Agent comparisons
OpenAI Dots vs Hermes Agent: OpenAI's Agent or Your Runtime
Choose OpenAI Dots if you already pay for ChatGPT Pro or Business Premium and want OpenAI to run an agent with its own cloud computer, thousands of app connectors, and approval rules you set. Choose Hermes Agent if you want to pick the model, read the skills and memory it works from, and talk to it in Telegram or Slack. A managed Hermes instance stays online like a dot does. It does not copy Dots' connectors or its per-action approvals.
Dots is OpenAI's always-on personal agent, set up in ChatGPT on desktop and powered by GPT-6 Astra. It is not a model you call from an API, and it is not the Agents SDK for developers. Hermes Agent is Nous Research's open-source agent runtime, not the Hermes model family. You can point Hermes at several providers, OpenAI included.
How do OpenAI Dots and Hermes Agent compare?
| Decision | OpenAI Dots | Managed Hermes |
|---|---|---|
| Who runs it | OpenAI, inside your ChatGPT plan | Lobsterland runs the instance; the runtime is open source |
| Model | GPT-6 Astra | A provider you connect: OpenAI, Anthropic, Google, OpenRouter, or MiniMax |
| Its own machine | A cloud computer and browser, separate from yours | Its own instance, plus an optional Hosted Browser |
| Apps | More than 4,000 through OpenAI's plugins | Skills you install and the accounts you sign into |
| Where you talk | ChatGPT, Slack, Microsoft Teams; limited US Pro texting beta | Built-in chat, Telegram, Slack; no Teams |
| Work while you are away | Background goals; read-only research while you are inactive | The instance stays up; no managed cron on Hermes |
| Before a risky action | Custom Rules: allow, require approval, or block | No per-action prompt; allowlists decide who can message it |
| What it learns | Shared ChatGPT memory plus its own retained context | A memory file and skill files you can read and edit |
| Price | First dot included in Pro and Business Premium | Hosting from $6.90/month; model access separate |
The Dots column draws on OpenAI’s announcement, its setup guide, and its memory and safety FAQ. The Hermes column is managed Hermes hosting on Lobsterland, not every option in the upstream project.
When is OpenAI Dots the better fit?
When you want OpenAI to supply the computer, app connections and action-review system. Its cloud browser and plugin catalog reduce the integration work you would otherwise do through skills. Microsoft Teams is another practical difference: Dots supports it; managed Hermes does not.
Dots’ Custom Rules can allow, block or require approval for actions. Auto-review checks consequential actions against your instructions and safeguards. These are documented controls, not a guarantee that an agent will never make a mistake. In Enterprise, workspace administrators control access, and Custom Rules are off by default. Enabling Dots does not grant access to every app.
If your eligible Pro or Business Premium account includes a dot, try it on the intended job before paying for a separate host. Hermes earns that extra cost when provider choice and readable skill files matter to the work.
Can a ChatGPT plan power Hermes instead?
It can supply the model, not the dot. Lobsterland supports ChatGPT/Codex login as a Hermes provider. The instance sends its model requests through your ChatGPT account, within your plan's limits. You get OpenAI models behind a runtime you configure.
That login does not bring over what makes Dots a product: its cloud computer, its app connectors, its Custom Rules, or what it has learned about you. It does not import ChatGPT conversations or memory either. You would be using the same subscription for two different things, and the two do not share state.
A plain OpenAI API key works too, and so do Anthropic, Google, OpenRouter, and MiniMax. That choice is the main thing Hermes gives you that Dots cannot: when a model gets worse at your task, or more expensive, you can switch providers and keep the agent's skills and memory.
What stops each agent before a risky action?
This is the biggest practical difference, and it favors Dots. Dots is built around per-action policy. On managed Hermes, Lobsterland turns off Hermes' command-approval prompt. A hosted agent working through a long task would otherwise stall on a prompt no one is there to answer. So Hermes runs commands it would normally flag, without asking.
The controls you do have come before and after the work. Before: an allowlist decides who can message the agent on Telegram or Slack, and in a group it can be set to reply only when mentioned. You also decide what it can reach: the sites you sign in to in its browser and the credentials you give its skills. After: the logs show what it did. None of that is a rule like "ask me before you send." It also means that anyone you let message the agent can make it run commands, so keep that list short.
So give Hermes work you can check, and hold back the access you cannot afford to have misused. Ask for a draft reply instead of a sent one, or a proposed change instead of a committed one. Keep payment methods out of the Hosted Browser profile unless the task needs them. If a workflow has to stop for approval, Dots' Custom Rules fit it better than anything we host today.
What happens when your agent lives in Slack?
Both products work in Slack, but message volume depends on the integration. Slack lets bots edit their own messages; the problem we investigated was Hermes posting separate updates. In the incident we documented on September 7, 2026, one user message produced about 44 bot messages. Two separate settings emitted tool progress and interim assistant messages.
