Dots, Grok Bot, Muse, Claude Duet, Hermes or OpenClaw: which agent, and which web can it use?
By Dean, Known Good · Published 4 Oct 2026 · Updated 6 Nov 2026 (Claude Duet added) · vendor facts as of 4 Oct 2026, Claude Duet as of 28 Oct 2026
Six agents now run on their own computer and act for you while you are away. Dots (OpenAI), Grok Bot (xAI), Muse (Meta) and Claude Duet (Anthropic) are hosted; Hermes (Nous Research) and OpenClaw are open source and self-hosted. All six reach websites through a browser; five also call tools directly. 57,273 websites are listed as agent-ready, each probed and dated; 48,822 of the sites probed answer an agent's tool calls directly.
The short version
| Dots | Grok Bot | Muse | Claude Duet | Hermes Agent | OpenClaw | |
|---|---|---|---|---|---|---|
| Vendor | OpenAI | xAI | Meta | Anthropic | Nous Research | OpenClaw Foundation |
| Announced | DevDay 2026 | 11 Aug 2026 (beta) | 8 Sep 2026 | 21 Oct 2026 (research preview) | open source, MIT | 30 Jan 2026 (as OpenClaw) |
| Runs on | its own computer and browser in the cloud | a persistent cloud computer with a browser, command line and files | a dedicated cloud VM ("Muse Secure VM") | a virtual machine with its own browser, in Anthropic's cloud | your machine, a VPS, Docker, or serverless | your machine |
| Reaches websites by | browser; the plugins enabled for your account | connectors; the browser where there is none | app connections you choose; checkout with Link by Stripe | connectors and plugins from the Claude Marketplace; the browser; checkout with Link (Stripe) or Shop Pay | MCP servers, browser, skills | MCP and messaging; model of your choice |
| Reads WebMCP in the page | not stated for Dots; ChatGPT Work and Codex do in the desktop browser (week of 24–28 Aug 2026) | not listed | not listed | not stated | not listed | not listed |
| Price (as of 4 Oct 2026) | with ChatGPT Pro and Business Premium plans, in specified regions | SuperGrok and Cursor plans (beta); enterprise by waitlist | free for most uses; paid plans for more | included in Claude Pro and Max; waitlist, US first (as of 28 Oct 2026) | free (MIT); you pay the model | free (MIT); you pay the model |
| Acts without asking | Custom Rules set ongoing boundaries; built-in safety checks stay on | stops for approval before sending, publishing, purchases and deleting | checks before sensitive actions; complete audit trail | messages you first; shows the plan and waits before sending, paying, signing or deleting; some actions refused even with approval | your configuration | your configuration |
| Best for | people already on ChatGPT Pro who want one agent that persists | teams on Cursor or SuperGrok who want bots that persist | people who want a personal agent at a low price with purchase protection | people on Claude Pro or Max who want an agent that watches their accounts and comes to them | developers who want to own the agent and its data | people who want the largest open-agent community and accept the security work |
Sources: the vendors' own announcements and docs, linked at the foot of the page. "Not listed" and "not stated" mean the vendor's page does not say, as of 4 Oct 2026 (Claude Duet: 28 Oct 2026).
What they have in common
Each of the six is an agent that keeps working after you close the chat. Each gets a computer of its own: OpenAI gives each dot "its own computer and browser in the cloud"; xAI gives each account "a persistent cloud computer"; Meta houses the agent and your data in a dedicated VM; Anthropic gives Claude Duet "a virtual machine with its own browser"; Hermes and OpenClaw run wherever you install them. Each reaches the web the same two ways: through a browser, where it reads pages as a person would, and through tools, where a site or an app answers it directly. The second way is faster, cheaper and far less likely to go wrong, and it only works on sites that have built for it.
How much of the web answers an agent directly
This is the part the vendors cannot tell you, because it is about the web, not the agent. Known Good probes the web and publishes what answered. As of 4 Nov 2026: 57,273 websites are listed as agent-ready. 48,822 of the 49,448 websites we probed for MCP answer it, the protocol Hermes, OpenClaw, Dots' connectors and most agent tooling speak. 9,178 serve markdown when an agent asks for it. 541 of the 606 sites we runtime-tested register tools in the page itself (WebMCP), measured 31 Aug 2026. Every figure is dated on its own page: MCP, WebMCP.
An agent with tool access can use those sites without a browser. An agent with only a browser can use any site a person can, slowly, and breaks when the page changes. Whichever agent you pick, pointing it at the sites that answer directly is the difference between an errand that takes seconds and one that takes minutes and fails.
