常見問題
關於 Hermes Agent 的常見問題,全部來自官方 README、發布說明和文件。
Hermes Agent 在 README 裡的自我介紹是「Nous Research 打造的自進化 AI 智能體」——內建一套學習閉環,會從實際使用中沉澱出 skill、使用過程中繼續打磨,還會檢索自己過去的對話,並且跨 session 逐步建立對你的使用者模型。專案採用 MIT 授權,原始碼公開在 github.com/NousResearch/hermes-agent。
Hermes Agent 由 Nous Research 開發。NousResearch/hermes-agent 倉庫列出 300+ 位貢獻者,但絕大多數 commit 來自 Nous Research 的核心維護者——光是頭號貢獻者就有 2,500+ commit。專案以 MIT 授權開源。
官方推薦的安裝方式是 README 那條 curl 指令:`curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash`。Script 會自動幫你裝好 uv、Python 3.11、Node.js、ripgrep 和 ffmpeg。完整文件見 hermes-agent.nousresearch.com/docs/getting-started/installation。
不行,官方不支援原生 Windows。README 明確列出的支援平台是 Linux、macOS、WSL2,以及透過 Termux 的 Android。Windows 使用者請走 WSL2——在 WSL2 發行版裡跑同一條 curl 指令即可。
Python 3.11,README 和倉庫的 pyproject.toml 都是這麼寫的。安裝 script 會透過 uv 自動把 Python 3.11 裝好,你不需要自己設環境——git 是唯一的前置條件。
README 裡列的內建 provider 包括 Nous Portal、OpenRouter(200+ 模型)、z.ai/GLM、Kimi/Moonshot、MiniMax、OpenAI 以及自訂 endpoint。v0.8.0 還新增了原生 Google AI Studio 支援,Anthropic 和 Mistral 也直接當依賴打包進來。裝完後跑 `hermes model` 就能挑一個用。
README 內建支援的部署對象包括 Telegram、Discord、Slack、WhatsApp、Signal、Email 和 CLI。v0.6.0 又新增了飛書/Lark 和企業微信。所有平台都透過同一個 `hermes setup` 精靈來設定。
Hermes Agent 本體完全免費,採用 MIT 授權。要付的錢是模型 provider 那邊的費用(Nous Portal、OpenRouter、OpenAI 等)。按 v0.8.0 的發布說明,Nous Portal 提供免費的 MiMo v2 Pro 額度,你也可以把 Hermes 指向本地的模型 endpoint。
安裝 script 預設會把倉庫 clone 到 `~/.hermes/hermes-agent/`,並在 `~/.local/bin/hermes` 建立 `hermes` 執行檔的軟連結。想換位置的話,跑 script 前設一下 `HERMES_INSTALL_DIR` 就行。
直接跑 `hermes update`——CLI 自帶升級器,官方安裝指南有說明。想自己控制的話,`cd` 進安裝目錄、`git pull`,然後在 venv 裡重新執行 `uv pip install -e ".[all]"`。
可以。README 明確支援透過 Termux 在 Android 上執行,安裝 script 會自動偵測 Termux 並裝 `.[termux]` extra。同一條指令,不需要 root。
README 把它描述成一個閉環:Hermes 會「從使用經驗中創建 skill、在使用過程中持續改進、主動把知識沉澱下來、檢索自己過去的對話、並在跨 session 間持續加深對你的使用者模型」。README 正是拿這一點來區分 Hermes 和一般的無狀態聊天助手。
skill 是可重用、可組合的能力單元。Hermes 原生支援 README 裡提到的開放目錄 agentskills.io,而且能在使用過程中自己產生並儲存新的 skill——這套會自我書寫的 skill 系統,正是學習閉環的一部分。
支援。v0.6.0(2026-03-30)加入了 MCP server 模式,讓 Hermes 能把自己的工具暴露給其他 MCP client。v0.8.0(2026-04-08)又補上了 MCP OAuth 2.1,用於帶鑑權的 MCP 連線——兩項都在官方發布說明裡列為重點特性。
v0.6.0(2026-03-30)新增了官方 Docker 容器,這是那次發布在 GitHub notes 裡標注的一項重點。在 v0.6.0 之前,唯一支援的安裝方式就是 curl script 加本地 Python 環境。
README 列了六種工具執行的 terminal 後端:local、Docker、SSH、Daytona、Singularity、Modal。從普通 shell 到短生命期的 serverless 容器都涵蓋到,可以依需要把 Hermes 的指令跑在不同隔離強度的環境裡。
v0.7.0(2026-04-03)把 memory 變成了一套可插拔的 provider 介面——第三方後端只要實作一個 provider ABC 並透過 plugin 系統註冊就行。Honcho 是參考實作,帶有 profile 維度的 host/peer 解析,也正是 README 裡描述的那套「辯證式使用者模型」後端。
