Meta Muse: from personal agent to portable hardware and everyday tasks

Muse deserves a topic because Meta is connecting a model, a cloud computer, shopping and work services, and hardware interfaces such as glasses, Charm, maker devices and Home Link. A person can hand over a task, let the agent progress in the background and return to the result. This topic follows how that chain enters everyday life.

At Connect on September 23, Mark Zuckerberg demonstrated Muse’s voice and avatar interaction, then revealed physical Charm hardware near the end. We begin with the capabilities and device, then connect firsthand experience and everyday scenarios, distinguishing demonstrations, available features and plans.

Meta launched Muse on September 8 as a personal agent powered by Muse Spark. It has its own cloud computer and browser, can continue after you close the app, and returns when it needs approval.

Muse Spark 1.3, released September 2, is a model update available through Muse Code and Meta Model API. Meta emphasizes sustained multi-step work and instruction following, with examples including an edited audio track. A model release does not establish which features every Muse user has received.

• See Spark 1.3 task inputs and outputs ↗Vendor examples, suitable for comparing requests with outputs; not an independent model comparison.

Meta's Wonder release includes a kitchen game with ingredients, chopping, plating and serving across three kitchens. We opened the public artifact and entered its first level, with controls and timed orders. Meta presents it as a 3D-game example built from customizable image inputs. This is a concrete work to explore; we did not rerun its generation or verify every level.

Two models were released October 12. Standard Muse Wonder is available through Meta Model API and Muse Code; Wonder Full is currently offered through a Muse Code subscription. Wonder in the personal Muse app remains forthcoming, starting with subscribers. A model release therefore does not establish that every phone, Charm or Home Link user has switched.

A complete reread of the Artificial Analysis index gives Full at max 56 and standard Wonder at max 52. For reference, the highest measured settings for Opus 5.5 score 58, while Astra and Fable 5.1 score 53. This places two Wonder models on one index without predicting a particular task's outcome. Both have distinct entries in our model directory and board.

• Open the official kitchen-game artifact ↗A public artifact with onscreen controls, without regenerating it through BitShovel.
• Standard Wonder: board and evaluations →
• Wonder Full: board and subscription access →

Active Memory concerns bringing relevant context back across conversations. Active Focus concerns sustained work: trying another approach, adapting to new information, asking when needed and checking before important actions. Alongside portable hardware, these address whether a previous request remains usable and how far work can progress after you leave. They are Meta's capability definitions; device interaction and account-level results still need their own observation.

The model report breaks these claims into tasks. MemoryBench places distractors and later changes in 64K–100K-token conversations, while FocusBench tests blockers and changing conditions in tasks such as housing and travel. Both are internal Meta evaluations. Their scores do not establish error-free memory or universal real-world completion.

Long projects also have external results with consequential settings. The report identifies Full's SWE-Marathon run as xhigh, while Astra and Fable 5.1 use Codex and Claude Code, respectively; that row is not a fully controlled model-only comparison. A concrete next step is to inspect a bounded task with supplied material and revisions: how requirements survive, a failure is handled and a reviewable result returns.

• Read results alongside their evaluation settings ↗
• Continue with an author's shared-task test ↓

On October 12, Piotr Olszewski ran standard Wonder and four rivals once on 23 tasks through OpenRouter at low effort. Tool use, recovery and long-document questions were included; automated scores reached 99–100%.

Wonder totaled $0.64 and 165 seconds; Opus 5.5 totaled $0.75 and 222 seconds. Wonder nevertheless misspelled a Polish amount-to-words output and inserted an unspecified letter date; the latter was not penalized.

Near-ceiling scores do not rank overall capability. Full and Chinese were untested. Low does not equal matched compute; costs and times cover this set. The author reports saved logs but links no download for this run.

• Read the author’s task and error report ↗
• Continue with the tools you already use ↓

Helicon is an unofficial Muse Code interface. On October 12, its author reported that older versions could start threads on Spark 1.3 despite Wonder being selected. His report concerns the Muse Code 1.4.6 session service.

The fix starts a session without a model, then sets it separately and records the requested choice. The author reports checking the model through session/read and turn items. The merged fix shipped in v0.21.9.

Release, menu availability and actual execution are separate integration steps. This is a specific developer report, not a diagnosis of all Muse clients. We read the code without installing or retesting it; a free interface does not remove service fees.

• See how the author checked the model that ran ↗
• Explore this unofficial interface ↗Muse Code account and service conditions still apply.

