on 跨平台消息Cross-platform messaging 定时任务Scheduled tasks MCPMCP 子代理Subagents RPCRPC 技能系统Skills system 记忆Memory 上下文压缩Context compression 权限审批Permission approval 自托管Self-hosted partial lifecycle: active 核验Checked 2026-10-23

Hermes Agent 是一个常驻的跨平台个人助手，会从经验里自己攒技能。

它不绑在你的笔记本上。

官方 README 说：

The self-improving AI agent built by Nous Research. It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. Run it on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It's not tied to your laptop — talk to it from Telegram while it works on a cloud VM.

⚠ 「the only agent」这句话，本站的态度是： 这是官方自我宣传语。本站未做跨产品对比测试来验证 "only" 是否成立。记录为官方声明，不是本站结论。

Hermes Agent is a resident cross-platform personal assistant that accumulates its own skills from experience. It is not tied to your laptop — the official README says:

The self-improving AI agent built by Nous Research. It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. Run it on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It's not tied to your laptop — talk to it from Telegram while it works on a cloud VM.

⚠ This site's position on "the only agent": that is the vendor's own marketing line. This site ran no cross-product comparison to test whether "only" holds. It is recorded as a vendor claim, not as this site's conclusion.

适合与不适合When to use it, when not to

适合这六类场景。

想要常驻的跨平台个人助手
放云 VM / serverless，不绑在自己的笔记本
需要定时任务无人值守
需要跨会话记忆与技能自积累
想用自建 endpoint 或多 provider
需要命令级审批 + DM 配对做安全边界

不适合这四类场景。

只做编程任务
需要当嵌入式 SDK
需要「状态 vs 上下文」分层保证
需要企业级多用户与审计

A fit for these six cases.

A resident cross-platform personal assistant (six platforms: Telegram / Discord / Slack / WhatsApp / Signal / Email)
Running it on a cloud VM or serverless, rather than tied to your laptop
Scheduled tasks that must run unattended
Cross-session memory and self-accumulating skills
Using your own endpoint or multiple providers such as OpenRouter
Command-level approval plus DM pairing as a security boundary

Not a fit for these four cases.

Coding tasks only — it is a general assistant; coding is one of many things
Using it as an embedded SDK — it is a complete runtime, not a library
A detailed "state vs context" layered guarantee — it has many mechanisms, but the vendor does not split them along those two layers

⚠ Unknown: enterprise multi-user and audit. The vendor provides no separate documentation.

Checked the security and messaging pages: both cover only single-owner allowlist / DM pairing authorization, not multi-tenant audit.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网直连实测 200。official site measured 200 direct.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 官方给了四种路径，并且强调没有锁定。

这是本站收录对象里，对模型自由度表述最直白的一个。

README 原文：

"Use any model you want — Nous Portal, OpenRouter, OpenAI, your own endpoint, and many others. Switch with hermes model — no code changes, no lock-in."

「your own endpoint」这一条对本站读者最有用。 自建网关、自托管推理都能直接接。

换模型不改代码，用一条运行时命令：/model [provider:model]。这条命令 CLI 与消息平台共用。（见站点上的 CLI vs Messaging 对照表）

完整 provider 清单已核验，来源是官方 docs/integrations/providers。官方一共列出 40+ 家：

Nous Portal —— 官方称 "300+ frontier agentic models"
OpenRouter、Anthropic、GitHub Copilot
xAI Grok、Google Gemini / Vertex AI
AWS Bedrock、Azure AI Foundry
DeepSeek、Hugging Face、NVIDIA Build
本地 / 自托管：Ollama、LM Studio，以及 Custom Endpoint

Custom Endpoint 的配置官方说：

"saved in config.yaml"

官方对那一页的定位原文：

"from cloud APIs ... to self-hosted endpoints like Ollama and vLLM, to advanced routing and fallback configurations."

⚠ 未知：换模型后的能力对齐度、本地量化模型在多后端下的行为差异，官方未给结论。

本站也没有实测。

核验：已查 providers 页与 configuration 页。
The official docs give four paths and stress there is no lock-in.

This is the most blunt statement of model freedom among everything we list.

README original:

"Use any model you want — Nous Portal, OpenRouter, OpenAI, your own endpoint, and many others. Switch with hermes model — no code changes, no lock-in."

The "your own endpoint" clause matters most to our readers. Self-built gateways and self-hosted inference connect directly.

Switching models needs no code change — one runtime command: /model [provider:model]. The same command works in the CLI and on messaging platforms. (See the CLI vs Messaging comparison table on the site.)

The full provider list is verified, from the official docs/integrations/providers. The official docs list 40+ in total:

Nous Portal — officially billed as "300+ frontier agentic models"
OpenRouter, Anthropic, GitHub Copilot
xAI Grok, Google Gemini / Vertex AI
AWS Bedrock, Azure AI Foundry
DeepSeek, Hugging Face, NVIDIA Build
Local / self-hosted: Ollama, LM Studio, and Custom Endpoint

For Custom Endpoint config the official docs say:

"saved in config.yaml"

The official positioning of that page, verbatim:

"from cloud APIs ... to self-hosted endpoints like Ollama and vLLM, to advanced routing and fallback configurations."

⚠ Unknown: capability alignment after switching models, and behavior differences of local quantized models across backends — the official docs give no conclusion.

We have not run measured tests either.

Verified: providers page and configuration page both checked.
运行位置Where it runs: 这是它证据最完整的一维：六种终端后端加插件后端，再加两个入口。

README 原文列举（2026-10-23 滚动核验版）：

"terminal backends — local, Docker, SSH, Modal, Daytona, and Vercel Sandbox — plus plugin backends such as Singularity/Apptainer."

⚠ 旧版 README（2026-10-01 核验时）写的是 "Seven terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox" —— Singularity 已被官方挪到 plugin backends 一类，本站按现口径记录。

官方还特别点出这两个后端：Daytona 与 Modal。它们带来 serverless persistence。

"your agent's environment hibernates when idle and wakes on demand, costing nearly nothing between sessions"

入口有两个。 README 有一张专门的 CLI vs Messaging 对照表：

终端 TUI：启动 hermes
消息平台：先跑 gateway，再从 Telegram / Discord / Slack / WhatsApp / Signal / Email，以及 Home Assistant 聊天

⚠ 这是本站最有价值的一格。 它是唯一明确说「agent 可以跑在云 VM 上、你从 Telegram 指挥它」的收录对象。

这和 OpenHands 的「关掉笔记本后 agent 仍在跑」是同一类诉求。区别是：OpenHands 把它做成可选部署项，Hermes 把它做成了默认形态。
It provides six terminal backends, plugin backends and two entry points.

