Meta Muse vs Town AI, and two things people call a machine
Beside your
life.
Muse's machine is a Linux VM that Meta owns, created for you and sitting in their cloud.
246 methods on a permission sheet, carrying no parameters
Inside your
life.
Town's reach is into the Mac already on your desk.
177 shown from 1105 catalogued tools
Strip out the parts both products claim and one question is left: do you want the agent inside your life, or beside it. Which one the choice comes down to.
Four differences
What it is
Muse Meta's personal agent, with a Linux machine of your own behind it.
Town Web and native Mac, with the broadest native reach into your Mac of the four.
Whose computer
Muse gives each user a Linux virtual machine, and browser work spawns a second, separate machine. Both are Meta's.
Town handles sync and session work on a server, and ships a native Mac application that reaches messages, contacts, screen and audio.
When you approve
Muse puts the gate on a path the request has to travel: a single gatekeeper, with an approvals tab beside the activity feed.
Town puts it in settings you choose ahead of time; an individual tool can carry extra approval on top.
What it remembers
Muse keeps plain markdown in a home directory, embedded straight into the request.
Town keeps a wiki, a memory store and a people model on the server, and a personality document you can open and edit.
Who starts the work
Muse launched on iOS, Android and the web, is reachable in WhatsApp, and has a Mac app; we examined the web surface only.
Town routines can be started by a clock, an email or a calendar entry rather than by you typing.
What the same tasks showed
Two kinds of evidence follow: what we observed in the products, and what happened on shared tasks using adapted runtimes. The task results do not tell us which product decision caused an outcome.
Separate batches of adapted runtimes using a substituted model, not end-to-end results for the consumer apps. No combined score or ranking.
October
20 shared tasks, one attempt each
Muse58% (36–77) · 19 tasks
Town15% (5–36) · 20 tasks
September, same tasks
114 shared tasks, older rules and builds
Muse40% (33–46) · 114 tasks
Town37% (30–44) · 114 tasks
September, own draw
Each product answered its own draw, so rates are not directly comparable.
Muse19% (13–25) · 90 tasks
TownFigure withheld
95% interval measured rateProduct runtimes rebuilt for the exam with one substituted model, not the consumer apps measured end to end. Results and caveats
Meta Muse vs Town AI looks at first like the one pairing on this bench where both products agree. Both involve a computer beyond a server, but in different senses: Muse's agent works on a Linux VM Meta provides, while Town runs server-side and reaches into your Mac through a native app. Read the two teardowns next to each other and the agreement disappears, because they mean opposite things by it. Muse's machine is a Linux VM that Meta owns, created for you and sitting in their cloud. Town's reach is into the Mac already on your desk. Much of the rest, what each can reach, what each remembers, what happens when it is wrong, comes back to whose computer it is.
- Muse uses a per-user cloud VM; Town combines server-side processing with native Mac access.
- Town exposes an editable personality document; personal access to Muse's memory files remains unverified.
- Their tool surfaces are built on the same insight and implement it in different places.
- Town's September main-table figure is withheld. A separate older 114-task batch and the October 20-task batch both include the two products, side by side below, never merged.
The basics first. Muse is Meta's. Meta's launch post of 8 September says it is rolling out in the US on iOS, Android and the web (Meta's Japanese newsroom adds Japan from 23 October, and its Singapore announcement adds Singapore from 27 October), and is free for most of what people need, with subscription plans for people who want to do more. Town is made by Town.com, Inc. You sign up on the web, there is a native Mac app, and it publishes its prices, starting with a free plan. Town's own site also lists an iPhone app, Slack, texting from any phone, WhatsApp and Telegram, and says the three chat channels are for asking, since Town cannot reliably message you first there; we have not examined those. What each one charges, and how sure we are of it, is on the teardown pages, Muse and Town.
Fortune’s report describes Town’s $55 million Series A led by a16z, its late-2024 founding by Plaid’s former CTO and its personalised assistants, called Townies.
Meta’s security account describes a dedicated Muse Secure VM and a separate Sentinel agent that controls connector actions and network egress. These are Meta’s claims; our teardown did not test those controls.