After roughly 1:04:15 in the keynote, Zuckerberg reveals the keychain-sized Muse Charm. It displays an avatar and is described as bringing real-time voice and avatar interaction to a portable device. A corner fingerprint sensor lets the user start speaking without first unlocking a phone and opening an app. The image above shows the physical reveal.

In a separate demonstration around 28:31, Zuckerberg discusses Project Charm’s materials, translucent styling and component layout with Muse, which responds and continues working. That is a software interaction about the design, distinct from the later hardware reveal; it does not establish that Charm independently completed the design task.

TechRadar’s report, updated September 25, adds a 2-inch OLED screen, front and rear cameras, 5G and Wi-Fi. Reporter Jacob Krol says Meta confirmed the cameras, but he had no hands-on opportunity. Cameras and connectivity could add visual context to a quick request; that is an interpretation of the reported hardware, not a test of recognition, task completion or continuity across devices.

The keynote says only a few devices exist, materials and component layout are still being refined, and shipment is planned for the December holidays. When checked on September 30, Meta’s product page still offers arrival notifications and requires a Muse account. Pricing, battery life, cellular fees and initial setup remain to be clarified, alongside what can be completed away from a phone and how the same task continues when returning to it.

• Watch the physical Charm reveal ↗
• Watch Zuckerberg discuss design with Muse ↗

Muse does not have to wait for Charm to gain a physical entry point. Meta’s new Muse Gadgets code supports ESP32 boards and Linux computers. Makers can start with a small screen, a button or a Raspberry Pi. A microphone-equipped board sends spoken requests to Muse; the current code returns text for the screen or app. Spoken replies require a text-to-speech integration. The person chooses the form and interaction, while the agent receives the information those parts send.

Muse Home Link fills a different gap. The 35 × 42 × 10 mm ESP32-C5 device plugs into USB power and stays on home Wi-Fi after phone pairing, letting Muse reach compatible devices and local interfaces. It provides a network route, rather than a Charm screen or a local large model. The repository’s community printer recipe checks formats, paper and device state before submitting a file, then follows the job. Submission and a printed page still need separate confirmation.

A button board conveys intent; Home Link lets the agent reach devices; the target device and recipe determine the action. Start with something you own: what would you like to ask, and where would you confirm the result? Connecting requires a Muse account, phone pairing and an SDK token. Screen, audio and home-network support vary by board. The source uses Apache 2.0; token access has separate personal, non-commercial terms.

Free Home Link claims are currently limited to active US Muse subscribers, one each. The page says October shipping, first come, first served; claiming reserves a place in line rather than placing an order. The device cannot be reflashed. This account is based on official documentation and code, without flashing a board or testing a household task. Real maker projects and confirmable results are the next things to follow.

• Explore Home Link hardware and setup ↗
• Explore the device code and maker entry points ↗

Muse has another screen. Meta's App Store version 10.0 release notes add iPad support; the U.S. listing checked here dates that version to October 7 and requires iPadOS 18 or later. This is the app version; Muse Spark's model changes remain in the earlier capability section.

The official iPad view places tasks, conversation and material together. Its example prepares a school-trip form, then asks about sending a confirmation email. The larger screen can be understood as a place to review material while directing a task. This is an official illustration, not our account-level test.

Windows is at another stage. Microsoft's October 7 announcement says a native Muse app is coming with MXC integration, which controls file and network access. It still says 'coming soon', without a public download date for the app. Native packaging does not establish fully offline model operation.

Alongside Charm and Home Link, portable devices offer ways to start a task while tablets and computers offer room to inspect material and continue the result. A useful follow-up is how tasks, sources and approvals carry across those entry points. Check each platform's availability separately; this U.S. listing does not establish access in every region.

• Read Muse's iPad release notes ↗
• Continue with the portable Charm entry point ↓
• Revisit reviewing and correcting memory ↓

Charm could shorten the distance between having an intention and handing it over: speak and hear a response immediately. AI glasses offer a hands-free route for visual context. Meta plans Muse on glasses in the coming months, acting on products, posters and lists while service connections expand.

Consider preparing for school: read a notice, organize dates and items, clarify gaps, then set reminders or proceed to purchasing. A phone suits reviewing a long list; glasses could reduce transcription from things in view; Charm suits spoken additions and questions. This is scenario analysis based on the interfaces, not an end-to-end test.

The next question is continuity across interfaces: can something spoken on the way home be found on the phone, can you see what the agent has done and who needs to confirm the next step, and do connected services actually complete the task?

• Watch the official glasses demonstration ↗

Meta’s Muse Connector Platform page shows a concrete example: someone asks Muse to summarize meeting notes, Muse proposes connecting to the note-taking service Granola, and a Connect button appears. Accessing material from another service still involves a connection choice. This is a product illustration, not proof that a summary was delivered or that the same task was completed on Charm.