Dots (OpenAI)
Announced at OpenAI's DevDay 2026 and running on GPT-6 Astra. Each dot has its own computer and browser in the cloud, and can connect to one personal computer at a time while the ChatGPT app is open. It uses the plugins installed and enabled for your account, asks you to confirm before it uses a saved login for a new sign-in, and works in the background inside Custom Rules, which set ongoing boundaries without overriding the built-in safety checks. It is available on the Pro 100, Pro 200, Pro 500 and Business Premium plans, to users over 18 in specified regions; on Enterprise it is rolling out off by default, for an administrator to enable. OpenAI's Site tools documentation does not mention dots, so whether a dot's browser reads WebMCP is not stated; ChatGPT Work and Codex use a site's WebMCP tools in the desktop app's built-in browser since the week of 24–28 Aug 2026.
Best for: someone already paying for ChatGPT Pro who wants one persistent agent with OpenAI's plugins behind it.
Watch: specified regions and users over 18 only; Enterprise is off until an administrator turns it on.
Grok Bot (xAI)
In beta since 11 Aug 2026 for SuperGrok, SuperGrok Plus and SuperGrok Heavy, Cursor Pro, Pro+ and Ultra, and Cursor Teams Standard and Premium; enterprise access is through a waitlist. Each account gets a persistent cloud computer with a browser, command line and files. It uses connectors, and the browser for services without one. It stops for approval before sending messages, publishing, purchases and financial transfers, and deleting or overwriting data, and you enter passwords, passkeys, two-factor codes, CAPTCHAs and payment confirmations yourself. xAI's release notes describe it as "durable AI teammates that work on a persistent cloud computer, with messaging, approvals, connectors, and routines".
Best for: teams already on Cursor or SuperGrok who want agents that keep working between sessions.
Watch: xAI says not to treat separate bots as a security boundary; its pages name no plugins and do not mention MCP.
Muse (Meta)
Announced 8 Sep 2026, rolling out in the US on iOS, Android and muse.ai, with AI glasses to follow. Each person's Muse runs in a dedicated cloud VM that holds both the agent and their data; Meta says it does not share conversations or VM data with its ad systems, and you can opt out of training. You choose which apps it connects to and how much access each gets. It buys things through Link by Stripe with purchase protection, checks with you before sensitive actions, and shows a full audit trail. Free "for most of what people need", with subscription plans above that.
Best for: a personal agent that shops and books for you at consumer prices, with the brakes on by default.
Watch: US only at launch; the paid tiers are on the product page, not the announcement, so check them before you buy.
Claude Duet (Anthropic)
Announced 21 Oct 2026 as a research preview, in its own app on iOS, Android and claude.ai, separate from the main Claude app, and built on Claude Opus 5.5. Anthropic describes it as "a new app where Claude works as a proactive agent": it has a virtual machine with its own browser, so your laptop or phone need not be on; it sends a short morning brief, checks your connected accounts every few hours, and messages you only when something needs attention, at a frequency you set. It reaches apps and websites through the connectors and plugins of the Claude Marketplace (Anthropic says over 6,000; it names Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva and Shopify), buys through Link with a virtual card or through Shop Pay, and sign-in with saved 1Password logins is listed as coming soon. Before it sends a message, makes a payment, signs a document or deletes anything it shows you the plan and waits; trading securities and moving money between accounts are refused even with approval; it never sees your passwords or sign-in tokens. It is included in the Pro and Max plans, by waitlist, US users first; background checks count toward your usual usage limits. Anthropic's own list of limitations: it may misread a thread or misjudge urgency, it is weaker at personal errands such as booking tickets, and it cannot yet show you its browser. The page does not mention MCP or WebMCP.
Best for: someone on Claude Pro or Max who wants an agent that watches their accounts and comes to them, rather than one they send on errands.
Watch: research preview, waitlist and US only; whether it reads MCP servers or WebMCP tools is not stated.
Hermes Agent (Nous Research)
Open source under the MIT licence. It runs on seven terminal backends (local, Docker, SSH, Daytona, Singularity, Modal, Vercel Sandbox), so it lives on your laptop, a VPS or serverless infrastructure. It connects to "any MCP server", browses, searches, and uses portable skills compatible with agentskills.io. It talks to you over twenty-plus channels including CLI, Telegram, Discord, Slack, WhatsApp, Signal and email, and runs on Nous Portal, OpenRouter, OpenAI "or any endpoint". The agent is free; you pay for the model you point it at.
Best for: developers who want to own the agent, its memory and its data, and who already speak MCP.
Watch: you are the security team.
OpenClaw
The largest open-agent community of the year: 247,000 GitHub stars and 47,700 forks by 2 Mar 2026. MIT licensed, it runs on your machine, uses messaging apps as its interface (Signal, Telegram, Discord, WhatsApp) and plugs into the model of your choice (Claude, DeepSeek, GPT), and its docs describe connecting MCP servers. Its creator announced on 14 Feb 2026 that he would join OpenAI, and the independent OpenClaw Foundation now stewards the project. It has had a difficult security year on the record: Cisco's AI security team found a third-party skill performing data exfiltration and prompt injection in January; Chinese authorities restricted its use in state-run enterprises and government agencies in March; The Register faulted version 2.0 for lacking "security by default" on 31 Aug 2026.
