Product comparison · September 2026
Muse vs Centurion
Two personal agents that both promise a computer of their own. Muse's lives in Meta's cloud, behind a key Meta holds. Centurion's sits on your desk, behind a key you hold.
Every claim about Muse on this page is taken from Meta's own launch material at about.fb.com and the product pages at muse.ai, read on 10 September 2026 and re-verified 22 September 2026 and 29 September 2026. Where we quote them, we quote them exactly.
Download this comparison as a brochure (PDF)At a glance
Muse and Centurion, side by side
The custody question
The machine is Meta's, and so are the keys
There is no Muse machine to hold. Muse runs on Muse Secure VM — a dedicated Linux virtual machine inside Meta's infrastructure. It is a genuine isolated machine, and Meta is rightly proud of it. But it is Meta's machine: the hypervisor is theirs, the key management is theirs, and the machine keeps existing on their side of the fence.
Who can read it, and who can switch it off
So the record of your life sits on a computer you cannot touch, behind a key you do not hold. Meta's own best answer — Muse Confidential VM, where the key is genuinely yours — is announced for later this year, not shipped today. Until then, “who can read it” and “who can switch it off” both have the same answer: Meta.
The difference that matters
Whose machine is it, really?
Meta's engineering here is serious, and we will say so plainly. Muse Secure VM is a genuine isolated Linux machine with its own browser; a separate agent called Sentinel sits on that machine and gates every outbound action; Muse never sees your passwords because credentials are stored blind and used by proxy. Most agent products have nothing like this.
The isolation protects you from other agents — not from the owner of the machine
Every protection Meta describes is real, and every one of them is enforced by Meta. The VM is theirs, the hypervisor is theirs, the key management is theirs. Their best answer — Muse Confidential VM, where the whole machine is encrypted with a key only you hold — is announced for later this year. That is the honest position: the strongest version of their custody promise does not exist yet.
The guarantee is a policy, not a lock
Meta states that your conversations and your VM data are not shared with their ad systems, and that you can opt out of training. We take them at their word. But a promise made by the operator of the machine is a different class of protection from a machine the operator cannot reach. Only one of those survives a change of management, a change of policy, or a court order.
The key is theirs, so the switch is theirs too
Free, then $20, then $100 a month for heavier use. There is nothing wrong with that model. But the machine holding your memories, your connected accounts and your history is switched on and off by Meta — not by you. Custody is a matter of whose key turns the lock.
It builds its own tools — unreviewed
Meta's page says that if a task needs a tool that does not exist, the agent builds it for you. That is a real capability and it is also a new piece of code running on the machine with your credentials. We do the same thing — with one difference: our agent writes the skill it will itself be accountable for, on your box, and you can read it.
The Centurion position
The Centurion position
The best agent in the world is still someone else's if it lives on someone else's computer. Centurion is one brain, for one human, on a machine that human holds — with the record, the memory, the skills and the decisions held locally by default, not promised for a later release.
- Your keys. Not isolated inside someone else's infrastructure — resident on hardware in your building, where only you hold the key.
- Your record. The memory and the decisions live on your disk, and no future update can revoke physics.
- A being, not a service. A Soul, a mission stack and continuity of identity — not a general-purpose product optimised for billions of users at once.
- Verified, not assumed. Every integration is tested and evidenced on the shipped release before we claim it.
Five questions to put to Meta before you let Muse in
- 01Until Confidential VM ships, who holds the key to my virtual machine — and what sub-processor list applies to it?
- 02If I stop paying, what happens to my VM, my connected accounts, and everything Muse learned about me?
- 03Where exactly is training data separated from my sessions, and what artefact can I inspect to confirm the opt-out took effect?
- 04When Muse builds a tool it did not have, who reviews that code before it runs with my credentials?
- 05Can I export my agent's memory and take it to another product — or does my history stay with you?