In its September 24 Connect developer recap, Meta says it has launched the Muse Connector Platform. Developers describe their product and its use, then submit it for functional, security and legal review plus end-to-end testing. Approved connectors can appear in Muse’s directory. Charm could shorten the step of expressing an intention; connectors determine which services the agent can use. Account access, actual results and continuity across devices still need to be observed together.

This is why conversation and task completion are different. For meeting notes, receiving a request on a portable device, reaching an authorized note service and producing a reviewable summary are separate steps. The official page illustrates the connection, not a public end-to-end record of that sequence. Meta AI Connectors, announced at the same event, are a separate developer preview for Meta AI conversations; they do not establish connector support on Muse Charm.

• See Meta’s connector page and original illustration ↗Look at the connection choice and developer review process rather than treating the illustration as a completed task.

These three examples come from the Muse product team: an India itinerary, a school information page and a packing checklist. They turn chat material into pages you can revisit and use. They are official examples; we have not rerun these tasks.

• Explore the team’s design walkthrough ↗Original interfaces, outputs and interaction design.

In a September 14 TechRadar account, the writer asked Muse to find shorts. It asked for a size, returned options and a sale after a few minutes, then requested an Amazon login to proceed. The writer found the delegation useful but felt uneasy about granting account access.

The situation then changed. On September 21, The Register asked Muse to find an office chair and reach checkout. The agent reported an Amazon browser block, with no product found or added. That later trial shows why an early shopping experience does not establish lasting access.

• Read the author’s Muse shopping experience ↗A third-party account describing both convenience and account-access concerns.

Service providers are also deciding how agents can work with them. Meta’s September 24 Connect recap describes Muse’s existing access to Shopify’s catalog and the addition of payment connections such as Shop Pay. This adds another part of the shopping story: an agent’s reach depends on the merchant’s product, checkout and payment access as well as the model.

On September 28, Shopify extended WebMCP to checkout, letting compatible browser agents read and update the current order. Product lead Gil Greenberg illustrates it with an address change: checkout returns revised delivery options and missing information, which the agent can act on next. The potential benefit is less repeated clicking and form-filling. This is an official workflow explanation, not evidence of a Muse purchase.

The final step still needs a clear handoff. Shopify requires the agent to show the current order and total and obtain the buyer’s approval before submitting it; a changed total requires another confirmation. Buyers handle Shop Pay login, payment challenges and required review screens. The tools do not accept new card details. A checkout being ready to pay is different from an order actually being completed.

Muse’s Shopify connection also depends on the entry point. Shopify’s help page, checked October 1, limits direct checkout within Meta to customers in the US, Canada and Mexico, and names Muse where available. If a merchant disables direct checkout, discovery can remain while purchasing moves to the store. This channel checkout and browser WebMCP are separate routes. The official integration is documented; a public record of Muse completing an order through the new WebMCP tools has not been established here.

• See Shopify’s address-change and buyer-handoff example ↗An official product explanation, not a Muse or BitShovel purchase record.
• Check the actual scope of the Meta channel ↗

Finding a carry-on and authorizing its purchase are different requests. Amex's October 13 Intent Intelligence framework connects that distinction to payment: the issuer receives authenticated purchase intent before an agent spends. This continues the Muse shopping question: after discovery, what did the person actually approve?

Approval can cover a specific cart or a conditional future purchase. Budget, merchant, item requirements, expiry and a maximum transaction count define the scope. The agent clarifies gaps, including whether shipping is included; the person selects a card, confirms the intent and authenticates. Dependencies can require a flight purchase before buying insurance. Understanding a request and accepting authority to spend remain separate steps.

An Intent ID identifies the approved record; a Proof of Intent Token represents its main constraints and can obtain a limited payment credential instead of exposing the card number. The issuer may still request final confirmation. Permission, payment and delivery are distinct: the official four-screen illustration shows a proposed flow, while an actual order and fulfillment establish what happened. The original intent can inform a later dispute.

Shopify's current-cart approval and Amex's durable advance intent offer two reference points. As checked October 14, ACE lists some specifications, but API materials require an application and confidentiality agreement; onboarding needs approval and an agreement. Registration and cart-context specifications remain in development. Planned error protection has card, registered-agent, authenticated-intent, return and review conditions. These sources do not establish a Muse–ACE integration or universal protection.