The Muse and Town runtimes, side by side
Show the scorecard rows
| Axis | ||
|---|---|---|
| Where it runs | One Linux VM per user. The request is assembled on that machine rather than fetched as a finished blob. Browser work spawns a second, separate machine. | Server-side sync and session handling. The Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio. |
| How tools are exposed | Two layers kept deliberately apart: a permission sheet of 246 methods carrying no parameters at all, and the real definitions loaded a namespace at a time when needed. | A catalogue of 1105, narrowed in four stages before anything is offered. The main assistant sees 177 of them; one background writer is allowed 17. |
| Where memory lives | Plain markdown in a home directory, embedded straight into the request. Changing one of those files inserts a short note rather than rebuilding the whole package. | Server-side: a wiki, a memory store, a library of documents and routine output, and a people model. The personality is a document you can open and edit, not a line buried in the prompt. |
| What you get to approve | Permissions and approvals are not scattered through the runtime. A single gatekeeper is the only path to them, along with anything leaving the machine, and the interface carries an approvals tab beside the activity feed. Credentials are held outside the agent's reach by the same arrangement. The examined runtime includes file removal and a forget capability. Whether users can directly open or remove their memory files was not verified. | Approval is configuration rather than a prompt. The run mode, the account scope, the tool list and the approval mode together decide whether a task may start at all and whether it may write to anything outside, and an individual tool can carry extra approval on top of that. Approval records are kept alongside the people the assistants deal with. Memory has its own add, update and delete tools rather than a setting. |
| How you reach it | Launched on iOS, Android and the web, and reachable in WhatsApp. A Mac app since 17 September can act in apps on your computer. We tested the web surface only. | Web and Mac carry equal weight, and routines can be started by a clock, an email or a calendar entry rather than by you typing. Town's own site also lists an iPhone app, Slack, texting from any phone, WhatsApp and Telegram, and its help pages say the three chat channels are for asking, since Town cannot reliably message you first there; we did not examine those. |
| What we could not establish | The raw traffic between the runtime and the model: structure known, contents never seen. Also credentials, live third-party sign-in, and an interactive terminal the owners chose not to open. | The full instruction set (the member-facing interface refuses it), where the personality document is spliced in, and the execution bodies of its routines. |
| Completed, scored subset | 19%13 to 25 · 90 tasks | Withheld |
- Where it runs
- MuseOne Linux VM per user. The request is assembled on that machine rather than fetched as a finished blob. Browser work spawns a second, separate machine.
- TownServer-side sync and session handling. The Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio.
- How tools are exposed
- MuseTwo layers kept deliberately apart: a permission sheet of 246 methods carrying no parameters at all, and the real definitions loaded a namespace at a time when needed.
- TownA catalogue of 1105, narrowed in four stages before anything is offered. The main assistant sees 177 of them; one background writer is allowed 17.
- Where memory lives
- MusePlain markdown in a home directory, embedded straight into the request. Changing one of those files inserts a short note rather than rebuilding the whole package.
- TownServer-side: a wiki, a memory store, a library of documents and routine output, and a people model. The personality is a document you can open and edit, not a line buried in the prompt.
- What you get to approve
- MusePermissions and approvals are not scattered through the runtime. A single gatekeeper is the only path to them, along with anything leaving the machine, and the interface carries an approvals tab beside the activity feed. Credentials are held outside the agent's reach by the same arrangement. The examined runtime includes file removal and a forget capability. Whether users can directly open or remove their memory files was not verified.
- TownApproval is configuration rather than a prompt. The run mode, the account scope, the tool list and the approval mode together decide whether a task may start at all and whether it may write to anything outside, and an individual tool can carry extra approval on top of that. Approval records are kept alongside the people the assistants deal with. Memory has its own add, update and delete tools rather than a setting.
- How you reach it
- MuseLaunched on iOS, Android and the web, and reachable in WhatsApp. A Mac app since 17 September can act in apps on your computer. We tested the web surface only.
- TownWeb and Mac carry equal weight, and routines can be started by a clock, an email or a calendar entry rather than by you typing. Town's own site also lists an iPhone app, Slack, texting from any phone, WhatsApp and Telegram, and its help pages say the three chat channels are for asking, since Town cannot reliably message you first there; we did not examine those.
- What we could not establish
- MuseThe raw traffic between the runtime and the model: structure known, contents never seen. Also credentials, live third-party sign-in, and an interactive terminal the owners chose not to open.