README 的快速開始流程是:先 `source ~/.bashrc` 重載 shell,然後 `hermes` 開始聊、`hermes model` 挑 provider、`hermes tools` 設定工具、`hermes setup` 走完整的設定精靈。這些都能在 README 的安裝後段找到。
官方倉庫是 github.com/NousResearch/hermes-agent——MIT 授權,目前約 5 萬 star、約 6500 fork,main 分支上有 3,700+ commit。首頁是 hermes-agent.nousresearch.com。
README 給出三個社群管道:Nous Research Discord(discord.gg/NousResearch),bug 回報走 github.com/NousResearch/hermes-agent/issues,更長的問題與 skill 分享則在同一個 repo 的 GitHub Discussions。
把 CLI 的軟連結和安裝目錄刪掉就行:`rm ~/.local/bin/hermes`,接著 `rm -rf ~/.hermes/hermes-agent`。如果安裝時設過 `HERMES_INSTALL_DIR`,就刪對應那個路徑。Script 不會動任何系統層級的套件。
不是。這是一個由粉絲架設的社群站點,不是由 Nous Research 營運、背書或與之存在任何關聯。站點內容全部來自公開資料——GitHub 倉庫、發布說明以及 hermes-agent.nousresearch.com 的官方文件。Hermes Agent 和 Nous Research 均為各自所有者的商標。
沒有。README 列出的內建介面是 CLI 和聊天閘道(Telegram、Discord、Slack、WhatsApp、Signal、Email),v0.8.0 不附帶官方瀏覽器介面。如果你想在瀏覽器裡用 Hermes,社群常見的兩條路徑是:(1) 將 Hermes 掛到 Matrix/Discord/Slack 工作區,用這些平台自己的 Web 用戶端;(2) 用相容 MCP 的編輯器連接到 Hermes 的 MCP server 模式(v0.6.0 新增)。
可以,Linux VPS 是最常見的部署方式。官方要求是:最低 2 GB 記憶體(跑本地模型需要更多)、約 1 GB 磁碟、Python 3.11(由 curl 指令碼自動安裝)。SSH 登入後跑同一條安裝命令即可。若擔心命令隔離,建議開啟 README 中的 SSH 或 Docker 沙箱後端,讓 Hermes 的 shell 工具跑在容器裡而不是直接落在你 VPS 的檔案系統上。
上游官方文件位址是 hermes-agent.nousresearch.com/docs,這是安裝、設定、工具、頻道和 CLI 參考的權威來源。本社群站(hermesagents.net)所有事實均來自這套官方文件、NousResearch/hermes-agent 的 README,以及 GitHub 發布說明。若上游與本站資訊不一致,請以上游為準。
不是。市面上有若干不相關的產品也叫「Hermes」——包括企業安全 agent、郵件用戶端、中介軟體等。本站只介紹 Nous Research 出品的 Hermes Agent:一個開源的自我改進 AI agent,倉庫在 github.com/NousResearch/hermes-agent。如果你找的是某款 Hermes 安全產品,請去那個產品的官網。
Teams launched in v0.12.0 (March 2026) as the first plugin-shipped messaging platform. v0.14.0 (May 2026) wired the full Microsoft Graph stack end-to-end: Graph auth + client foundation, a webhook listener that receives Teams events, a pipeline plugin runtime, and outbound delivery via the Teams adapter. Register a Microsoft Graph app, paste the credentials into `hermes gateway setup`, and the bot can read messages and post back in any Teams channel, DM, or group.
Both were added as first-class platforms in v0.14.0 (May 16, 2026). LINE runs via the official LINE Messaging API and brings Hermes to Japan, Korea, and Taiwan markets where LINE dominates. SimpleX Chat is the privacy-focused decentralized messenger with no user IDs — Hermes can run on it without exposing identifiers. They bring total messaging support to 22 platforms.