README original list (2026-10-23 rolling-verification edition):

"terminal backends — local, Docker, SSH, Modal, Daytona, and Vercel Sandbox — plus plugin backends such as Singularity/Apptainer."

⚠ The older README (verified 2026-10-01) said "Seven terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox" — Singularity has since been moved by the official docs into plugin backends; we record the current wording.

The official docs also single out two backends: Daytona and Modal. They bring serverless persistence.

"your agent's environment hibernates when idle and wakes on demand, costing nearly nothing between sessions"

There are two entry points. The README has a dedicated CLI vs Messaging comparison table:

Terminal TUI: launch hermes
Messaging platforms: run the gateway first, then chat from Telegram / Discord / Slack / WhatsApp / Signal / Email, plus Home Assistant

⚠ This is the most valuable cell on our site. It is the only listed product that explicitly says "the agent can run on a cloud VM while you direct it from Telegram".

That is the same need as OpenHands' "the agent keeps running after you close the laptop". The difference: OpenHands makes it an optional deployment, Hermes makes it the default form.
本地文件Local files: 本地文件能力是间接获得的：它靠终端后端，而不是框架自己。

隔离机制见 permissions 维度。

工作区由所选后端决定。 官方 terminal.backend 默认是 local。

官方给的标注原文："Run on your machine (default)"

terminal.backend 决定工作区。工作目录由 cwd 决定。

这些后端各自提供一层隔离：

Docker / SSH
Modal / Daytona / Vercel
Singularity（README 现口径归入 plugin backends）

Context Files 是官方的一等概念，格式与优先级已核验。 （文档 features/context-files）支持的文件名如下：

.hermes.md / HERMES.md
AGENTS.md、AGENTS.override.md
CLAUDE.md
SOUL.md
.cursorrules、.cursor/rules/*.mdc

优先级是 first match wins。 顺序原文如下：

.hermes.md → AGENTS.override.md → AGENTS.md → CLAUDE.md → .cursorrules。 SOUL.md 是身份文件，始终单独加载。

在 git 仓库里，整条 AGENTS.md 链按 git root → 工作目录 合并。会话中还会渐进发现子目录 context。每个文件的字数上限是 floor 20,000 / ceiling 500,000。

文件访问有边界。 安全页第 8 层管的是 terminal 后端的工作目录参数：

官方原文："validated against an allowlist to prevent shell injection"

安全页第 3 层说，write_file 与 patch 有 denylist。另有可选 write sandbox。
Local-file capability comes indirectly: through terminal backends, not the framework itself.

Isolation mechanisms are covered under the permissions axis.

The workspace is determined by the chosen backend. The official terminal.backend defaults to local.

Official label, verbatim: "Run on your machine (default)"

terminal.backend decides the workspace. The working directory is decided by cwd.

Each backend provides one layer of isolation:

Docker / SSH
Modal / Daytona / Vercel
Singularity (current README wording files it under plugin backends)

Context Files are an official first-class concept; format and precedence verified. (docs features/context-files) Supported filenames:

.hermes.md / HERMES.md
AGENTS.md, AGENTS.override.md
CLAUDE.md
SOUL.md
.cursorrules, .cursor/rules/*.mdc

Precedence is first match wins. Order, verbatim:

.hermes.md → AGENTS.override.md → AGENTS.md → CLAUDE.md → .cursorrules. SOUL.md is the identity file and is always loaded separately.

In a git repo, the whole AGENTS.md chain merges git root → working directory. Subdirectory context is also discovered progressively during a session. The per-file word cap is floor 20,000 / ceiling 500,000.

File access has boundaries. Layer 8 of the security page governs the terminal backend's working-directory parameter:

Official wording: "validated against an allowlist to prevent shell injection"

Layer 3 of the security page says write_file and patch have a denylist. There is also an optional write sandbox.
关机后的任务Tasks after shutdown: 官方列出了三种后台任务机制。

① 内建 cron 调度器。 官方特性表原文：

"Built-in cron scheduler with delivery to any platform. Daily reports, nightly backups, weekly audits — all in natural language, running unattended."

（文档 features/cron）

② serverless 休眠后唤醒。 Daytona / Modal / Vercel 后端空闲时休眠、需唤醒，官方原话是这两个词： 官方原文："hibernates when idle and wakes on demand"

③ 常驻网关进程。 gateway 一旦运行，就能从多个消息平台随时触发。

补偿与补跑机制官方已写明（features/cron）：

有 misfire catch-up 扫描
resume 之后 "the next tick fires one catch-up run"
补跑可关：cron.catch_up_missed: false
全局 hermes pause 期间，托管 cron 的 fire webhook 返 503 Retry-After: 60，让调度器重投

官方称 "nothing is lost: due work catches up on the first tick"
运行中的任务 "never killed"

网关侧还有 event-loop liveness watchdog。它连续 3 次探针无响应就以 code 75 退出，由 supervisor 重启。另有 60s 心跳。

⚠ 未知：跨休眠 / 唤醒的状态一致性，官方未说明。

核验：已查 features/cron、features/tools、user-guide/messaging 三页。
The official docs describe three background-task mechanisms.

① Built-in cron scheduler. Official feature-table wording:

"Built-in cron scheduler with delivery to any platform. Daily reports, nightly backups, weekly audits — all in natural language, running unattended."

(docs features/cron)

② Serverless hibernate-and-wake. Daytona / Modal / Vercel backends hibernate when idle and wake on demand; the official phrase is exactly these words: Official wording: "hibernates when idle and wakes on demand"

③ A resident gateway process. Once the gateway runs, multiple messaging platforms can trigger it any time.

Catch-up and re-run mechanics are officially documented (features/cron):

There is a misfire catch-up scan
After resume, "the next tick fires one catch-up run"
Catch-up can be disabled: cron.catch_up_missed: false
While globally paused via hermes pause, the hosted cron's fire webhook returns 503 Retry-After: 60 so the scheduler re-delivers

Officially: "nothing is lost: due work catches up on the first tick"
Running tasks are "never killed"

The gateway side also has an event-loop liveness watchdog. After 3 consecutive unresponsive probes it exits with code 75 and the supervisor restarts it. There is also a 60s heartbeat.

⚠ Unknown: state consistency across hibernate/wake — the official docs do not state it.

