Muse Teardown: How the Same Claw Core Can Interact with People in a Different Way. Author |Rhea
Editor | Li Na, Cen Feng
Around September 20, a post was widely shared on X. AI app co-founder Ansh Nanda asked Meta’s newly launched personal agent, Muse, why its system files were almost identical to those of the open-source project OpenClaw. He captioned the screenshot with the commentMuse is LITERALLY OpenClaw for normies. Muse is simply OpenClaw for ordinary people.
The post “Muse is LITERALLY OpenClaw for normies” spread widely on X, and Friedman responded the next day. Leifeng Network(WeChat official account: Leifeng Network)Developers subsequently found SOUL.md and MEMORY.md in Muse’s workspace, bearing the same names as those in OpenClaw.The next day, Nat Friedman, product lead at Meta Superintelligence Labs, responded: Muse’s code had been built from scratch, and the product was deeply inspired by OpenClaw. As for why those files were carried over, he explained that OpenClaw author Peter Steinberger had “gotten these things right.”
Sarah Perez's “Meta admits Muse's likeness to OpenClaw isn't a coincidence”After using Muse for some time, we found that the app's most significant innovations and breakthroughs came from the very things the tech crowd tends to dismiss. Developers can often build many tools by hand themselves, but they overlook one fact:most people won't first learn an entire suite of tools just to gain a single capability.Muse and OpenClaw use the same personal-agent paradigm, and there is actually no generational gap between them in terms of capabilities. The main change lies in the way people interact with agents:OpenClaw makes people learn the machine; Muse makes the machine learn people.“Learning people” has two meanings. One is learning the user: it comes to know that person through their existing social accounts and the memories of previous assistants. The other is learning how people interact with one another, replacing those earlier terminals, configuration files, and logs with an always-present conversation, a character whose appearance you can see, and a response that arrives just in time.In the past, users had to learn all of these things themselves first. Combining our hands-on testing with a static analysis of the publicly available resources on the muse.ai website and the Muse Android APK, let's see how Muse shifts this learning burden away.
muse Web login pageface
01
The same SOUL.md:
What Muse inherited
Muse launched in the United States on September 8, 2026. You can give it tasks via the app, on the web, or through WhatsApp, and it keeps working in the background after the app is closed.Under the hood, Muse and OpenClaw have the same architecture:an agent runtime framework, an internet-capable browser, scheduled tasks, memory stored in Markdown, and a personality file that defines “who it is.”Even the Android version carries over OpenClaw's terminology for phones: the code calls the phone a “node”; the cloud issues commands, and the phone executes them. So all of these parts have been around for a long time; Meta simply has a near-genius product manager who got every one of them right.A product experience that is good enough can attract more people. The familiar story is that, within two weeks of launch, Muse reached No. 1 on the U.S. App Store's Top Free chart.While logged out, we read the pages, scripts, and interface translation table publicly served by the muse.ai website, without calling any authentication endpoints or running the client application. The Simplified Chinese interface copy table contains 6,350 entries, nearly 90% of which have already been translated. In other words, although Muse has not yet launched in Chinese-language markets, the Chinese-language interface is essentially ready.This copy deck is like a dictionary for the product, showing how it intends to speak to users. Later, the author also obtained Muse’s Android installation package (12.0.0.14.169) and, again, performed only static analysis—it was neither installed nor run. Nearly 90% of the Chinese copy in the package had also been translated, though it had not yet been copy-edited. Interestingly, in several places, agent was translated as “customer service.”
The right-hand side shows the beginning of the copy deck, with one interface string per line. The 89% on the left was calculated from those 6,350 strings.This inheritance extends all the way to the ritual of a first encounter. OpenClaw’s BOOTSTRAP.md has the agent awaken in a conversation, work out a name, personality, and signature emoji with the user, and then write them into IDENTITY.md. Muse’s identity file stores the same few things; its project codename is Hatch, meaning “incubation.”