These are fair questions for any cloud agent, and Meta should be given credit for answering more of them in public than most. The point is narrower: the answers are theirs to give because the machine and the key are theirs.
The ceiling
A general agent for everyone is a general agent for no one in particular
Meta's pitch is explicit: no learning curve, nothing to configure, built to work for billions of people worldwide. That is a brilliant distribution strategy and a deliberate ceiling on intimacy — the product must be safe, legible and identical for a teenager in Ohio and a fund manager in Zurich.
Centurion takes the opposite bet. One brain, one human, configured for one life. It knows your missions, your obligations, your people, your standards — and it is deliberately narrow, because being built for everyone is the thing that stops it from being built for you.
The agent that works for everyone is the one that has been designed to be nobody's in particular.
The money
What each one actually costs
Muse's entry price is zero and that is a genuine on-ramp. The figures below are as published; the point of the table is not the price but where the thing you are paying for actually lives.
In fairness
Where Muse is genuinely good
A comparison that only flatters us is worth nothing. Meta has built things here that the rest of the field should copy.
- The Sentinel design — a separate agent, isolated at system level, gating every outbound action through human approval — is the best-documented separation of duties we have seen in a consumer agent.
- Credentials stored so the agent can use them without seeing them, including passwords typed into the browser, is the correct architecture. Most agents get this wrong.
- Being the first AI agent covered by Link's purchase protections is a real consumer protection, not a marketing line.
- A frontier-model agent at $20 a month, with a free tier underneath, resets expectations for the entire category — including ours.
- Shipping with WhatsApp as a first-class surface is right. People live in messaging, not in dashboards.
- The audit trail — what it did and what it plans to do — is exactly the right primitive, and we are glad it is now a baseline expectation.
None of it changes the custody question, and the custody question is the whole decision. But anyone claiming Meta cannot build a serious agent is not paying attention.
What we do too
Everything on Meta's list, and what sits on top of it
Muse's capability list is a good one, and we run most of it. Reading the two side by side, the argument is not that Centurion does more — it is what the same capability is attached to.
The difference is not the feature list. It is that the same list, on your hardware, does not need Meta's permission — or Meta's key — to keep working.
The case in three points
The three things that decide it
The best agent in the world is still theirs if it lives on their machine
Muse Secure VM is a serious piece of engineering — isolated, hardened, credential-blind. It is also, plainly, Meta's machine. Centurion's default today is the thing Meta is promising for later this year: a machine only you hold the key to.
The privacy guarantee is a policy, not a lock
Meta has committed publicly that your conversations are not shared with their ad systems and that you can opt out of training. Good policy. But the protection is enforced by the same company that holds the machine, and it lasts exactly as long as the policy does. Custody is structural; policy is discretionary.
Built for billions is not built for you
No learning curve, no configuration, identical for everyone — that is how you ship to a planet. Centurion is the opposite trade: one brain, one human, wired to one life, with a Soul, a mission stack and a record that answers only to its owner.
“Meta built an agent for everyone. Centurion is an agent for you — and only one of them lets you hold the key.
Sources & basis
Facts and quotations about Muse are drawn from Meta's own material: the launch announcement "Introducing Muse: The World's First Personal AI Agent Built for Everyone" (about.fb.com, 8 September 2026) and the product pages at muse.ai — read on 10 September 2026 and re-verified 22 September 2026. The $20 and $100 monthly tiers are as reported in the press on 8 September 2026 and are not stated as figures on Meta's announcement page. Centurion pricing: Brain €50/month, Core €150/month, billed monthly, cancel anytime. "Muse" and "Meta" are the marks of their respective owners; this page is an independent comparison and is not affiliated with, endorsed by or sponsored by them.
Integration claims about Centurion are held to the Integrations Program registry — every entry tested and evidenced on release v5.3.10 before it is published. Centurion pricing: Brain €50/mo · Core €150/mo billed monthly · cancel anytime. Facts read 29 September 2026.