Verified: features/cron, features/tools, user-guide/messaging — three pages checked.
工具与扩展Tools & extensions: 官方说它有 40+ 工具，并且明确支持 MCP。

文档目录里有独立的 "Tools & Toolsets" 页，副标题原文：

"40+ tools, toolset system, terminal backends"

（文档 features/tools）

MCP Integration 也是独立文档页，原文：

"Connect any MCP server for extended capabilities"

（文档 features/mcp）

还有一项很特别的能力：子代理 RPC。 官方原文：

"Spawn isolated subagents for parallel workstreams. Write Python scripts that call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns."

⚠ 「zero-context-cost」本站这样理解： 把多步管线移到上下文之外执行。也就是产物只回传结果。 ⚠ 官方没有给出具体的上下文节省量，本站标记为推断。

社区还维护了一个 MCP server，用途是控制 Linux 桌面。名字是 computer-use-linux。它用 AT-SPI 无障碍树 + Wayland/X11 输入 + 截图。

README 明确说它是给 "Hermes and other MCP hosts" 用的。
The official docs say it has 40+ tools and explicitly support MCP.

The docs directory has a dedicated "Tools & Toolsets" page, subtitle verbatim:

"40+ tools, toolset system, terminal backends"

(docs features/tools)

MCP Integration is also its own docs page, verbatim:

"Connect any MCP server for extended capabilities"

(docs features/mcp)

One capability is quite unusual: subagent RPC. Official wording:

"Spawn isolated subagents for parallel workstreams. Write Python scripts that call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns."

⚠ How we read "zero-context-cost": moving multi-step pipelines out of the context for execution — i.e. only the result is returned to the context. ⚠ The official docs give no concrete context-savings figure; we mark this as inference.

The community also maintains an MCP server for controlling the Linux desktop, named computer-use-linux. It uses the AT-SPI accessibility tree + Wayland/X11 input + screenshots.

The README explicitly says it is for "Hermes and other MCP hosts".
上下文与记忆Context & memory: ⚠ 这一维是本站最看重的，也恰好是本对象宣传力度最大的，必须分开核实。

先看官方声明。

README 特性表 "A closed learning loop" 一行原文：

"Agent-curated memory with periodic nudges. Autonomous skill creation after complex tasks. Skills self-improve during use. FTS5 session search with LLM summarization for cross-session recall. Honcho dialectic user modeling. Compatible with the agentskills.io open standard."

但把官方给的机制逐个归位，会发现它其实是三个维度。

官方说法本站归类说明
periodic nudges（自我提示）状态面机制触发器，不是压缩算法
FTS5 session search + LLM summarization 上下文面 ✅ 检索 + 摘要，这是真正的上下文管理
Skills（自主创建 + 使用中自我改进）独立一层程序性记忆，与对话上下文不同
Honcho dialectic user modeling 独立一层对用户的长期建模

关键区分点：压缩是用户显式触发的。 它有明确的 /compress 命令， CLI 与消息平台都有，对照表里与 /usage、/insights 并列。也就是说：不触发就不会压缩，它不是自动的。

还有一组能落到文件名的证据。 来自 "Migrating from OpenClaw" 章节，这是少见的官方一手清单：

SOUL.md —— persona
MEMORY.md 与 USER.md —— 记忆条目
AGENTS.md —— 工作区指令
~/.hermes/skills/openclaw-imports/ —— 技能目录

这说明记忆是明文的 markdown 文件，可读、可迁移。 对本站读者来说这是重要优点。
⚠ This is the dimension our site cares about most, and also the one this product markets hardest — they must be kept apart.

First, the official claims.

README feature table, the "A closed learning loop" row, verbatim:

"Agent-curated memory with periodic nudges. Autonomous skill creation after complex tasks. Skills self-improve during use. FTS5 session search with LLM summarization for cross-session recall. Honcho dialectic user modeling. Compatible with the agentskills.io open standard."

But placing each official mechanism into its proper bucket shows this is not one dimension but three.

Official claim Our categorization Note
periodic nudges (self-prompting) state plane a mechanism trigger, not a compression algorithm
FTS5 session search + LLM summarization context plane ✅ retrieval + summarization — this is genuine context management
Skills (autonomous creation + self-improvement during use) its own layer procedural memory, distinct from conversation context
Honcho dialectic user modeling its own layer long-term modeling of the user

Key distinction: compression is explicitly user-triggered. There is a dedicated /compress command, in both the CLI and messaging platforms, listed alongside /usage and /insights in the comparison table. That means: no trigger, no compression — it is not automatic.

There is also a set of evidence that comes down to filenames. From the "Migrating from OpenClaw" section, an unusually concrete first-party official list:

SOUL.md — persona
MEMORY.md and USER.md — memory entries
AGENTS.md — workspace instructions
~/.hermes/skills/openclaw-imports/ — skills directory

This shows memory is plaintext markdown files — readable and portable. For our readers this is an important advantage.
权限与限制Permissions & limits: 官方文档里有一节独立的 "Security"，把权限面拆成了八层。

八层原文列的是：

user authorization（allowlists、DM pairing）
dangerous command approval
file write safety
container isolation
MCP credential filtering
context file scanning
cross-session isolation
input sanitization

逐项核验（官方 user-guide/security）：

① Command approval（默认值已核验）。 approvals.mode 有三档，默认 smart。

smart —— 用辅助 LLM 评风险；低危命令自动批准，真正危险自动拒绝，不确定的升级为手动

manual —— 每次都问
off —— 等于 --yolo

超时默认 300 秒，未回复即 fail-closed。

另有始终生效的 hardline blocklist。它包含：

rm -rf /
fork bomb
dd 写块设备等

--yolo 或 approvals.mode: off 都不能越过它。

还有用户可编辑的 deny 规则：approvals.deny。它用 fnmatch glob、大小写不敏感。可以写成 "yolo 但除这些之外"。

② DM pairing（默认拒绝）。 授权检查按这个顺序走：

平台 allow-all
DM pairing 已批准名单
平台 allowlist（TELEGRAM_ALLOWED_USERS 等）
全局 allowlist
全局 allow-all
Default: deny

配对流程是 8 位码：

hermes pairing approve <platform> <code> unauthorized_dm_behavior 默认 pair。另可设 ignore / decline。配对码 1 小时过期、限速，5 次失败锁 1 小时。

③ Container isolation（默认关）。 terminal.backend 默认 local。容器隔离只在改用特定后端时才启用：

docker / ssh / singularity
modal / daytona / vercel

启用后，官方列出一整套 Docker 加固 flag：

--cap-drop ALL
--security-opt no-new-privileges
--pids-limit 256
noexec tmpfs 等

迁移脚本会碰密钥。 hermes claw migrate 会导入五家 API keys：

Telegram
OpenRouter / OpenAI
Anthropic
ElevenLabs

官方给了 --dry-run 预览。

⚠ 未知：gateway 端口安全，官方未单列一节。

dashboard 侧有 OAuth / 自托管 OIDC 鉴权。
The official docs have a standalone "Security" section that splits the permissions surface into eight layers.

