# How do Muse, dots, Claude Duet, OpenClaw and Hermes compare?

Date: 2026-11-01 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/personal-agents-compared-2026/

**Answer:** Muse, dots and Claude Duet are managed services that run on the company's cloud computers, aimed respectively at consumers, paying ChatGPT users and professionals. OpenClaw and Hermes Agent are free open-source programs you run yourself with a model of your choice, which gives you more control and leaves more of the safety work to you.

A **personal agent** is an AI assistant that keeps running when you close the app, connects to your accounts and takes actions for you. This fall, the choice widened from two open-source projects to several big-company products. They look alike in demos. The differences that matter are where the agent runs, who can get it and what happens before it acts.

Everything below reflects what each company has published or what reporters have described as of early November 2026. These products change monthly, so treat the table as a starting point and check the current terms.

## Side by side

| | Muse (Meta) | dots (OpenAI) | Claude Duet (Anthropic) | OpenClaw | Hermes Agent |
|---|---|---|---|---|---|
| **Launched** | Sept. 8, 2026 | Sept. 29, 2026 | Oct. 21, 2026, research preview | Jan. 2026 under this name | Feb. 2026 |
| **Runs on** | A dedicated cloud VM | Company's cloud, within ChatGPT or Codex | A cloud VM with its own browser | Your machine or server | Your machine, server or container |
| **Model** | Muse Spark | GPT-6 Astra | Claude Opus 5.5 | Your choice | Your choice |
| **Aimed at** | General consumers | Individuals and companies | Professionals, small business owners | Technical users | Technical users |
| **Price** | Free, or $20 or $100 (US) a month | Available on Pro and Business Premium plans | Included in Pro, Max, Ultra, sharing usage limits | Free software; you pay for the model | Free software; you pay for the model |
| **Where** | US and Canada, plus others being added | "Eligible markets" | US only, by waitlist | Anywhere | Anywhere |
| **Reach it through** | App, web, WhatsApp | ChatGPT, Codex, Slack, Teams | Its own app, web | Many chat apps | Terminal and several chat apps |
| **Approval** | Before sensitive actions; separate Sentinel gatekeeper | Controls include a read-only mode | Before sending, paying, signing, deleting | Configurable pairing and tool limits | Command approval, pairing |

A **virtual machine** (VM) is a software-defined computer running on a company's server. Giving each user one lets the agent keep working when your phone is off, and keeps it apart from other users' agents. It does not, by itself, keep your data away from the company running it.

Google's Gemini Spark belongs in this group too. Google unveiled it in May as an agent meant to run around the clock, starting with testers and a planned beta for US Ultra subscribers. Other entrants include X's Grok Bot and Instinct. They are left out of the table because their current terms are harder to pin down.

## How they differ in practice

**Muse** goes furthest toward everyday use, with travel, bills, forms and shopping. Meta says purchases can use a one-time card through Stripe's Link, and that the agent never sees your passwords. According to [The Batch](https://www.deeplearning.ai/the-batch/how-to-secure-agents-for-the-masses), the free tier allows up to 100 million tokens a week, with 500 million at $20 and 3 billion at $100. TechCrunch notes that you need a payment card on file to start. Your interactions train Meta's models unless you opt out, and Meta says it is planning a Confidential VM later this year, encrypted with a key only you hold.

**dots** come from OpenAI's push to make agents work continuously toward a goal. You name your primary dot and message it through ChatGPT, Codex or workplace tools such as Slack and Teams. OpenAI says they connect to more than 4,000 apps, per [Al Jazeera](https://www.aljazeera.com/economy/2026/9/30/openai-launches-dots-personal-ai-assistant-built-to-handle-everything), and that it envisions teams of dots later. Access starts with Pro and Business Premium users. Usage terms for each plan were still settling after launch, so check OpenAI's current page.

**Claude Duet** is the most work-centred. It sends a morning brief, checks your connected accounts every few hours and can use more than 6,000 connectors for tools like QuickBooks, Gusto and Docusign. Anthropic says outright that personal errands such as booking tickets are weaker, and that it refuses trades and transfers between accounts even with approval. Its background checks draw down your plan's usage limits. Because it checks periodically, Anthropic warns it can miss something that arrives in between. A [reviewer at Every](https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent), writing with early access, rated its email drafts below Dot's.

**OpenClaw and Hermes** remain the option for people who want to choose the model, keep files on their own disk and read the code. You can run them with a local model so that nothing leaves your machine, but that means a smaller model and more work. Our [detailed comparison](/en/learn/openclaw-hermes-agent-personal-agent-fits/) covers how they differ from each other, and we have setup guides for [OpenClaw](/en/learn/set-up-openclaw-safely-own-computer/) and [Hermes](/en/learn/set-up-hermes-agent/).

## A worked example

Three illustrative people, in Canada:

| Person | Wants | Likely fit | Catch |
|---|---|---|---|
| Sophie, a parent in Gatineau | Family logistics, reminders, some shopping | Muse | Trusting Meta with these accounts; opt out of training if you wish |
| Karim, an accountant in Montréal | Inbox triage, unpaid invoices, client data | Duet's design fits, but it is US-only for now | Client confidentiality rules may rule out any of these; check first |
| Lin, a developer | An agent that learns her deployment routine | Hermes or OpenClaw on a small server | She has to patch and watch it herself |

Karim's case shows why the availability row matters. The best match on paper may not be open to you yet, and a tool that is available still has to pass your professional obligations and, in Quebec, Law 25.

> **Try it:** Write your top three tasks for an agent. Mark each one R (reading only), D (drafting) or A (acting without you). Most of the benefit often sits in R and D, which any product above can handle at far lower risk.

## What none of them has solved

The same weaknesses run through all five. All of them read outside content, so all of them are exposed to **prompt injection**, hidden instructions that the model may follow. Vendors' defenses are real, but not fully independently tested. Reuters reported that testers inside Meta [found flaws](https://www.carriermanagement.com/news/2026/09/09/291788.htm) in Muse, and Anthropic says Duet may misread a thread or misjudge urgency. Early user reports on reliability are mixed for several of these products. Published evaluations with shared measures barely exist, which makes a fair ranking impossible for now. Independent figures for how often any of them makes a costly mistake do not seem to exist yet.

## What should you check before connecting your first account?

Go through this list for the specific account, not the product in general:

1. **Read or act?** Grant read-only access if the product offers it. Muse, for example, lets you choose whether it reads your email or can also send.
2. **Can you undo it?** Connecting a calendar is a small step. Keep your main email, banking and password manager out at first, for the reasons given in [our guide to agent risks](/en/learn/go-wrong-personal-agent-that-has-accounts/).
3. **Is your data used for training?** Look for the setting on the day you sign up. Muse requires an opt-out. Check the others' current terms.
4. **Can you see what it did?** Look for an activity log. Meta advertises an audit trail, and Anthropic says you can see what Duet has done and steer it.
5. **Can you disconnect and leave?** Test disconnecting one service on day one. Then read [who owns what your agent remembers](/en/learn/who-owns-your-personal-ai-memory/) before it has learned much about you.

## Sources

- Introducing Muse: The World's First Personal AI Agent Built for Everyone: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
- Introducing Claude Duet: https://claude.com/resources/articles/claude-duet
- OpenAI launches dots, its bubbly agentic avatar: https://techcrunch.com/2026/09/29/openai-launches-dots-its-bubbly-agentic-avatar/

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/personal-agents-compared-2026/, no commercial use without permission, and keep the same licence.