Hermes Agent

English

Overview

Hermes Agent is an open-source AI agent from Nous Research. It is designed to run as a persistent assistant on a laptop, server, VPS, or cloud backend while users interact with it from a terminal or messaging channels such as Telegram, Discord, Slack, WhatsApp, Signal, email, or CLI.

Its 2026-facing positioning is broader than a coding autocomplete tool. Hermes combines a terminal agent, messaging gateway, tool system, persistent memory, skills, scheduled jobs, model routing, and isolated subagents. The project emphasizes self-hosting, model choice, and a learning loop where the agent can preserve useful knowledge and improve reusable skills over time.

Why it matters

Most coding assistants are tied to one surface: an IDE, a web app, or a vendor-hosted cloud workspace. Hermes is useful when the agent should keep working somewhere durable and reachable from multiple channels. A developer can run it on a low-cost server, connect tools and providers, schedule reports or audits, and continue conversations from a phone while the agent executes in its configured environment.

Hermes also matters for agent research and advanced automation. It supports tool-calling workflows, MCP connections, terminal backends, subagent parallelism, skills, trajectory generation, and long-lived memory. The tradeoff is operational responsibility: because Hermes is self-hosted and tool-capable, users must manage secrets, command approvals, isolation, updates, and provider costs carefully.

Architecture/Concepts

• CLI/TUI: the hermes command starts an interactive terminal interface with conversation history, slash commands, streaming tool output, and model/tool configuration.
• Messaging gateway: one gateway process can connect Hermes to chat platforms, so users can delegate work from messaging apps while execution happens on the configured host.
• Model routing: Hermes is model-agnostic and can use providers such as Nous Portal, OpenRouter, OpenAI, Anthropic, Kimi/Moonshot, MiniMax, z.ai/GLM, Hugging Face, and compatible custom endpoints.
• Tools and toolsets: tools cover terminal, filesystem, browser/search, media, MCP, and other agent capabilities. Toolsets help control what is enabled for a context.
• Persistent memory and skills: Hermes can store durable memory, search previous sessions, and use skills as reusable procedural knowledge.
• Subagents: large tasks can be split into isolated child agents with their own workspace and tool context.
• Cron scheduling: natural-language recurring jobs can run unattended and report back through configured channels.
• Execution backends: terminal sessions can run locally or through Docker, SSH, Daytona, Singularity, Modal, and related environments.

Practical usage

Install from the official repository instructions:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.zshrc
hermes

Common commands:

• hermes starts an interactive session.
• hermes setup runs guided configuration.
• hermes model selects provider and model.
• hermes tools configures enabled tools.
• hermes gateway manages messaging integrations.
• hermes update updates the installation.
• hermes doctor checks common setup problems.

Good first workflows:

• Run Hermes locally with filesystem and shell permissions constrained to a test project.
• Configure one model provider and verify cost controls before enabling more tools.
• Add a messaging gateway only after the CLI workflow is stable.
• Use cron for low-risk recurring reports before scheduling write-capable jobs.
• Use Docker, SSH, or another isolated backend for tasks that run untrusted commands.

Learning checklist

• Install Hermes and run hermes doctor.
• Configure one model provider and switch models with hermes model.
• Start a CLI session and learn /new, /model, /skills, /usage, and /compress.
• Enable only the tools needed for one small project.
• Configure the messaging gateway for one private channel.
• Create or use one skill for a repeatable workflow.
• Schedule a simple report with the cron system.
• Run a task in an isolated backend such as Docker or SSH.
• Review security settings, command approvals, and secret handling.
• Track release notes before upgrading because the project is changing quickly.

繁體中文

概覽

Hermes Agent 是 Nous Research 推出的開源 AI Agent。它可以長時間運行在筆電、伺服器、VPS 或雲端後端，使用者則能透過終端機或 Telegram、Discord、Slack、WhatsApp、Signal、email、CLI 等通道與它互動。

它在 2026 年的定位不只是程式補全工具，而是結合終端機 Agent、訊息 Gateway、工具系統、持久記憶、Skills、排程、自動化、模型路由與隔離 Subagents 的自架式 Agent Runtime。Hermes 強調自架、模型選擇，以及能把經驗保存成記憶與可重用技能的 learning loop。

為什麼重要

許多 coding assistant 綁定在單一介面，例如 IDE、Web app 或供應商雲端工作區。Hermes 適合需要 Agent 長時間待在可控環境中工作的情境。開發者可以把它部署在低成本伺服器上，接上工具與模型供應商，安排報告或稽核任務，並從手機訊息通道持續追蹤工作。

Hermes 也適合 Agent 研究與進階自動化。它支援 tool calling、MCP、terminal backends、subagent parallelism、skills、trajectory generation 與長期記憶。代價是使用者必須自己負責營運安全：包含 secrets、command approval、隔離、更新與模型成本。

架構/概念

• CLI/TUI： hermes 命令啟動互動式終端介面，支援對話歷史、slash commands、串流工具輸出與模型/工具設定。
• Messaging Gateway： 單一 gateway process 可連接多個聊天平台，讓使用者從訊息 App 指派工作，而實際執行發生在設定好的主機上。
• 模型路由： Hermes 不綁定單一模型，可使用 Nous Portal、OpenRouter、OpenAI、Anthropic、Kimi/Moonshot、MiniMax、z.ai/GLM、Hugging Face 與相容 endpoint。
• Tools 與 Toolsets： 工具涵蓋 terminal、filesystem、browser/search、media、MCP 等能力；toolsets 用來控制不同情境可用工具。
• 持久記憶與 Skills： Hermes 可保存長期記憶、搜尋過去 sessions，並用 skills 保存可重複使用的程序知識。
• Subagents： 大型任務可拆給彼此隔離的 child agents，各自有自己的工作區與工具脈絡。
• Cron 排程： 可用自然語言設定定期工作，並把結果送回指定通道。
• 執行後端： 終端環境可在 local、Docker、SSH、Daytona、Singularity、Modal 等後端執行。

實務使用

依官方 repository 安裝：

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.zshrc
hermes

常用命令：

• hermes 啟動互動 session。
• hermes setup 執行設定精靈。
• hermes model 選擇 provider 與模型。
• hermes tools 設定可用工具。
• hermes gateway 管理訊息平台整合。
• hermes update 更新安裝。
• hermes doctor 檢查常見設定問題。

建議入門流程：

• 先在本機以受限 filesystem 與 shell 權限測試一個專案。
• 只設定一個模型供應商，確認成本與限制後再擴充。
• CLI 穩定後再加入 messaging gateway。
• 先用 cron 做低風險定期報告，再安排可寫入或可改動系統的工作。
• 對會執行不可信命令的任務使用 Docker、SSH 或其他隔離後端。

學習檢核表

• 安裝 Hermes 並執行 hermes doctor。
• 設定一個模型供應商，並用 hermes model 切換模型。
• 啟動 CLI session，熟悉 /new、/model、/skills、/usage、/compress。
• 只為一個小專案啟用必要工具。
• 為一個私人通道設定 messaging gateway。
• 建立或使用一個可重複 workflow 的 skill。
• 用 cron 排程一個簡單報告。
• 在 Docker 或 SSH 等隔離後端執行任務。
• 檢查安全設定、command approvals 與 secrets 管理。
• 升級前追蹤 release notes，因為專案仍快速變動。

References

• Hermes Agent GitHub repository - NousResearch
• Hermes Agent releases - GitHub
• Hermes Agent overview
• Nous Research GitHub organization