Agents · late September 2026智能體 · 2026 年 9 月下旬
Hermes and OpenClaw are open-source personal agents you run yourself. Claude Code, Codex, Cursor, and Dots are vendor products aimed at code or at a day that continues without you. They are not substitutes. They share a risk: a long-running process that holds keys and can touch your accounts.Hermes 與 OpenClaw 是你自己跑的開源個人智能體。Claude Code、Codex、Cursor 與 Dots 是廠商產品，瞄準程式，或瞄準你不在時仍繼續的一天。它們不能互相取代。它們共享一種風險：一個長時間執行、握有金鑰、碰得到你帳號的行程。
| Agent智能體 | Who maintains it誰維護 | Center of the design設計中心 | Where it runs在哪裡跑 | Models模型 | Reach碰得到哪 |
|---|---|---|---|---|---|
| Hermes Agent | Nous Research | A learning loop. It stores memory and can write its own skills as Markdown.一個學習迴圈。它儲存記憶，並能把技能寫成 Markdown。 | Your machine. Python. MIT license.你的機器。Python。MIT 授權。 | Many providers, plus a local OpenAI-compatible endpoint.許多供應商，加上相容 OpenAI 的本機端點。 | Telegram, Discord, Slack, WhatsApp, Signal, email, a terminal, and a desktop app. Several sandbox backends.Telegram、Discord、Slack、WhatsApp、Signal、電子郵件、終端機與桌面應用。好幾種沙盒後端。 |
| OpenClaw | OpenClaw Foundation | A local gateway: sessions, tools, events, and channels in one control plane.一個本機閘道：工作階段、工具、事件與頻道，在同一個控制平面。 | Your machine. TypeScript. MIT license. No paid tier from the foundation.你的機器。TypeScript。MIT 授權。基金會沒有付費層。 | Many providers, plus local models. Ollama can launch it.許多供應商，加上本機模型。Ollama 可以啟動它。 | A very wide chat list, native companion apps, voice, and device nodes.很寬的聊天清單、原生陪伴應用、語音，以及裝置節點。 |
| Claude Code | Anthropic | A coding agent in the repo. Not a general life assistant.儲存庫裡的程式智能體。不是通用生活助理。 | Terminal, IDE, web. Commercial.終端機、IDE、網頁。商業。 | Claude. Opus 5.5 is the strongest checked coding model this week.Claude。本週已核對的程式模型裡，Opus 5.5 最強。 | The repository, tests, and the review tools around them.儲存庫、測試，以及周圍的審查工具。 |
| Codex | OpenAI | A coding agent, including cloud sandboxes.程式智能體，包含雲端沙盒。 | Terminal, IDE, cloud. Commercial.終端機、IDE、雲端。商業。 | GPT-6 family. Sol for value, Astra for the top rate.GPT-6 家族。Sol 看價效，Astra 是最高費率。 | The repository and OpenAI's cloud computer for the task.儲存庫，以及 OpenAI 為該任務準備的雲端電腦。 |
| Dots | OpenAI | Always-on agents that keep a goal moving. Announced 29 Sep 2026.讓目標繼續走的常駐智能體。2026 年 9 月 29 日宣布。 | OpenAI's cloud. Tied to paid ChatGPT plans.OpenAI 的雲端。綁在付費的 ChatGPT 方案。 | OpenAI models.OpenAI 的模型。 | Apps you connect. New enough that the safe move is a narrow goal and a separate account.你連接的應用。新到安全做法是窄目標與分開的帳號。 |
| Cursor | Anysphere | An editor where you watch the diff.你可以看著差異的編輯器。 | Desktop IDE. Commercial.桌面 IDE。商業。 | Several vendors, including Cursor's own Composer models.好幾家供應商，包括 Cursor 自己的 Composer 模型。 | The files you have open. You stay in the loop.你開著的檔案。你留在迴圈裡。 |
| Aider and Cline | Open source開源 | Repo edits from a terminal (Aider) or an editor (Cline), easy to point at a local model.從終端機（Aider）或編輯器（Cline）改儲存庫，很容易指向本機模型。 | Your machine.你的機器。 | Whatever speaks an OpenAI-compatible API, including Ollama.任何說相容 OpenAI API 的東西，包括 Ollama。 | The repository. Best when the model fits in your hardware.儲存庫。模型放得進硬體時最好。 |
Project details checked against public comparisons and docs in the last week of September 2026. Star counts move and are a poor way to choose, so they are not listed. Repositories: Hermes, OpenClaw, Aider, Cline.專案細節依 2026 年 9 月最後一週的公開比較與文件核對。星星數會動，也不是好的選擇方式，所以不列。儲存庫：Hermes、OpenClaw、Aider、Cline。
Research sprints, a lab notebook, a support process you refine. Hermes. Let it write a skill after the third time you explain the same procedure. Review the skill file. It is code.研究衝刺、實驗筆記、你會反覆改的支援流程。Hermes。 第三次解釋同一套程序之後，讓它寫下技能。審查那個技能檔。它是程式。
WhatsApp, Telegram, Slack, iMessage, a phone. OpenClaw. Start with one channel and an allowlist. Add voice only after the text path is boring and safe.WhatsApp、Telegram、Slack、iMessage、一支手機。OpenClaw。 從一個頻道與允許清單開始。文字路徑無聊且安全之後，再加語音。
Legal notes, unpublished numbers, health files. Hermes or OpenClaw on a local model from the advisor. No cloud fallback. A folder the agent may read, and no other folder.法律筆記、未公開數字、健康檔案。用顧問選的本機模型跑 Hermes 或 OpenClaw。不要雲端後援。智能體可以讀一個資料夾，沒有第二個。
A real repository and a reviewer. Claude Code or Codex for the long task, Cursor when you want to see every edit. Require tests. Read the diff.真正的儲存庫與一位審查者。長任務用 Claude Code 或 Codex；想看見每一次修改用 Cursor。要求測試。讀差異。
Cursor or Aider with a cheaper model, and you apply the changes. An always-on agent hides the lesson and spends the budget.用較便宜的模型配 Cursor 或 Aider，變更由你套用。常駐智能體會把課藏起來，並花掉預算。
Dots if you accept OpenAI's computer and can state a finish line. Otherwise do not. "Keep working" without a finish line is how agents send the wrong email.若你接受 OpenAI 的電腦，也能說出終點，用 Dots。否則不要。沒有終點的「繼續工作」，就是智能體寄錯信的方式。
A chat model for the script, then a tool from the image and video page. A coding agent is the wrong hire for a storyboard.腳本用聊天模型，然後用影像頁的工具。分鏡不該雇一個寫程式的智能體。
GitHub Copilot's coding agent, or Claude Code / Codex opening pull requests. The review culture you already have matters more than the brand.GitHub Copilot 的程式智能體，或由 Claude Code／Codex 開拉取請求。你已經有的審查文化，比品牌重要。
Both Hermes and OpenClaw had published security advisories in 2026. So did plenty of IDE plugins. The shape of the risk is the same: messages from other people, a shell, and your API keys in one process.Hermes 與 OpenClaw 在 2026 年都有公開的安全通報。很多 IDE 外掛也有。風險的形狀相同：別人的訊息、一個殼層，以及你的 API 金鑰，在同一個行程裡。
More on the security page.更多在安全頁。