The eight layers, verbatim:

user authorization (allowlists, DM pairing)
dangerous command approval
file write safety
container isolation
MCP credential filtering
context file scanning
cross-session isolation
input sanitization

Item-by-item verification (official user-guide/security):

① Command approval (defaults verified). approvals.mode has three settings, default smart.

smart — a helper LLM assesses risk; low-risk commands auto-approved, genuinely dangerous auto-denied, uncertain cases escalated to manual

manual — ask every time
off — equivalent to --yolo

Timeout defaults to 300 seconds; no reply means fail-closed.

There is also an always-active hardline blocklist. It includes:

rm -rf /
fork bombs
dd to block devices, etc.

Neither --yolo nor approvals.mode: off can bypass it.

There are also user-editable deny rules: approvals.deny. They use fnmatch globs, case-insensitive. You can write "yolo except these".

② DM pairing (deny by default). The authorization check runs in this order:

platform allow-all
DM pairing approved list
platform allowlist (TELEGRAM_ALLOWED_USERS etc.)
global allowlist
global allow-all
Default: deny

The pairing flow uses an 8-digit code:

hermes pairing approve <platform> <code> unauthorized_dm_behavior defaults to pair. It can also be ignore / decline. Pairing codes expire after 1 hour, are rate-limited, and 5 failures lock for 1 hour.

③ Container isolation (off by default). terminal.backend defaults to local. Container isolation activates only when switching to specific backends:

docker / ssh / singularity
modal / daytona / vercel

When enabled, the official docs list a full set of Docker hardening flags:

--cap-drop ALL
--security-opt no-new-privileges
--pids-limit 256
noexec tmpfs, etc.

The migration script touches secrets. hermes claw migrate imports API keys from five services:

Telegram
OpenRouter / OpenAI
Anthropic
ElevenLabs

The official docs provide --dry-run preview.

⚠ Unknown: gateway port security — the official docs have no dedicated section.

The dashboard side has OAuth / self-hosted OIDC authentication.
适合什么任务Best for: 适合这六类场景。

想要一个常驻的跨平台个人助手 （六个平台：Telegram / Discord / Slack / WhatsApp / Signal / Email）

想把它放在云 VM 或 serverless 上，不绑在自己的笔记本
需要定时任务无人值守地跑
需要跨会话记忆与技能自积累
想用自建 endpoint 或 OpenRouter 等多 provider
需要命令级审批与 DM 配对做安全边界

不适合这三类场景。

只做编程任务 —— 它是通用助手，编程只是其一
需要拿它当嵌入式 SDK —— 它是完整运行时，不是库
需要详细的「状态 vs 上下文」分层保证 —— 它机制多，但官方没按这两层分

⚠ 企业级多用户与审计，官方未单独提供文档。

已查 security 与 messaging 两页：二者只讲单 owner 的 allowlist / DM pairing 授权，未讲多租户审计。
Six scenarios it fits.

You want a resident cross-platform personal assistant (six platforms: Telegram / Discord / Slack / WhatsApp / Signal / Email)

You want it on a cloud VM or serverless, not tied to your laptop
You need scheduled tasks running unattended
You need cross-session memory and self-accumulating skills
You want a self-built endpoint or multi-provider setups like OpenRouter
You need command-level approval and DM pairing as the security boundary

Three scenarios it does not fit.

Coding-only work — it is a general assistant; coding is just one thing it does
Using it as an embedded SDK — it is a full runtime, not a library
You need detailed "state vs context" layering guarantees — it has many mechanisms, but the official docs do not organize them along those two layers

⚠ Enterprise multi-user and auditing: the official docs provide no dedicated documentation.

Security and messaging pages checked: they cover only single-owner allowlist / DM pairing authorization, not multi-tenant auditing.

头号误解Top misconception

当成编程框架用 —— 它是通用个人助手，编程只是它能做的事之一，形态与 LangGraph/CrewAI 完全不同Using it as a programming framework — it is a general personal assistant; coding is one of many things it does, and its form is nothing like LangGraph/CrewAI
相信「the only agent with a built-in learning loop」是本站结论—— 那是官方宣传语，本站未做跨产品对比验证Taking "the only agent with a built-in learning loop" as this site's conclusion — that is the vendor's marketing line; this site ran no cross-product comparison
把终端后端清单理解成部署方式清单 —— 官方说的是 terminal backends（工具执行的后端），且其中 Modal/Daytona 提供的是休眠不是常驻；清单本身也在变（Singularity 已从主列表挪到 plugin backends）Reading the terminal backend list as a deployment-mode list — the official text says terminal backends (the backend that executes tools), and Modal/Daytona there provide sleep, not always-on; the list itself changes (Singularity moved from the main list to plugin backends)
以为 /compress 是自动的 —— 它是显式命令，不触发就不会压缩Assuming /compress is automatic — it is an explicit command; nothing is compressed unless you trigger it
忽略 DM pairing 的重要性 —— 一个能接六个消息平台、还能在文件系统动手的 agent，谁能指挥它就是最关键的安全问题Overlooking DM pairing — for an agent that reaches six messaging platforms and touches your filesystem, who can command it is the central safety question
跑 hermes claw migrate 前不看 dry-run —— 该命令会接触 API keys（README 列出五家）Running hermes claw migrate without reading the dry-run first — the command touches API keys (the README lists five providers)
不确认就迁移 —— Hermes 与 OpenClaw 的关系本站未核验，只记录 README 的一手事实Migrating without confirming — this site's view of the Hermes / OpenClaw relationship was not verified; it records only the README's first-hand facts

未知项清单Known unknowns

serverless 后端的唤醒延迟与跨休眠状态一致性
「zero-context-cost」的实际上下文节省量
skills 自我改进的边界（会不会越学越歪）
Honcho 用户建模的数据去向与隐私
Hermes 与 OpenClaw 的确切关系（团队、是否改名、迁移完整性）
Nous Portal 的官方档位与额度数值（官网不公开，第三方口径未获确认）
Nous Portal 额度制的实际消耗速度（长 agent 循环能几天烧完一档，第三方说法未实证）
Nous Portal 与 OpenRouter 的差异（同样走 credit，但模型目录与工具网关不同）
本地 / 自托管模型在各后端下的可靠性（完整 provider 清单已核验）