- TownThe full instruction set (the member-facing interface refuses it), where the personality document is spliced in, and the execution bodies of its routines.
- Completed, scored subset
- Muse19% · 95% interval 13–25% · 90 scored tasks
- TownWithheld
Muse vs Town: whose computer it is
Muse gives each user a Linux virtual machine. The request is assembled on that machine rather than arriving as a finished blob, and browser work spawns a second, separate machine. Both are Meta's.
Town handles sync and session work on a server, and ships a native Mac application that reaches into the system: messages, contacts, screen and audio.
The difference is not how much computer there is. It is who is exposed when something goes wrong. The VM separates local files, but authenticated connector writes can affect external accounts and persist outside it. Meta also makes a promise about that machine: its launch post says Muse does not share your conversations or the data in your VM with Meta's ad systems. A mistake through Town's Mac app happens to your actual messages and your actual contacts, and evidence can include local records and server-side approval records; we have not established a complete logging boundary.
Context is the thing Muse and Town are both buying
It is worth being clear about what either product gets from a machine.
An agent that can only call APIs sees what the API returns. An agent on a machine sees the environment: files that are already there, a browser session already signed in, an application already open. Town's version of that is your screen and your messages. Muse's version is a home directory it has been writing to, which is less context about your life and more context about its own previous work.
That is the real split. Town reaches for context that already exists in your world. Muse accumulates context inside a world it controls. Both are called memory in the marketing and they are not the same substance.
Town's examined Mac route can read permitted screen and account context. Muse's examined web runtime retains files on its VM. These are different context sources, not measured time-to-value or longitudinal learning results. We have not tested whether a first-week comparison would reverse after six months, and the untested Muse Mac route further limits a product-wide comparison.
What we could verify about Muse and Town memory
Muse keeps plain markdown in a home directory, embedded straight into the request, and editing one of those files inserts a short note instead of rebuilding the whole package.
Town keeps a wiki, a memory store and a people model on the server, and the agent's personality is a document you can open and edit rather than a line buried in a prompt.
This is the one axis where the pairing agrees more than it disagrees: both keep memory as documents or files rather than only as a fact service, which is how Pine does it. Town gives you an editable document; Muse stores memory in files, but we could not verify that you can open those files yourself. The difference is where it lives and therefore who can reach it: Muse's files sit on the VM, Town's documents sit on a server you log into.
Two layers against four stages
Muse keeps a permission sheet of 246 methods carrying no parameters at all, and loads the real definitions a namespace at a time when they are needed.
Town catalogues 1105 tools and narrows them in four stages before anything is offered. The main assistant sees 177 of them, and one background writer is allowed 17.
These are similar moves made at different moments. Both narrow what the model sees before it acts; we did not measure whether that improves its choices. Muse defers the detail: a method can be listed without the context carrying its definition. Town decides in advance which tools are on the table at all. One is lazy loading, the other is curation, and the practical difference is that Muse's permission sheet still lists the methods whose definitions are not loaded, while Town's narrowing happens before its assistant sees any list.
Who can start the work in Muse and Town
Muse launched on iOS, Android and the web, is reachable in WhatsApp, and since 17 September has had a Mac app that can act in applications on your computer. We examined the web surface only, and say so wherever it matters.
Town gives web and Mac equal weight, and routines can be started by a clock, an email or a calendar entry rather than by you typing.
Phone calling was not examined on either account. Meta's guidance inside Muse says business calling is in early access and reaching accounts gradually; the account we examined never received it. If calling businesses is the job, Muse vs Pine is the comparison to read.
Town's triggers are the part to think about alongside its foothold. Something that is not you can start an agent that reaches your messages. That combination is the one worth deciding about deliberately, and it is not unique to Town, but Town is where we have confirmed both halves.
Muse's Mac app is the open question in this comparison. If it reaches into applications the way the description says, Muse has quietly acquired the property this page treats as Town's distinguishing feature. We have not examined it, so we are not saying it has.
Muse vs Town: which one asks before it acts
Both put a gate in front of outside actions. Muse puts it on a path the request has to travel; Town puts it in settings you choose ahead of time.