Muse carried over the core persona files from the OpenClaw workspace.The process is the same in both cases: OpenClaw handles it in the terminal, while Muse moves it into a character-customization interface. Meta kept the framework the open-source community had worked out, but almost the entire experience outside that framework was rewritten. We can summarize briefly: the rewritten parts can be broken down into five interaction costs taken off users’ shoulders.- Installation cost
- Choice cost
- Learning cost
- Monitoring cost
- Self-introduction cost
▎Installation cost: A computer users cannot see
OpenClaw’s first hurdle is a computer: you have to supply, configure, and maintain it yourself. Zuckerberg’s assessment was that billions of people would not buy a Mac Studio to set up at home.Muse gives every user a computer, but users do not have to operate and maintain it themselves (heaven knows how much time maintaining OpenClaw takes!). All you have to do is log in with a Meta account, and Facebook, Instagram, and Threads under the same Accounts Center connect automatically. The cloud computer behind the scenes normally stays out of sight, surfacing only when needed as a browser card or a button asking the user to take over.The only thing users need to do is open the muse.ai website or download the Muse app. (Of course, users in China need to do far more than this.)
The cloud computer is normally invisible and appears only when needed, either as a browser card or as a takeover request.▎Choice cost: For the first task, let Muse speak first
The broader the range of capabilities, the harder it is to fill in a blank input box. Nearly every agent product faces this problem: what exactly should the agent do first?Muse’s approach is for it to speak first. The “Ideas” page lists things it can do for the user; click “Let’s do it” and it gets to work. “Feed” contains the short posts and podcasts it writes at its own pace; “Goals” is for long-term work—it asks clarifying questions first, then lays out a plan and provides regular updates.
Muse proactively identifies tasks it can complete for the user based on the information it already hasThe feed also offers an excellent cold-start experience. The first time users see the feed, they are shown a message explaining what it does and what value it offers. When users browse with no user information at all, they see feature descriptions and status messages, but rather than presenting a blank slate, the feed cleverly introduces its features in the form of the posts that will appear later. The feed also supports editing the prompt, meaning users can see how it works and can even directly change the way it operates.
The feed avoids a blank slate during cold start, introduces itself in the form of future posts, and lets users modify how it worksMeta’s first task for users has to do with money. Chief AI Officer Alexandr Wang launched #MuseMoneyChallenge on X, asking Muse to help users save $1,000. Users do not need to figure out what an agent is first; they only need to be able to read a bill.
Meta designed “helping users save money” as Muse’s first taskThrough static reverse engineering of the web page, we found that even before the user had said anything, Muse had already set itself a default goal while creating the virtual machine. The blank page’s tone was also reminiscent of making a friend: “It will gradually take shape as we get to know each other.”
The blank page’s copy frames the beginning of a relationshipTAG300@@▎Learning cost: like sending a message to a friendTAG309
TAG311@@Muse design director Alex Cornell said the team had consistently calibrated the experience around one question:TAG315@@“If the user were talking to another person, how should this work?”TAG319@@The Agent was therefore placed at the top of the screen, like a contact in an address book. There is only one main chat, just as in WeChat, where you have a single window for one person. After a request is sent, Muse first replies with a ?, then goes to work in the background.
TAG323@@Muse confirms receipt before executing in the background, making the interaction much like sending a message to a friendThis trade-off is most evident in the wording. I tallied the keywords in 6,350 Chinese-language interface strings:
TAG330@@Based on 6,350 Chinese-language interface strings. Technical senses such as “target position” have been excluded from “goal”; the statistics reflect interface wording, not usage frequencyAll the developer concepts remain, but they have simply been buried deeper and deeper. MCP appears only six times in total, all in forms where third parties submit connectors for review; the word “plugin” does not appear even once. When the required tool does not yet exist, the website puts it this way:“The agent will create that tool for you”TAG340@@.TAG345@@The exact wording of this line is already baked into the Android installation package. When a user says, “I want to create a space about…,” Muse goes ahead and builds a page. Once it is ready, Muse puts it in the app’s tab bar, and the prompt reads, “Create your own dedicated app.”This March, at the height of the OpenClaw craze, I made one judgment: skill and MCP are developer concepts that ordinary users do not need to understand.Yet OpenClaw and several domestic agent products still built skill stores. This is actually understandable: people are always eager to map familiar concepts onto new things, because product managers at these big tech companies have long mastered the traditional internet concepts of “stores” and “ecosystems.” But new technology calls for entirely new ways of thinking, not a crude application of old ideas.