Wake latency and cross-hibernation state consistency on serverless backends
The actual context savings of "zero-context-cost"
The boundary of skill self-improvement (can it learn itself crooked?)
Where Honcho user-modeling data goes, and privacy
The exact relationship between Hermes and OpenClaw (team, rename, migration completeness)
Nous Portal's official tiers and quota numbers (not published on the official site; third-party figures unconfirmed)
The actual burn rate of the Nous Portal credit system (how many days of long agent loops drain one tier — third-party claims unproven)
The difference between Nous Portal and OpenRouter (both credit-based, but different model catalogs and tool gateways)
Reliability of local / self-hosted models across backends (the full provider list is verified)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-23。

类型Type	名称Name	链接Link
repo	NousResearch/hermes-agent · 仓库（256,242★，MIT，滚动核验 2026-10-23）NousResearch/hermes-agent · repository (256,242★, MIT, rolling verification 2026-10-23)	https://github.com/NousResearch/hermes-agent
docs	主 README（定位、六项特性表含终端后端清单与学习闭环、CLI vs Messaging 对照表、文档目录、安全与 MCP 页、OpenClaw 迁移清单）Main README (positioning, six-item feature table with terminal-backend list and learning loop, CLI vs Messaging comparison table, docs directory, security and MCP pages, OpenClaw migration list)	https://github.com/NousResearch/hermes-agent/blob/main/README.md
docs	官方文档首页Official docs homepage	https://hermes-agent.nousresearch.com/docs/
docs	LLM and Model Providers（官方 provider 清单 40+ 家、Nous Portal 300+ models、Ollama / LM Studio / vLLM / Custom Endpoint）LLM and Model Providers (official provider list of 40+, Nous Portal 300+ models, Ollama / LM Studio / vLLM / Custom Endpoint)	https://hermes-agent.nousresearch.com/docs/integrations/providers
docs	Tools & Toolsets（40+ tools、terminal backends 七种、container security、background 进程管理）Tools & Toolsets (40+ tools, seven terminal backends, container security, background process management)	https://hermes-agent.nousresearch.com/docs/user-guide/features/tools
docs	Security（Command approval / DM pairing / container isolation）Security (Command approval / DM pairing / container isolation)	https://hermes-agent.nousresearch.com/docs/user-guide/security
docs	Memory（Persistent memory、user profiles、best practices）Memory (Persistent memory, user profiles, best practices)	https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
docs	MCP Integration（Connect any MCP server）MCP Integration (Connect any MCP server)	https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp
docs	Tools & Toolsets（40+ tools、toolset system、terminal backends）Tools & Toolsets (40+ tools, toolset system, terminal backends)	https://hermes-agent.nousresearch.com/docs/user-guide/features/tools
docs	Cron Scheduling（Scheduled tasks with platform delivery）Cron Scheduling (Scheduled tasks with platform delivery)	https://hermes-agent.nousresearch.com/docs/user-guide/features/cron
docs	Context Files（Project context that shapes every conversation）Context Files (Project context that shapes every conversation)	https://hermes-agent.nousresearch.com/docs/user-guide/features/context-files
docs	Skills System（Procedural memory、Skills Hub）Skills System (Procedural memory, Skills Hub)	https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
docs	Nous Portal（可跳过 API key 收集）Nous Portal (can skip API-key collection)	https://portal.nousresearch.com
pricing	Nous Portal 首页（核验 2026-10-01：只有定性描述、无任何价格数字；原文含 "with free options and Portal-only discounts"；`/pricing` 路径返 404，故本站判定官方不公开定价）Nous Portal homepage (verified 2026-10-01: qualitative wording only, zero price numbers; original text includes "with free options and Portal-only discounts"; the `/pricing` path returns 404, so we conclude the official docs do not publish pricing)	https://portal.nousresearch.com/
docs	agentskills.io开放标准（官方称兼容）agentskills.io open standard (officially claimed compatible)	https://agentskills.io
changelog	Releases（v2026.9.24 @ 2026-09-24）Releases (v2026.9.24 @ 2026-09-24)	https://github.com/NousResearch/hermes-agent/releases

实测记录Measured runs

本档未跑集成与性能实测；结论只覆盖 API 与官方文档披露的能力面。

实测建议（按通用助手 + 常驻形态定制）：

#	测什么	为什么值得测
1	DM pairing 与 command approval 的默认状态	部署安全的前提，也是本站最大的缺口
2	`/compress` 压缩后信息保留的实际效果	它是本站最看重的维度，但压缩是手动触发的
3	Modal/Daytona 休眠后唤醒的延迟与状态一致性	官方强调省成本，但唤醒体验决定能否日用
4	用自建 endpoint 接本地量化模型跑一轮	README 明确支持，本站读者关心的组合
5	子代理 RPC 的「zero context cost」实际省了多少	官方说法未量化
6	cron 任务在网关重启后的行为	无人值守场景的核心可靠性问题
7	只用它做编程任务 vs 专用编程框架的体验差	它定位是通用助手，本站要给出诚实判断

This site has not completed measured runs. The test protocol is at protocols/framework/_protocol.md.

Suggested runs (tailored to the general-assistant + resident form):

#	What to test	Why it is worth testing
1	Default state of DM pairing and command approval	the precondition of deployment safety, and our biggest gap
2	Actual information retention after `/compress`	our most-watched dimension, but compression is manually triggered
3	Wake latency and state consistency after Modal/Daytona hibernation	the official docs stress cost savings, but wake experience decides daily usability
4	One round with a local quantized model via a self-built endpoint	the README explicitly supports it; a combination our readers care about
5	How much subagent RPC's "zero context cost" actually saves	the official claim is unquantified
6	Cron task behavior after a gateway restart	the core reliability question for unattended scenarios
7	Using it for coding-only vs a dedicated coding framework	it positions itself as a general assistant; we owe an honest judgment

选型第一提醒：它不是编程框架First selection reminder: it is not a coding framework

这是本页最需要先说清的一件事。 Hermes 与本站收录的其它 harness 不在同一层：

对象	定位	入口
LangGraph / Google ADK / CrewAI	编程任务的编排框架	Python 代码
Deep Agents	编程用 batteries-included harness	Python 代码
OpenHands	自托管 agent 控制中心	Web UI
Hermes Agent	跨平台常驻的个人助手	TUI + 六个消息平台