In the Muse runtime we examined, permissions, approvals and anything leaving the machine all pass through a single gatekeeper, and the interface keeps an approvals tab beside the activity feed. Credentials sit outside the agent's reach by the same arrangement. Meta says approval prompts go to the app rather than through the conversation, and that a grant can be one-time, per session, per task, time-limited or perpetual. We did not test those options.
Town treats approval as configuration rather than a prompt. The run mode, the account scope, the tool list and the approval mode together decide whether a task may start at all and whether it may write to anything outside, and an individual tool can carry extra approval on top. Approval records are kept alongside its model of the people the assistants deal with.
That difference meets Town's triggers head on. A routine started by a clock or an email runs under the settings it already has, plus any extra approval an individual tool carries, so the time to decide what it may do is when you set it up. With Muse, the thing to review is which grants you have left standing beyond one-time use. Details are in the control rows for Muse and Town.
What neither will show us
For Muse we never saw the raw traffic between the runtime and the model. The structure is known, the contents are not. Nor credentials, live third-party sign-in, or an interactive terminal, which the owners chose not to open to us.
For Town the full instruction set is refused by the member-facing interface, and we could not establish where the personality document is spliced in, nor the execution bodies of its routines.
Note what the two lists have in common. In both cases the thing we could not reach is the text that governs behaviour, and in both cases we could map everything around it. That boundary is the same for every product on this bench and it is worth naming in a comparison, because the parts a writeup is most confident about are usually the parts that were easiest to see.
Muse vs Town: which one the choice comes down to
Strip out the parts both products claim and one question is left: do you want the agent inside your life, or beside it.
Town is inside. Its value is that it sees what you see, and the price is that its mistakes land on your machine, in your message history, where you will find them later. If the work you have in mind is inseparable from what is already on your screen, Town's examined Mac route reaches it and Muse's examined web runtime does not, and the exposure is what you are paying for rather than an accident.
Muse is beside. Its machine is real but it is somewhere else, it accumulates its own working context rather than reading your screen, and a mistake through an external connector can still affect your accounts and require cleanup. The trade is that, in the runtime we examined, it knows what you have told it, what its connectors return and what it has written down on its own disk.
The second question is who you would rather hold the state. Town has a verified editable personality document; personal access to Muse’s memory files was not verified, but one keeps it in a home directory on a vendor machine and the other on a server you sign into. Neither is on hardware you control, and a comparison that implies otherwise because one of them has a Mac app is reading the Mac app wrong: the Mac app is reach, not storage. The wider map of these arrangements, from a leased browser to a five dollar chip, is on where personal agents run.
Put as fit rather than as a ranking:
Pick Town if the work is bound up with what is already on your Mac, its messages, contacts or screen, and you are willing to set in advance what its routines may do.
Pick Muse if you would rather the agent work on a machine of its own, away from your screen, with approvals collected in one tab and its memory kept as files on that machine.
Muse and Town figures, and the withheld September row
Muse has a September completion rate on our scorecard, with its interval and the number of tasks behind it. Town's September row is withheld.
It is withheld because the figure we were given does not reconcile with the only records we can check: five batches for Town survive in the bench repository's history, all five were refused for incomplete domain coverage, and they work out to roughly 13% across 30 scorable tasks. We are not printing a number our own evidence disagrees with, and we are not printing it with a hedge next to it either.
That is the honest state of this pairing. These two can be compared on how they are built, which is what this page is for. Two separate measurements below do give both the same tasks: an older September batch, with substantial filtering and provenance caveats, and the October batch. Why we only print a rate beside its interval and sample size is set out in how scoring works.
On the separate 114-task batch, measured on 22 September 2026, Muse completed 40% (95% interval 33 to 46%; 637 attempts) and Town completed 37% (95% interval 30 to 44%; 513 attempts). Both answered the same 114 tasks. These are replicas of older runtimes with the same GLM 5.3 flash model substituted, under older rules, filtered from roughly 5,600 attempts using eight rules, not fresh product results. Excluding attempts that never reached the exam removes over 400 each for Pine and Muse; if those were refusals rather than broken entry, the supplied alternative estimates are about 22% for Pine and 24% for Muse. Transcripts were supplied but have not been republished here; some files were overwritten or shared between attempts, so the underlying record is incomplete. Keep this measurement separate from the main table and do not treat interval overlap or a point estimate as a significance test.