The author dares to guess that many products imitating Muse will appear next. Someone will certainly keep using the concept of a “store,” whether it is a skill store, an MCP store, or an Agent store. A big ship is hard to turn around, and human thinking works the same way. We tend to find a sense of security in familiar things, which is perfectly understandable. In a contest between masters, you make your moves and I make mine; either way, the market will vote with its feet.
The premise of the App Store is that users themselves are the assemblers. In the agent era, assembly is delegated to the agent, and then the skill store is placed before the user. “Humans learning from machines” thus sneaks back in through the back door, much like a restaurant asking guests to choose their spatulas before ordering.
Muse’s consumer-facing layer does not contain even a single skill. After using it for a while, we discover that we do not actually need to choose a model, or a tier such as “Fast” or “Expert” at all.
Yet skills have not disappeared. In the Android installation package, the code responsible for connectors is called skills and consists of no fewer than 647 files. In the interface, everything is renamed “Connectors,” and the words “skill,” “plugin,” and “MCP” never appear even once. Not a single spatula has been removed; they have all been moved into the kitchen.
Operations have been moved into the kitchen as well. For example, when asking an agent to send a text message from a phone, OpenClaw’s documentation requires the administrator to add sms.send to the gateway configuration’s whitelist first. Muse for Android merely pops up a message on the phone saying, “Muse wants to use your phone to send a text message.” One tap grants permission, and you can also choose “Always Allow.”
Now back to that birth ceremony. Muse’s default character, Jolly, is a cream-colored plush character with rosy cheeks. Users can rename it, change its outfit, and choose its voice.
Muse lets users name and dress the agent, with the configuration process designed as an act of character customization.
During our reverse engineering, we observed that the voice-selection flow was written in Muse’s first person: “Let’s find my voice!” A configuration action was reframed as the character searching for itself.
Choosing a voice was written as the process of Muse searching for its own voice.
When I first started using Muse, it guided me through setting up an avatar. It not only created several selectable styles, but also quietly turned my avatar into an animated anime-style cartoon character in the background. The experience far exceeded my expectations. The result generated in a single pass was already absolutely stunning.
▎Monitoring overhead: Jolly holding a notebook
Most agents use a spinning animation and a scrolling thought process to indicate that they are “processing.” Muse divides this into three layers: ordinary users see expressions, those who want to know more see a task list, and engineers inspect the raw logs. The top layer is entrusted to Jolly, which picks up a small notebook while executing tasks and holds up a glowing orb when generating media.
During our reverse engineering, we also found interface copy containing “Internal Settings” and voice options labeled “Meta Only,” covering engineering parameters such as the speech-recognition backend and call-tool model, as well as debugging tools.
Beneath the three-layer state is another control plane reserved exclusively for the internal team▎Onboarding cost: it knows who you are without getting too familiar
OpenClaw gets to know users through a USER.md template, either filled in by hand or gradually accumulated through conversations. Muse also keeps this USER.md, and the Goals page of the Android app reads the user’s profile from it. During our reverse engineering, we observed Muse prepares a prompt for users who are switching assistants, asking the old assistant to “compile a portable memory snapshot about me,” with each entry being a one-sentence, evidence-based fact and uncertain items labeled “inference.” The new assistant writes a handoff letter to the old assistant on the user’s behalf.
The migration prompt Muse writes for users even spells out what should be left outActually, knowing the boundaries is harder than getting to know the user. The same prompt also states that information such as health status, financial details, and mental health should not be collected unless the user explicitly requests it. When importing old memories, Muse is likewise instructed to “skip anything transient or sensitive.”Those boundaries are also built into its mobile tools. The description of the Android dialer tool makes this clear: the calling tool that speaks on the user’s behalf and talks to the other party is for merchants only. When calling family and friends, Muse only places the call; the user still does the talking. Text messaging works the same way: for mass texts or messages with attachments, it only drafts them, and the user presses send in the end.
