OpenAI dots vs Meta Muse vs OpenClaw: a lab's agent or yours

Side by side, as of October 2, 2026

Both labs’ agents are new (dots launched at OpenAI’s DevDay on September 29, Muse on September 8) and both are rolling out gradually, so expect the rows on price and reach to move. The explainers on what OpenAI’s dots are and what Meta’s Muse is carry the detail this comparison leaves out.

What you need to be paying for already

A dot is not sold on its own. It comes with ChatGPT Pro, which since DevDay is three plans at $100, $200 and $500 a month, or with a ChatGPT Business Premium seat at $100 a user a month billed yearly ($125 billed monthly); Enterprise, Edu and Healthcare workspaces get a beta their admin switches on. OpenAI’s announcement says the first dot is included “at no extra cost” and talking to it doesn’t count against your ChatGPT limits, while work it starts in Codex or ChatGPT Work draws on those as usual. For the first month after launch, OpenAI’s release notes say something different: dots usage “won’t count” toward Pro, Business and Enterprise plan allowances, with each plan’s usage terms to follow. More dots, and paying for a faster or busier one, are described as coming later.

Muse starts free, with a usage limit that refreshes. Meta’s subscription page lists a $20 Power plan with 500 million Muse tokens a week and a $100 plan with 3 billion. OpenClaw costs nothing to install. It needs a computer that stays on and a model to think with: a spare machine and a pay-as-you-go key, a server rented for a few dollars a month, or the ChatGPT plan you already pay for, which OpenClaw can sign in with. The full cost breakdown itemizes each route.

Where you can get one

On Pro, dots are for adults outside the European Economic Area, the UK and Switzerland. Business Premium is rolling out worldwide, so in those countries a dot currently comes through a company’s ChatGPT workspace. Muse is available to people 18 and over in the US and Canada, and Meta says it has no further timing to share for other countries. OpenClaw runs on any computer you can install it on, in any country. A host has its own terms: Everpod’s exclude countries under US sanctions or export controls.

Which model, and whose apps

A dot thinks with OpenAI’s GPT-6 Astra and Muse with Meta’s Muse Spark. Neither lab’s pages describe a way to run its agent on another company’s model. OpenClaw’s provider directory runs from OpenAI, Anthropic and Google to DeepSeek, Mistral and open-weight models on your own hardware through Ollama or LM Studio, and the model can change per task or the day a better one ships.

The apps split the same way. You reach a dot in ChatGPT on desktop and the web, on mobile once it’s set up on desktop, by voice call, or in Slack and Teams; texting is announced, and OpenAI’s help center describes a limited beta for some Pro users in the US. Muse lives in its own app on iPhone, Android and Mac, at muse.ai, and inside WhatsApp, and Meta has announced AI glasses and an email address of Muse’s own. OpenClaw joins the messaging apps you already have open: Telegram, WhatsApp, Signal, iMessage, Discord, Slack, Teams and more than twenty others, so the agent is a contact rather than a new app.

What it can reach, and who decides

A dot works on its own cloud computer, in the apps you connect through ChatGPT’s plugins, and, if you allow it, on one personal computer of yours, which has to be online with the ChatGPT app open. Muse has a terminal and a browser on its own machine and the connectors you pick, and on a Mac you can give it Full Disk Access and let it read, or also act in, apps such as iMessage, Notes and Calendar. OpenClaw reaches whatever its computer and the accounts you hand it can reach.

Both labs build a floor you can’t lower. A dot starts with OpenAI’s rules for when to act and when to ask. Custom Rules can let it act without asking, make it ask first, or hand a step back to you, but they “cannot turn off core safety requirements”, and changing a password or moving money always stays with you. OpenAI’s monitoring “can pause or stop the dot’s work.” Muse sends everything that reaches the outside world, connector actions and network requests alike, through a separate Sentinel agent that the main agent cannot override, and keeps passwords and card numbers out of the model’s sight.

These are careful designs, and they are the vendor’s. On a computer of your own, OpenClaw’s approvals, sandboxing and tool permissions are settings you choose, which also makes getting them right your job; the model provider’s terms are the outside rule that still applies.

Who can see what it does

OpenAI’s dots privacy FAQ says human review “may occur in limited circumstances, including safety-related cases, even when model improvement is turned off”; on personal plans a setting decides whether a dot’s conversations and work train OpenAI’s models. Meta trains on Muse interactions by default, with a switch to opt out, and its security write-up says the current design “does not prevent Meta from accessing data when necessary to support, secure or operate the service”; an encrypted Muse Confidential VM is promised for later this year. With OpenClaw the question becomes who runs the hardware and which model provider sees your prompts, which the cloud-or-local question works through.

What happens to its memory if you leave

This is where the three part furthest. OpenAI’s help center says you “cannot view, delete or directly modify individual dot memories”: the way to clear them is to delete the dot, and OpenAI’s dots help pages describe no way to export them, though depending on where you live a privacy request can get you a copy of your personal data. Files, Codex threads and ChatGPT chats a dot created are kept separately and survive its deletion, and a dot shares memory with ChatGPT in both directions. Muse keeps its memory, personality and identity in files named Memory.md, Soul.md and Identity.md that you can edit, and a “Download your Muse data” control gives you a copy of chats, files and other agent information.

Those file names will be familiar to anyone who has run OpenClaw, whose workspace holds MEMORY.md, SOUL.md and IDENTITY.md. In OpenClaw those files are the agent: plain text on its own computer that you can copy, keep in a private git repository, and carry to another machine or another model without losing who it is. Keeping an agent’s personality and memory in files you own covers how.

What owning one asks of you

The open agent’s costs are the parts the labs do for you. Someone has to install OpenClaw, keep the machine patched and its control panel off the open internet, update OpenClaw as the project moves, back up the workspace, and pay for a model. On a laptop that sleeps, the agent sleeps with it; on a server you rent, you are its administrator. For people who enjoy that work, nothing is more flexible.

A host is the way to hand those parts over and keep the open agent, which is what open-source AI agent hosting means. Everpod is one such host: it runs OpenClaw on a private computer for each agent, never shared with other customers, set up and secured before you meet it, backed up every day and kept updated, and unreachable from outside unless you switch on a public address. It costs $29 a month and includes up to $10 of model usage a month, with the current models from the big labs and the open-source leaders behind it, and your own key welcome with nothing added on top. What you say to the agent and what it remembers live on its computer, not in Everpod’s systems, and if you leave, a full copy of the pod’s data is yours on request before the pod is deleted.

Grok Bot, the one before them

xAI’s Grok Bot, in beta since August 2026, has the same shape with one difference worth knowing: every bot on an account shares one cloud computer, and its FAQ says, “Do not use separate Bots as a security boundary.” It comes with paid Cursor plans or a linked SuperGrok subscription and lives in its own apps on desktop and phone. Grok Bot against OpenClaw and the three-way comparison with Hermes cover it in full.

Picking one

If you already pay for ChatGPT Pro, spend your day in ChatGPT and Slack, and want an agent working this afternoon, a dot is included and well made, and OpenAI has put real thought into what it may do without asking. If you are in the US or Canada and want to try an agent for nothing, Muse is the easiest start, and its downloadable memory files make it the easier of the two to leave. Both are polished products, and for many people one of them is the right answer.

Choose OpenClaw when some part of the arrangement has to be yours: the model, whether an open-weight one or whichever leads this month; the apps, if your life runs through Telegram, WhatsApp or Signal; the rules; a country the labs don’t serve yet; or memory you can carry to the next agent. Then choose where it runs, between a Mac mini, a server and a managed pod, since that decides whether you also become its administrator.