• Read Amex's purchase-intent framework ↗
• Revisit Shopify's buyer approval and order completion →

A bag found in Search and a sweater found in Gemini can now remain in one place. Google began rolling out Universal Cart in the U.S. across those two services on October 15, with saved products, price and restock alerts, and card-perk and promotion suggestions. The official illustration keeps Ulta Beauty and Quince items grouped by store. It addresses shopping progress scattered across tabs and screenshots; saving is still an intermediate step.

Checkout has two routes: Google Pay inside the cart with supported merchants, or transferring items to a merchant’s cart and finishing on its website. The pictured Ulta Beauty button shows the latter. Merchants remain the sellers of record and provide confirmations and support. A shared cart reduces repeated finding and adding; the final items, total and delivery and return terms still need review before purchase.

This adds a reference point to Muse, Shopify and Amex: preserving shopping progress, authorizing spending and fulfilling orders are distinct roles. May’s preview mentioned YouTube, Gmail and broader agent scenarios; this launch confirms U.S. Search and Gemini. It does not establish a Muse integration. As checked October 16, the announcement lists no separate subscription fee or complete eligibility table. Pictured discounts are simulated examples, and BitShovel has not completed a U.S. checkout.

• Read the U.S. launch and checkout routes ↗
• Compare the broader May preview ↗

Everyday tasks often depend on relationships: planning a dinner involves a friend’s dietary needs; handling an invoice may require recognizing a plumber you previously hired. Remembering those connections can reduce repeated questions and connect earlier work to the next step. Meta told WIRED that Muse gathers this context from public information and material users choose to share. People also need to inspect that understanding: did it confuse two names, has a preference changed, and what supports an inference?

WIRED’s October 3 report adds detail: internal instructions extracted by researchers describe a process for organizing relationship pages. An AI-generated public review linked by the article uses a September 24 export, with its page revised through October 3. Files and templates show a design direction; they do not establish that every current user has contact profiles or that the process runs at one shared frequency. BitShovel has not observed those accounts in operation.

Start with what you can inspect. Meta’s current help says memory combines conversations, preferences and patterns observed in interactions, connected-service context and file-based memory. You can ask in chat what it has saved and request a correction, or tap the assistant icon, open Identity, then Memory to edit or delete Memory.md directly. This makes saved understanding visible, but one Memory.md should not be assumed to contain every relationship page and related record.

Forgetting, deleting a message and disconnecting a service affect different parts. Meta says a forget request makes Muse look for relevant memories and supporting files and remove them to the best of its ability. Deleting a source message or file may leave information learned earlier. Disconnecting a Connector stops further data exchange, while existing memories and chat history may remain. For a relationship correction or removal, name the specific subject, ask what remains and inspect visible records; disconnection alone does not establish complete removal.

Use a task you actually do as the reference. For a trip, do you want advice, a shareable itinerary, or continued price checking, communication and booking? Those are different deliverables. Compare what your current tool completes with the particular step Muse might add.

Astra’s 3D examples turn descriptions into editable scenes; the home-inventory project turns item records into searchable household information. Muse adds the question of continuing from information and outputs into action. This connects use cases, rather than ranking performance on a shared test or prescribing a subscription change.

• Explore Astra’s editable creations →From a description to a concrete creation.
• See how a home-inventory assistant organizes belongings →From household information to a searchable result.

Availability changes after launch: the September 8 release specified the US; around 20:51 in the September 23 official keynote, Zuckerberg says Canada has just been added, with more countries to follow. This corrects the topic’s earlier US-only wording. Check your own region and account access on iOS, Android or muse.ai.

If you have access, start with something concrete and verifiable: turn supplied school dates into a one-page checklist and list missing information. Meta says sensitive steps such as sending email or purchasing require approval, and users choose service access. Review the result before delegating further steps.

A patched Mac-client flaw makes service access more concrete. In Patrick Wardle’s September 21 disclosure, software already running as the local user could redirect Muse dictation and abuse access granted to the assistant. The attack requires code execution on the Mac; it is not a way to take over any computer merely because it is online.

On September 22, Meta’s David Singleton said a hotfix had been issued for the Muse Mac app and confirmed the local-code prerequisite. These are the researcher’s demonstration and Meta’s response; BitShovel has not reproduced the attack or checked any user’s installed version. Later coverage on September 25 does not make it a newly discovered or still-unpatched flaw.

This adds context to the hardware story: the entry point may be a phone, Mac or future Charm, while account connections and device permissions also determine what the assistant can do. Before delegating mail or calendar work, update the official client and choose connections for that task. Approving one send or purchase differs from ongoing permission to read information. The Mac finding is not a test result for Charm.

• Open the official Muse entry point ↗Availability depends on your region and account.