02
After an instant rapport, can a lasting bond develop over time?
TAG464@@
In Sensor Tower’s September data, Muse’s next-day retention was nearly twice ChatGPT’s, but it also dropped off quickly and was overtaken by days 8 and 9. In other words, the first experience of handing over control may be great, but people’s affairs are fragmented; serving them consistently still depends on how the AI’s own capabilities are managed.Recently, over a weekend, I happened to have a travel plan for Hohhot.While browsing the feed, updates about traveling to Hohhot would surface from time to time among the various types of news I was interested in. This made it very natural for me to come across the information while scrolling, rather than passively receiving a dull notification.Muse also turns this information directly into visual cards. They aren’t particularly good-looking, but they are genuinely good enough. News and notifications are cleverly combined here, making for a very comfortable experience. The pacing of the feed is also handled well: it doesn’t flood you with a large amount of unseen new content every time you manually refresh. Clearly, the goal here is not retention.
Travel information naturally appears among the news I’m interested in; news and notifications become one thingWhen it needs to interrupt me, it doesn’t hold back either. Muse found information in my inbox that I needed to be notified about and notified me directly, so accurately that I was taken aback by how precisely it knew what to tell me.
Muse picks out the one email in the inbox worth interrupting the user for
03
WorkBuddy is at the workstation,
while the personal Agent is on the road
WorkBuddy's current layout is built around office scenariosThe office agent with the highest DAU in China is Tencent's WorkBuddy (estimated by Citi Research at more than 13 million).I inspected its installation package, and its core files are named after Tencent's coding agent, CodeBuddy. Muse's Android installation package likewise has the program recognizing the same personality files from OpenClaw. Both grew out of developer tools. The difference is that WorkBuddy focuses on workplace superpowers, dividing capabilities into three tiers—Skills, Experts, and Expert Teams—and users have to act as the supervisor who staffs projects. This makes sense at the workstation, but at home, few people are willing to play manager for a while just to book a ticket.
Both installation packages show that they grew out of developer toolsThe two come closest in how they handle payments, meaning that agents will eventually help people make purchases or payments, so this should be considered from the outset; let's look at the thinking and actions of the two.WorkBuddy has a built-in AI-exclusive WeChat Pay card, linked to the user's WeChat Pay account and kept separate from the balance. Payment instructions are handled by a deterministic program, while the raw instructions are not exposed to the model.Among the three office agents I tested side by side, this was the most carefully designed set of guardrails.Muse's Android installation package also includes a Link virtual card, and the agent does not have access to your credit card. When money is involved, both choose to open a separate account.But actually ordering a cup of coffee is still difficult. Earlier, when we tested WorkBuddy, I directly asked WorkBuddy whether it could help order coffee. It thought for a long time, then, after deducting 14 points, answered that it could not. I connected “Meituan Life Assistant” myself, linked my account, and went back and forth for several rounds. What I ended up with was an in-store group-buy voucher; even things that are normally very easy to choose, such as sugar level and ice amount, had to be typed out. The user has to know in advance which service to connect and which account to link, as if going through a procurement process.
What WorkBuddy ultimately managed to order was an in-store group-buy voucherBeyond money, the two have different rules.WorkBuddy's rules apply to files and the wallet. It runs on the user's own computer; the sandbox allows it to read the entire hard drive, writes are limited to the temporary directory, and access to actual files must be authorized individually. Muse adds further rules concerning people. As mentioned earlier, it can speak on the user's behalf to merchants, but can only make calls to family and friends.WorkBuddy was built to mirror an organizational chart, while Muse was built to mirror a map of relationships based on closeness and distance.