编程只是 Hermes 能做的事之一。官方对它的描述是：

"runs anywhere, not just your laptop"

消息入口有六个：Telegram、Discord、Slack、 WhatsApp、Signal、Email。它也能接 Home Assistant。

⚠ 另有一条本站必须记的前身信息。 README 有独立的 "Migrating from OpenClaw" 章节。 hermes setup 会自动检测 ~/.openclaw 并提供迁移。

⚠ 未知：Hermes 与 OpenClaw 的关系（是否同一团队、是否改名）本站未核验。 本站只记录 README 的一手事实。

This is the single thing this page most needs to say up front. Hermes is not on the same layer as the other harnesses we list:

Object	Positioning	Entry
LangGraph / Google ADK / CrewAI	orchestration frameworks for coding tasks	Python code
Deep Agents	batteries-included harness for coding	Python code
OpenHands	self-hosted agent control center	Web UI
Hermes Agent	a resident cross-platform personal assistant	TUI + six messaging platforms

Coding is just one thing Hermes can do. The official description of it:

"runs anywhere, not just your laptop"

The messaging entry points are six: Telegram, Discord, Slack, WhatsApp, Signal, Email. It can also connect to Home Assistant.

⚠ One more piece of predecessor information we must record. The README has a standalone "Migrating from OpenClaw" section. hermes setup auto-detects ~/.openclaw and offers migration.

⚠ Unknown: the relationship between Hermes and OpenClaw (same team? a rename?) is not verified by us. We record only the README's first-party facts.

两种形态都是一等公民Both forms are first-class citizens

README 有一张完整的 CLI vs Messaging 对照表。两个入口是：

操作	CLI	消息平台
开始	`hermes`	`hermes gateway setup` + `gateway start`，然后给 bot 发消息
新会话	`/new`、`/reset`	同
换模型	`/model [provider:model]`	同
压缩上下文	`/compress`	`/compress`
用量与洞察	`/usage`、`/insights [--days N]`	同
技能	`/skills` 或 `/<skill-name>`	`/<skill-name>`
打断当前工作	`Ctrl+C` 或发新消息	`/stop` 或发新消息

/compress 是两边都有的显式命令 —— 详见下方 context 一节。

The README has a full CLI vs Messaging comparison table. The two entry points:

Operation	CLI	Messaging platforms
Start	`hermes`	`hermes gateway setup` + `gateway start`, then message the bot
New session	`/new`, `/reset`	same
Switch model	`/model [provider:model]`	same
Compress context	`/compress`	`/compress`
Usage & insights	`/usage`, `/insights [--days N]`	same
Skills	`/skills` or `/<skill-name>`	`/<skill-name>`
Interrupt current work	`Ctrl+C` or send a new message	`/stop` or send a new message

/compress is an explicit command on both sides — see the context section below.

终端后端清单：这是它最独特的一格The terminal-backend list: its most unique cell

README 原文（2026-10-23 滚动核验版）：

terminal backends — local, Docker, SSH, Modal, Daytona, and Vercel Sandbox — plus plugin backends such as Singularity/Apptainer. Daytona and Modal offer serverless persistence — your agent's environment hibernates when idle and wakes on demand, costing nearly nothing between sessions.

⚠ 三点要说清。

① 这是 terminal backend，指的是「工具在哪执行」，不是部署方式清单。 local 就是本机执行，Docker 是容器，SSH 是远端机， Modal/Daytona/Vercel 是 serverless。

② 清单本身在变。 2026-10-01 核验时 README 写「Seven terminal backends」并把 Singularity 列在主列表；2026-10-23 滚动核验时 Singularity 已挪到 plugin backends （such as Singularity/Apptainer），主列表剩六种。引用这类清单要带核验日。

③ 「空闲时休眠」不等于「常驻」。 官方用词是 hibernate，唤醒有延迟。休眠期间的状态一致性本站未核验。

对比 OpenHands：它也主打「关掉笔记本后 agent 仍在跑」。区别是那是可选的部署形态，而 Hermes 把它做成了默认形态。

README original (2026-10-23 rolling-verification edition):

terminal backends — local, Docker, SSH, Modal, Daytona, and Vercel Sandbox — plus plugin backends such as Singularity/Apptainer. Daytona and Modal offer serverless persistence — your agent's environment hibernates when idle and wakes on demand, costing nearly nothing between sessions.

⚠ Three things to make clear.

① These are terminal backends — meaning "where tools execute", not a deployment-method list. local runs on your machine, Docker is a container, SSH is a remote machine, Modal/Daytona/Vercel are serverless.

② The list itself changes. When verified on 2026-10-01 the README said "Seven terminal backends" and listed Singularity in the main list; on the 2026-10-23 rolling verification Singularity had moved to plugin backends (such as Singularity/Apptainer), leaving six in the main list. Cite such lists with the verification date.

③ "Hibernates when idle" is not the same as "always resident". The official word is hibernate; waking has latency. State consistency during hibernation is not verified by us.

Compare OpenHands: it also leads with "the agent keeps running after you close the laptop". The difference is that there it is an optional deployment form, while Hermes makes it the default form.

记忆与上下文：机制最多，但官方没按本站关心的两层分Memory and context: the most mechanisms, but the official docs don't split the two layers we care about

这是本对象宣传力度最大的维度，也是本站必须仔细拆的地方。

官方声明（特性表 "A closed learning loop" 一行原文）：

Agent-curated memory with periodic nudges. Autonomous skill creation after complex tasks. Skills self-improve during use. FTS5 session search with LLM summarization for cross-session recall. Honcho dialectic user modeling. Compatible with the agentskills.io open standard.

把官方给的机制逐个归位，会发现这是三个不同维度混在一句话里：

官方说法	本站归类	说明
periodic nudges（自我提示）	状态面	机制触发器，不是压缩算法
FTS5 session search + LLM summarization	上下文面 ✅	检索 + 摘要，这是真正的上下文管理
Skills（自主创建 + 使用中自我改进）	独立一层	程序性记忆，与对话上下文不同
Honcho dialectic user modeling	独立一层	对用户的长期建模

关键区分点：压缩是显式触发的。 /compress 在 CLI 与消息平台都有。它与 /usage、/insights 并列。 不触发就不会压缩。

记忆是明文的 markdown 文件（这是重要优点）

来自 "Migrating from OpenClaw" 章节的迁移清单 —— 这是少见的官方一手文件级证据：

文件 / 目录	内容
`SOUL.md`	persona 文件
`MEMORY.md`、`USER.md`	记忆条目
`AGENTS.md`	工作区指令
`~/.hermes/skills/openclaw-imports/`	用户创建的技能

记忆是可读、可 diff、可手改的 markdown，不像数据库里的黑盒。

另有一个独立的 Context Files 概念。（文档 features/context-files）官方定义是：

"Project context that shapes every conversation"

即项目上下文文件会影响每一次对话。

它的格式与优先级已核验：first match wins。 .hermes.md → AGENTS.override.md → AGENTS.md → CLAUDE.md → .cursorrules。 SOUL.md 作为身份始终单独加载。（见 axes.local_files）

This is the dimension this product markets hardest, and the one we must carefully take apart.