Yes — v0.14.0 introduced `hermes proxy`. Run it once and you get a local OpenAI-compatible HTTP endpoint backed by whichever OAuth provider you're signed into — Claude Pro, ChatGPT Pro, or SuperGrok. Codex CLI, Aider, Cline, Continue, and anything else that accepts an OpenAI-compatible endpoint can hit it. One subscription, every tool, no API key required.
Yes — v0.14.0 made `/handoff` actually transfer sessions live. Mid-conversation, hand off to a different model, personality, or profile and Hermes moves the entire session — every message, every tool call, every piece of context — without dropping anything. Useful for switching from a fast model to a deep-reasoning one mid-debugging, or passing a session between profiles for different parts of a task.
Yes, via `/goal`. You specify a target and success criteria; the agent runs in a loop — worker proposes, a separate judge LLM evaluates — iterating until the judge confirms the criteria are met. v0.14.0 added `/subgoal` so you can append new criteria to an active loop mid-flight without restarting (#25449). The pattern is called the Ralph loop internally; it works best for tasks with concrete, gradable success conditions and is a poor fit for fuzzy goals where the judge can't grade consistently.
v0.14.0 (May 16, 2026) added an early-beta native Windows build that installs via a PowerShell one-liner — no WSL needed. It bundles MinGit (~45 MB, no admin) for shell command execution; the CLI, gateway, TUI, and tools all run natively. WSL2 remains the production-recommended path because it's been battle-tested longer; native Windows is functional for casual use but has ongoing platform-specific fixes. The only Hermes feature that currently needs WSL2 specifically is the browser-based dashboard chat pane, which uses a POSIX PTY.
Use the built-in cron scheduler: `hermes cron add`. You specify a name, a schedule (cron syntax like `0 8 * * MON-FRI` or natural language), a natural-language description of the job, and delivery targets. Hermes spawns an isolated agent session at the scheduled time, runs the job, and delivers the output to your messaging gateway. v0.14.0 added `deliver=all` (#21495) to fan output to every connected platform at once — useful for nightly backups, weekly audits, daily reports.
Yes. The `computer_use` tool drives the host's GUI via a cua-driver backend. v0.14.0 made this work with non-Anthropic models (#21967, #24063) — previously it was locked to Claude. Any vision-capable model (GPT-5, Gemini, Grok-vision) can now take screenshots, click, type, and drag. Focus-safe operations prevent the agent from fighting your active window. The driver refreshes automatically on `hermes update`.
No. Hermes supports OAuth login for Claude Pro, ChatGPT Pro, and SuperGrok — sign in with your existing subscription, no API key, no separate billing. v0.14.0 added `hermes proxy` (#25969) which exposes those OAuth providers as a local OpenAI-compatible endpoint, so tools like Aider, Cline, and Codex CLI can use them too. For other providers (Nous Portal, OpenRouter, NovitaAI, NVIDIA NIM, etc.), API keys remain the standard path.
v0.14.0's debloating wave (#24220, #24515) made heavyweight backends lazy-install on first use — the `[all]` extras dropped everything covered by lazy-deps, so a base install is meaningfully smaller. The cold start wave (#22138 and ~10 related PRs) shaved ~19 seconds off `hermes` launch — the `hermes tools` All Platforms screen alone dropped from 14 seconds to under 1.5. Concrete numbers depend on which extras you install, but the post-v0.14.0 experience is fast where it used to be sluggish.
Run `hermes claw migrate`. The migrator imports your SOUL.md (persona), MEMORY.md/USER.md entries, user-created skills (to `~/.hermes/skills/openclaw-imports/`), command allowlist, messaging platform configs, API keys (Telegram, OpenRouter, OpenAI, Anthropic, ElevenLabs), TTS assets, and workspace AGENTS.md. Two presets: `full` (default, includes secrets) and `user-data` (no secrets). Always run with `--dry-run` first to preview. The setup wizard auto-detects `~/.openclaw` and offers to migrate before first config.
Layered defenses. Container isolation is the primary boundary — the agent runs shell commands inside a sandbox (Docker, Daytona, Modal, SSH, Singularity, Vercel, or local — seven backends total). A command-approval workflow flags dangerous operations (rm -rf, sudo, curl|sh) for explicit user approval. v0.14.0 added three specific hardening fixes: a sudo brute-force block (#23736), three closures of known dangerous-command bypasses (#26829), and tool error sanitization (#26823) that strips injection attempts from error strings before they re-enter the model context. The overall security policy was rewritten in v0.14.0 around OS-level isolation as the primary trust boundary (#20317).