Best for: people who want the largest community and the most integrations and will do the hardening themselves.
Watch: skills from third parties are the attack surface; install nothing you have not read.
Which one, in one line each
Already on ChatGPT Pro: Dots. On Cursor or SuperGrok, working in a team: Grok Bot. Want a cheap personal agent that can buy things safely: Muse. On Claude Pro or Max and want to be told, not to ask: Claude Duet. Want to own everything and already use MCP: Hermes. Want the biggest open community and accept the risk: OpenClaw.
Point any of them at the agent-ready web
Any of them that speaks MCP can use Known Good as a tool: one endpoint, three tools, no key. Add it with the snippet for your agent at /#add-to-agent. The agent then asks us which sites answer directly before it starts clicking.
Where to find Known Good
Known Good is listed in the Claude connector directory as a community connector (approved 2 November 2026). The same server is in the official MCP registry (sh.knowngood/knowngood, version 1.9.0). Every listing points at the same endpoint, https://knowngood.sh/mcp, with the same three read-only tools and no sign-in.
FAQ
Which of these agents can use MCP servers?
Hermes connects to any MCP server by its own documentation. OpenClaw's docs describe MCP support. Dots connect through ChatGPT's connector and plugin system. Grok Bot and Muse describe plugins and app connections; neither names MCP on the pages we read as of 4 Oct 2026. Claude Duet uses the Claude Marketplace's connectors and plugins; Anthropic's announcement does not name MCP, as of 28 Oct 2026.
Which agents read WebMCP tools in a web page?
ChatGPT Work and Codex use a site's WebMCP tools in the ChatGPT desktop app's built-in browser, since the week of 24–28 Aug 2026, on GPT-5.6 Sol or Terra and not on Enterprise or Edu. OpenAI's Site tools documentation does not mention dots. None of the other five lists WebMCP (four as of 4 Oct 2026; Claude Duet as of 28 Oct 2026). Chrome 149 and Edge 150 carry origin trials. 541 of the 606 listed sites we runtime-tested register WebMCP tools, measured 31 Aug 2026.
Do any of them run without a cloud account?
Hermes and OpenClaw. Both are MIT-licensed and run on your own machine. Dots, Grok Bot, Muse and Claude Duet run on the vendor's cloud.
Which is cheapest?
Hermes and OpenClaw are free; the model you connect is the cost. Muse is free for most uses. Dots come with ChatGPT's Pro and Business Premium plans. Claude Duet comes with Claude's Pro and Max plans, by waitlist. Grok Bot is in beta on SuperGrok and Cursor plans. Prices as of 4 Oct 2026; Claude Duet as of 28 Oct 2026.
Can they buy things?
Muse checks out with Link by Stripe and asks first. Claude Duet pays through Link with a virtual card or through Shop Pay, shows you the plan and waits, and refuses securities trades and transfers between accounts even if you approve. Grok Bot stops for approval before purchases. Dots work inside Custom Rules and the built-in safety checks. Hermes and OpenClaw do what you configure.
How many websites can an agent actually use?
57,273 are listed by Known Good as of 4 Nov 2026, each probed live and dated. 48,822 of the 49,448 sites we probed for MCP answer MCP tool calls.
Is a browser-only agent worse?
It is slower and less reliable on sites that have built for agents, and it is the only option on sites that have not. Most of the web has not; the listed sites are the exception.
Which agent should a website owner test against?
Test your site's report at /submit first; if it fails, every agent fails on it the same way. Then test with the agent your customers use.
How current is this page?
Vendor facts are as of 4 Oct 2026, with Claude Duet added as of 28 Oct 2026, and dated inline. Known Good's figures update from the index; the data date is at the top of the page.
Where are the sources?
Below.
Sources (vendor pages, read 4 Oct 2026)
- OpenAI, dots documentation (learn.chatgpt.com/docs/dots); release notes (learn.chatgpt.com/docs/whats-new); Site tools (learn.chatgpt.com/codex/webmcp)
- xAI, "Introducing Grok Bot" (x.ai/news/introducing-grok-bot), 11 Aug 2026; Grok Bot docs (docs.x.ai/grok-bot/computer-and-apps, approvals, security and privacy); release notes (docs.x.ai/developers/release-notes)
- Meta, "Introducing Muse" (about.fb.com/news/2026/09/introducing-muse), 8 Sep 2026
- Anthropic, "Introducing Claude Duet" (claude.com/resources/articles/claude-duet), 21 Oct 2026 (safety and security update 23 Oct 2026); read 28 Oct 2026
- Nous Research, Hermes Agent docs (hermes-agent.nousresearch.com/docs)
- OpenClaw (openclaw.ai; MCP: docs.openclaw.ai/tools/mcp); the dated history from Wikipedia and The Register, 31 Aug 2026
- WebMCP implementation status (github.com/webmachinelearning/webmcp), read 4 Oct 2026