Official claim (feature table, the "A closed learning loop" row, verbatim):

Agent-curated memory with periodic nudges. Autonomous skill creation after complex tasks. Skills self-improve during use. FTS5 session search with LLM summarization for cross-session recall. Honcho dialectic user modeling. Compatible with the agentskills.io open standard.

Placing each official mechanism into its proper bucket shows three different dimensions mixed into one sentence:

Official claim	Our categorization	Note
periodic nudges (self-prompting)	state plane	a mechanism trigger, not a compression algorithm
FTS5 session search + LLM summarization	context plane ✅	retrieval + summarization — this is genuine context management
Skills (autonomous creation + self-improvement during use)	its own layer	procedural memory, distinct from conversation context
Honcho dialectic user modeling	its own layer	long-term modeling of the user

Key distinction: compression is explicitly triggered. /compress exists in both the CLI and messaging platforms. It sits alongside /usage and /insights. No trigger, no compression.

Memory is plaintext markdown files (an important advantage)

From the migration list in the "Migrating from OpenClaw" section — unusually concrete first-party file-level evidence:

File / directory	Content
`SOUL.md`	persona file
`MEMORY.md`, `USER.md`	memory entries
`AGENTS.md`	workspace instructions
`~/.hermes/skills/openclaw-imports/`	user-created skills

Memory is readable, diffable, hand-editable markdown — not a black box inside a database.

There is also a separate Context Files concept. (docs features/context-files) The official definition:

"Project context that shapes every conversation"

i.e. project context files affect every conversation.

Its format and precedence are verified: first match wins. .hermes.md → AGENTS.override.md → AGENTS.md → CLAUDE.md → .cursorrules. SOUL.md is always loaded separately as the identity file. (see axes.local_files)

权限：六个平台入口让「谁能指挥它」成为核心问题Permissions: six platform entry points make "who can command it" the core question

README 的文档目录里有一节独立的 Security，副标题原文：

Command approval, DM pairing, container isolation

逐项解读。

① Command approval —— 命令级审批。 默认 smart：辅助 LLM 评风险，低危自动批准、危险自动拒绝、不确定转手动。超时 300 秒 fail-closed。另有始终生效的 hardline blocklist，它不受这两个开关影响。

用户还能自己写 deny 规则：approvals.deny。它用 fnmatch glob、大小写不敏感。

这与别家的做法不在同一层。

OpenHands 用 API key 机制
Codex SDK 有四种 ApprovalMode

② DM pairing —— 私聊配对，即只有配对过的用户能指挥它。 默认拒绝：授权检查顺序里，pairing 已批准名单排在平台 allowlist 之前。全未命中即 deny。配对流程是 8 位码：

hermes pairing approve <platform> <code>

这一项在六个消息平台入口的前提下极其重要。 一个能通过 Telegram / WhatsApp 接指令、还会在文件系统上动手的 agent， 「谁能给它下指令」就是最关键的安全问题 —— 这也是本站把它单独列出来的理由。

③ Container isolation —— 容器隔离。默认关。 terminal.backend 默认是 local。改用下面这组后端时才会启用隔离：

docker / ssh
singularity
modal / daytona / vercel

The README's docs directory has a standalone Security section, subtitle verbatim:

Command approval, DM pairing, container isolation

Item by item.

① Command approval — command-level approval. Default smart: a helper LLM assesses risk; low-risk auto-approved, dangerous auto-denied, uncertain escalated to manual. Timeout 300 seconds, fail-closed. There is also an always-active hardline blocklist, unaffected by those two switches.

Users can also write their own deny rules: approvals.deny. They use fnmatch globs, case-insensitive.

This is not on the same layer as other products' approaches.

OpenHands uses an API-key mechanism
Codex SDK has four ApprovalMode values

② DM pairing — private-chat pairing: only paired users can command it. Deny by default: in the authorization-check order, the pairing approved list sits before the platform allowlist. If nothing matches, deny. The pairing flow uses an 8-digit code:

hermes pairing approve <platform> <code>

With six messaging entry points, this item is extremely important. For an agent that takes instructions over Telegram / WhatsApp and acts on the file system, "who can give it orders" is the most critical security question — which is why we list it separately.

③ Container isolation — off by default. terminal.backend defaults to local. Isolation activates when switching to this backend group:

docker / ssh
singularity
modal / daytona / vercel

工具：40+ 且支持 MCP，还有一项很特别的能力Tools: 40+ with MCP support, plus one very unusual capability

README 文档目录（原文副标题）：

文档节	官方副标题
Tools & Toolsets	"40+ tools, toolset system, terminal backends"
MCP Integration	"Connect any MCP server for extended capabilities"
Cron Scheduling	"Scheduled tasks with platform delivery"

特别能力：子代理 RPC。

官方原文：

Spawn isolated subagents for parallel workstreams. Write Python scripts that call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns.

⚠ 「zero-context-cost」本站读作「把多步管线移到上下文之外执行，产物只回传结果」 （这是对它的合理解释）。 但官方未给出具体的上下文节省量，标记为推断而非官方表述。

社区还维护了一个 Linux 桌面控制 MCP server：computer-use-linux。它的能力清单是：

AT-SPI 无障碍树
Wayland/X11 输入
截图
窗口定向

README 明确说它是给 "Hermes and other MCP hosts" 用的 —— 这是 MCP 生态的一个真实案例。

README docs directory (subtitle verbatim as verified on 2026-10-08):

Docs section	Official subtitle
Tools & Toolsets	"40+ tools, toolset system, terminal backends"
MCP Integration	"Connect any MCP server for extended capabilities"
Cron Scheduling	"Scheduled tasks with platform delivery"

The unusual capability: subagent RPC.

Official wording:

Spawn isolated subagents for parallel workstreams. Write Python scripts that call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns.