Of the six, only two have similar origins; the other five each go their own way.Looking at the domestic market, the org-chart model has long been public infrastructure, and anyone can connect to it through WeCom, DingTalk, or Lark. In Qwen Office’s proprietary code, WeCom appears nearly twice as often as Alibaba’s own DingTalk.According to reports, “Xiaowei,” which can send messages to people on the user’s behalf in WeChat, began its gray rollout in June and had still not reached all users by mid-October. The new version of Ant’s “Lingguang” and ByteDance’s “Xiaodou” have also not officially launched. I infer that three domestic office agents have already launched and that a personal agent will appear soon.
04
589@@
Where the Next Breakthrough Lies
Muse’s success lies in simplifying the complex concepts behind Agents into something that everyone can use. As more users get started at low cost and go deeper, we may get a glimpse of several future directions for personal agents:- First, the new app store will grow into the shape of “Ideas.” What sits on the shelves will shift from individual skills to things already thought through for the user. Whoever can turn the Ideas page into a slot merchants scramble to occupy will control the distribution entry point in the agent era.
- Second, multi-agent collaboration will be hidden behind a single contact. Users want one fixed conversational partner, not a pile of assistants. The Android installation package already includes subagents, which reveal only a single line in the status bar, such as “3 subagents working.” The body is no different: phones, glasses, and small household devices are all connecting one by one, while the name speaking to the user remains the same.
- Third, personal agents will move from “mine” to “ours.” The interface copy shows that Meta built a shared-agent system, branching off a bounded version of a user’s Muse for others to use; the option to create one has now been disabled. On October 20, Muse’s sharing links were also suspended after the cybersecurity company Claroc found that more than 6,800 sharing pages could be located through public DNS records, according to Ars Technica. It was a door opened and then closed again, and it was not the only one to close; yet the Android installation package already has new ones ready, allowing users to invite others to view their “space.” Demand objectively exists. After all, few matters at home concern only one person.
- Fourth, long-term retention depends on frequent everyday details. A promotion offering $1,000 in savings leans more toward the operational side and can win a single reshare. If users are still willing to open it on the ninetieth day, that must come from improved experiences across countless small things in daily life.
The creation entry point for shared agents has been disabled; existing ones can only be viewed and revoked.▎TAG639@@Three Realms of Seeing the Mountain
TAG645@@For developers who had spent a few months running Hermes, openclaw, or opencode themselves, nothing Muse does is new, yet it was still highly successful. This once again proves that new technology must be translated into a mass-market product, with an enormous gulf in between that only a very few people can cross first.In an era of monumental technological progress, where the experience of most products remains somewhat fragmented, Meta does come in like a shaft of light, cutting through and illuminating a space of its own. This brings to mind the classic three stages of seeing a mountain:
- Seeing a mountain as a mountain: Muse is OpenClaw, sharing the same SOUL.md and the same birth ritual.
- Seeing a mountain as not a mountain: One requires users to learn it first; the other has already learned its users.
- Seeing a mountain as a mountain again: Muse is still OpenClaw, except this time Meta has built a path at the foot of the mountain that ordinary people can walk up, too.
That SOUL.md is right there on Muse's profile page—you can open it to read or edit it, but most users will never click it in their lifetime. Perhaps this is exactly what Meta got right.
Click keyword to view the corresponding hands-on test results
AI Product Reviews
AgnesCode丨Fliggy AIWorkBuddyMore AI Product ReviewsMultimodal ReviewsMinimax H3 vs Seedance 2.0 Fast
Foundation Model ReviewsGPT‑6 Astra, Fable 5.1 and Sol
TAG781@@More foundation model evaluations@@
All aboard—take a tour of the highlights of the world's top AI conferences@@
Exclusive access:@@
Expert presentation slides@@
Full conference reports@@
Popular paper analyses@@
Interviews with rising academic stars@@
Scan the QR code above@@
or click@@“@@Read the original article@@”@@to follow the dedicated section.@@
Reposting in any form on websites, forums, or communities without authorization from “AI Tech Review” is strictly prohibited!@@
For reposts via the WeChat official account, please first leave a message in the “AI Tech Review” backend to obtain authorization. When reposting, credit the source and include this official account’s card.@@
Original article by Leiphone. Unauthorized reproduction is prohibited. See@@Reposting Notice@@.@@