Hermes' own Slack default is quiet, but a global setting written at boot overrode it. Every managed Hermes instance now pins Slack progress updates, reasoning, and interim messages off. The periodic "Working" notice stays on, so a long task still shows it is alive. Our Hermes Slack guide explains both causes and the fix for self-hosted setups.
We cannot tell you how a dot behaves in a busy Slack channel, because we have not run one. Before you add either agent to a shared channel, test it in a private one with a long task.
How do memory and skills differ?
Dots can share memory with ChatGPT, but its own saved context is separate. OpenAI’s memory FAQ, checked October 16, 2026, says you cannot directly view or edit individual dot memories. Deleting the dot removes its context; shared ChatGPT memories and files stored elsewhere have separate controls. Disconnecting an app stops new access without erasing context already retained.
Hermes keeps two things you can inspect. Its curated memory is a small file of facts the agent chose to keep. Its skills are SKILL.md files that spell out how to do a task, and you can install, read, and edit them from the dashboard. The Hermes skills guide shows how.
That memory has a size limit, and it has caused real failures. The Hermes version in the incident we documented on September 2, 2026 capped curated memory at 2,200 characters. When the file is nearly full, new writes fail. So does the consolidation meant to free space, because it is checked against the same cap. One of our production agents lost eight of nine memory writes that way. Managed Hermes now sets the limit to 8,000 characters. The Hermes memory limit fix covers the self-hosted case.
What do access and cost look like?
As of October 16, 2026, OpenAI’s setup guide includes a first dot with eligible Pro and Business Premium plans. Deeper work has an allowance, with extended limits for the first month after launch; this is not unlimited work. Pro excludes the EEA, Switzerland and UK, while Business Premium covers supported ChatGPT regions. Enterprise, Edu and Healthcare need administrator-enabled beta access.
You can initiate a voice call to your dot, but it cannot initiate calls at launch. Texting is a limited beta for some US Pro users, not Business or Enterprise. Check the controls in your account before choosing Dots for a phone workflow.
Lobsterland's Hermes hosting starts at $6.90 per month, with a seven-day trial. A card is required up front, and the hosting charge only happens if you keep the plan after the trial. Model access is separate: use a provider API key or supported ChatGPT/Codex login within your plan’s limits. For Claude, use an Anthropic API key as the default. A connected Claude subscription is not a promise of included hosted usage; Anthropic restricts third-party subscription use, and hosted traffic may incur additional metered charges. The hosting pricing page lists the capacity tiers. If you already pay for Pro, Dots may cost you nothing extra. Hermes would then be an added cost you pay for control over the model and method.
How do you test Hermes against a Dots-style task?
Start with a bounded alert investigation that needs no account access. Paste a redacted alert and the relevant log lines. Then ask: "Tell me what most likely broke, quote the lines that support it, list what you would check next, and change nothing."
- Launch Hermes and connect one model. ChatGPT/Codex login is fine if you have it. Run the prompt in built-in chat.
- Check the quotes. Every conclusion should point to a line you pasted. A confident diagnosis without quotes fails the test.
- Save the method as a skill. Once the answer has the shape you want, ask the agent to save it and read the file. On the next alert, it should follow that file.
- Connect Slack or Telegram last, in a private channel first. Give it more access only after you have read the logs from a few runs.
Here is everything managed Hermes includes: built-in chat, Telegram, Slack, skills, logs, upgrades, Paperclip, and an optional Hosted Browser. The native dashboard needs image v2026.6.5 or later. There is no managed cron, workspace-file editor, multi-agent routing, usage analytics, WhatsApp, or Teams. If background schedules and several agents are what drew you to Dots, OpenAI Dots vs OpenClaw compares the runtime that has them. OpenClaw vs Hermes explains how to choose between the two runtimes.
If you are comparing plain ChatGPT and its Work agent rather than Dots, read ChatGPT vs Hermes Agent.
Frequently asked questions
Is OpenAI Dots the same thing as the OpenAI Agents SDK?
No. Dots is OpenAI's always-on personal agent, launched inside ChatGPT on September 29, 2026. The Agents SDK is a developer library for building agents into your own product. This page compares Dots with Hermes Agent, Nous Research's open-source agent runtime.
Can my ChatGPT plan power Hermes instead of Dots?
It can supply the model. Lobsterland supports ChatGPT/Codex login for Hermes, subject to your plan's limits. That login does not bring your dot, its cloud computer, its app connectors, or its approval rules into Hermes.
Does managed Hermes ask before it runs a risky command?
No. Lobsterland ships Hermes with its command-approval prompt turned off, so hosted agents can finish work without waiting for you. Your controls are who may message the agent, which accounts you connect, and the logs you review afterwards.
Is OpenAI Dots available in the UK and the EU?
OpenAI’s setup guide, checked October 16, 2026, excludes the European Economic Area, Switzerland, and the UK from Pro access. Business Premium covers supported ChatGPT regions. Enterprise beta access requires an administrator to enable it. Availability is subject to rollout.