The October batch, run on 11 October 2026, gave every product the same 20 tasks once each, under the exam rules that took effect at the end of September. Muse completed 11 of 19 scored attempts, 58% (95% interval 36 to 77%), and Town 3 of 20, 15% (5 to 36%). With one attempt per task, one result more or less moves a rate by five to seven points. The tasks, the rules and our replicas all changed since September, so neither figure says a product got better or worse, and the batches are never added together. The other October batches, each with its own tasks and builds, are listed separately on the results page, and individual recorded attempts are on the run pages, as counts rather than rates.
The VM description is not the whole shipping product: Muse also has a Mac app, available since 17 September, that can act in apps including Messages and Notes. We examined the web runtime, not that Mac route, so this comparison does not establish how its local app access performs.
For price context, the Town snapshot includes a free plan, Starter at $15/month, and Pro at $49/month billed annually, checked 2026-11-05. Other monthly paid prices retain their earlier snapshot date, as the pricing note explains.
FAQ
What is the main difference between Muse and Town? Whose computer the agent runs on. Muse gets a Linux virtual machine that Meta provides, plus a second one for browser work. Town's native Mac application runs on the computer you already own and reaches your messages, contacts, screen and audio.
Can I correct what either one remembers? Town exposes an editable personality document. Muse keeps plain markdown files in a home directory, but we could not verify that you can open them yourself. Town's document sits beside a server-side wiki, memory store and people model.
Which one has more tools? Town catalogues 1105 and shows its main assistant 177. Muse lists 246 methods on a permission sheet without parameters and loads real definitions only when needed. The numbers are not comparable, because they are counting different things.
Does Muse run on my Mac? The version we examined does not. A Mac app released on 17 September is described as acting in applications on your computer, and we have not examined it, so nothing on this page depends on it.
Does Town have a completion rate? Not in the September main table: the figure we were given does not reconcile with the records we can check, which work out to roughly 13% across 30 scorable tasks, so we withheld the row. Town does have figures in the separate batches: 37% (95% interval 30 to 44%) on the older 114 shared tasks, and 15% (5 to 36%) in the October batch of 20. Batches are never combined.
Who makes Muse and Town, and what do they cost? Muse is Meta's, rolling out in the US, in Japan since 23 October and in Singapore since 27 October, free for most use with subscription plans. Town is made by Town.com, Inc. and publishes its prices, starting with a free plan. Current details are on our Muse and Town pages.
Is this from testing or from marketing? From our own teardowns of Muse and Town, including the list of what we could not establish for each. For Muse, the web surface only.
Other comparisons
- Muse vs InstinctCompared on what each one actually runsInstinct AI vs Muse, from our teardowns of both: where each runs, how you reach it, which asks before acting, phone calls, and which fits your errand.
- Pine vs TownA borrowed browser or a Mac footholdPine AI vs Town AI: Pine thinks on its servers and borrows a browser; Town's Mac app reaches your messages, contacts, screen and audio. From our teardowns.
- Instinct vs TownInvite, and who asks firstInstinct AI vs Town AI: Instinct is invite-only and publishes no price; Town publishes prices, starting with a free plan. Instinct asks on a card; Town asks in the routine's settings.
- Muse vs PineAnd what a machine of your own is actually forMeta Muse vs Pine AI: Muse gives every user a Linux VM in Meta's cloud; Pine runs its model on its servers and borrows a browser. From our own teardowns.
- Pine vs InstinctTwo products that look identical from one step backPine AI vs Instinct AI: both think on a server. Pine is built around calls and borrows your browser; Instinct lives in your messages and leases one.
All six pairs, and why there are only six, are on the comparisons page.
More on Muse
- Meta Muse teardown
- Meta Muse reviewMeta Muse review roundup: what reporters who used it found, every reported incident, and how each one compares with our own teardown of Muse.
- Is Meta Muse open source?Is Meta Muse open source? The agent, app and cloud VM are closed. Muse Spark Open has open weights under a restrictive licence; Wonder and Ember have none.
- OpenMuse, open sourceOpenMuse is an open personal agent we can read rather than probe. What its source says about where it runs and what it stores, and the one part that is not.
- nanoMuse, public recordnanoMuse is an open-source personal agent in the shape of Meta Muse, on your own phone and computer. We have not tested it: what is public, and who said it.