⚠ We read "zero-context-cost" as "multi-step pipelines execute outside the context; only the result is returned" (a reasonable reading of it). But the official docs give no concrete context-savings figure — marked as inference, not an official statement.

The community also maintains a Linux-desktop-control MCP server: computer-use-linux. Its capability list:

AT-SPI accessibility tree
Wayland/X11 input
screenshots
window targeting

The README explicitly says it is for "Hermes and other MCP hosts" — a real-world case of the MCP ecosystem.

定时任务：常驻形态的第三重保障Scheduled tasks: the third guarantee of the resident form

README 特性表原文：

Built-in cron scheduler with delivery to any platform. Daily reports, nightly backups, weekly audits — all in natural language, running unattended.

常驻这件事官方给了三重机制：

① 内建 cron
② serverless 休眠后唤醒
③ 常驻网关进程（多平台随时触发）

cron 的补偿 / 补跑已核验。

有 misfire catch-up 扫描
补跑可关：cron.catch_up_missed
全局 hermes pause 期间，托管 cron 的 fire webhook 重投 503 Retry-After: 60
运行中的任务 "never killed"

网关侧有 event-loop liveness watchdog 与 60s 心跳。

⚠ 未知：跨休眠 / 唤醒的状态一致性，官方未说明。

README feature table, verbatim:

Built-in cron scheduler with delivery to any platform. Daily reports, nightly backups, weekly audits — all in natural language, running unattended.

The official docs give three layers for residency:

① built-in cron
② serverless hibernate-and-wake
③ a resident gateway process (multi-platform triggering any time)

Cron catch-up / re-run is verified.

There is a misfire catch-up scan
Catch-up can be disabled: cron.catch_up_missed
While globally paused via hermes pause, the hosted cron's fire webhook re-delivers with 503 Retry-After: 60
Running tasks are "never killed"

The gateway side has an event-loop liveness watchdog and a 60s heartbeat.

⚠ Unknown: state consistency across hibernate/wake — the official docs do not state it.

迁移脚本会碰密钥：用 dry-runThe migration script touches secrets: use dry-run

hermes claw migrate 涉及的迁移清单里有：

API keys — allowlisted secrets (Telegram, OpenRouter, OpenAI, Anthropic, ElevenLabs)

也就是说，迁移脚本本身会接触五家服务的密钥。 官方给的手段是：

hermes claw migrate --dry-run              # 预览会迁移什么
hermes claw migrate --preset user-data     # 不导入密钥

跑之前先 dry-run —— 这是官方自己设计的用法。

The migration list covered by hermes claw migrate includes:

API keys — allowlisted secrets (Telegram, OpenRouter, OpenAI, Anthropic, ElevenLabs)

In other words, the migration script itself touches keys from five services. The official remedy:

hermes claw migrate --dry-run              # preview what would migrate
hermes claw migrate --preset user-data     # do not import keys

Dry-run before running — that is the usage the official docs themselves design.

核验说明Verification notes

confidence: partial 的依据：

为什么不是 verified。

model_access 的换模型后能力对齐度未核验
background 的休眠 / 唤醒延迟与跨休眠状态一致性官方未说明
「zero-context-cost」与 Honcho / skills 的实际效果未核验

2026-10-08 已补：permissions 三机制默认值、local_files 的 context files 格式与优先级、 cron 补偿语义均已核验，见上。

已核验：

仓库存在与星数（256,242 · 滚动核验 2026-10-23，本站收录对象里最高）
许可（MIT，经 license API）、最近推送（2026-10-22，仍活跃）
最新版 v0.22.4（releases @ 2026-10-21）——⚠ 版本命名已从日期式（v2026.9.24）换成 semver（v0.22.x），换轨发生在 2026-10-08 至 2026-10-12 之间（v2026.9.24 之后直接是 v0.21.6）
README 全文 243 行。它含定位段、六项特性表与终端后端原文（现口径：六种主列表 + plugin backends）
它还有完整的 CLI vs Messaging 对照表
十四节文档目录、OpenClaw 迁移清单、社区与许可段
仓库有中文与西语等 README 分支、Nous Portal 入口

未核验：

换模型后的能力对齐度与本地模型可靠性
休眠唤醒延迟与跨休眠状态一致性
「zero-context-cost」的量化值
Honcho 用户建模与 skills 自改进的实际效果
Hermes 与 OpenClaw 的确切关系

2026-10-01 补充：Nous Portal 计费的官网状态已核验 —— 官网不公开定价（首页只有定性描述、无一个数字；/pricing 返 404）。第三方口径流传但未获官方确认——按本站的报价记录约定本站不记录这些数字。

这条的状态： 从「完全未核验」变成「已核验到官网不公开」。

Basis for confidence: partial:

Why not verified.

model_access: capability alignment after switching models not verified
background: hibernate/wake latency and cross-hibernation state consistency not stated by the official docs
"zero-context-cost" and the real-world effect of Honcho / skills not verified

2026-10-08 additions: the three permission mechanisms' defaults, the context-files format and precedence under local_files, and the cron catch-up semantics are all verified, see above.

Verified:

Repository existence and star count (256,242 · rolling verification 2026-10-23, the highest among everything we list)
License (MIT, via the license API), last push (2026-10-22, still active)
Latest release v0.22.4 (releases @ 2026-10-21) — ⚠ version naming has switched from date-style (v2026.9.24) to semver (v0.22.x); the switch happened between 2026-10-08 and 2026-10-12 (v2026.9.24 is followed directly by v0.21.6)
README full text 243 lines. It contains the positioning paragraph, the six-item feature table with terminal-backend wording (current wording: six in the main list + plugin backends)
It also has a full CLI vs Messaging comparison table
A fourteen-section docs directory, the OpenClaw migration list, community and license sections
The repo has Chinese and Spanish README variants, and a Nous Portal entry

Not verified:

Capability alignment after switching models, and local-model reliability
Hibernate/wake latency and cross-hibernation state consistency
The quantified value of "zero-context-cost"
The real-world effect of Honcho user modeling and skill self-improvement
The exact relationship between Hermes and OpenClaw

Added the 2026-10-08 check (2026-10-01): the official-site status of Nous Portal billing is verified — the official site does not publish pricing (the homepage has only qualitative wording, not a single number; /pricing returns 404). Third-party figures circulate but are not confirmed by the official docs — per our price-recording policy we do not record those numbers.

Status of this item: moved from "entirely unverified" to "verified that the official site does not publish pricing".