Muse agent users on the web are about to get a clearer window onto the computer that Meta’s assistant works on. A new build of the Muse web app includes a dedicated tab for watching the agent’s virtual machine (VM) screen, and a redesigned voice mode that looks and behaves like a phone call. TestingCatalog, which tracks unreleased features in AI apps, spotted both changes and summed them up on 28 September 2026: “Muse agent will get a dedicated tab for permanent browsing of its VM screen. In addition, the voice mode experience for Muse on the web has been updated and now appears as a voice-calling widget.”
A second clip posted later the same day showed a third change on the way. According to TestingCatalog, users “will soon be able to take over and operate the Muse web browser if needed”. Meta has not announced any of the three features and has not given a release date. Features seen in test builds can still change or be dropped before they reach users.
This article explains what each change does, why a permanent screen view matters for an agent that works in the background, and how the build fits with what Meta said about Muse at launch and at Connect. It also compares the Muse agent with ChatGPT agent and Operator, separates what is confirmed from what is not, and sets out what users and IT teams should do now.
Table of contents
- What the Muse Agent Web Build Adds
- Why a Permanent Screen View Matters for the Muse Agent
- Taking Over the Muse Agent’s Browser
- Muse Agent Voice Mode, From Connect Stage to Browser Tab
- How the Muse Agent Compares With ChatGPT Agent and Operator
- The Muse Agent Timeline in Numbers
- What Is Confirmed About the Muse Agent and What Is Not
- What Muse Agent Users and IT Teams Should Do Now
- Frequently Asked Questions About the Muse Agent Update
- References
What the Muse Agent Web Build Adds
The changes are small in interface terms, but they touch the two things that make an agent different from a chatbot: it acts on its own computer, and it keeps working while you talk to it. TestingCatalog first described them on 27 September in an article on the new Muse web build, then posted a screenshot and a short video on Threads the next day.
A dedicated tab for the agent’s screen
Every Muse agent account comes with its own cloud computer, and the agent uses a browser on that machine to fill in forms, compare prices and make bookings. The new build gives that screen a permanent home. According to TestingCatalog, the screen view “would live in a dedicated tab alongside areas such as identity and updates”. When Muse is carrying out a task through its virtual computer, users can open the tab and watch what is happening. TestingCatalog’s own reading is that this “could make longer computer-use tasks easier to follow without relying only on status messages”.
Voice mode becomes a calling widget
The second change is to voice. TestingCatalog describes a web interface that “resembles a phone call with the agent”, letting people talk to the Muse agent directly from the browser while it continues working. Meta presented Muse Voice Mode at Connect on 23 September as a way to hold long conversations while the agent works in the background. The web widget is the first sign of how that will look outside the phone apps. TestingCatalog also notes that Muse already has a Mac app, so “the same calling experience could also appear there”.
A takeover control follows a day later
The third change arrived in a separate Threads post with a video on 28 September. It says users will be able to take over and operate the Muse browser “if needed”. That is the natural partner of a screen tab. Watching is only half of supervision, and the other half is being able to step in. The clip does not show when the control will ship or whether it will reach the mobile apps at the same time.
| Change | What the build shows | First seen | Meta’s position |
|---|---|---|---|
| Dedicated VM tab | A tab beside identity and updates where users watch the agent’s screen while it works | 27 September (article), 28 September (screenshot) | Not announced |
| Voice-calling widget on the web | Voice mode styled as a phone call in the browser, while the agent keeps working | 27 September (article), 28 September (screenshot) | Voice Mode announced at Connect; web widget not announced |
| Web browser takeover | Users take over and operate the agent’s browser “if needed” | 28 September (video) | Takeover described in Meta’s 8 September security design |
Why a Permanent Screen View Matters for the Muse Agent
A Muse agent is designed to work for long stretches without a person at the keyboard. Alexandr Wang, Meta’s chief AI officer, put it plainly at Connect, as TechCrunch reported: “You can walk away from your computer and it keeps working for you on all the jobs you lined up.” An agent like that needs a way for people to check its work without interrupting it, and a screen tab is the most direct answer.
From a beta widget to a standing tab
The browser view is not new in itself. Screenshots from the closed alpha, reported by TestingCatalog on 7 September, showed “an interactive browser-window widget that the user can open as a link”, which Muse used to open a cinema website and pick seats. That widget carried a “Beta” label. At launch on 8 September, Muse also showed a short snippet of active work under its avatar and a full activity log of actions and approved permissions.
What changes now is placement. A widget belongs to one message in one task. A dedicated tab stays put, so the Muse agent’s screen becomes something you can return to at any time rather than something you have to scroll back through the chat to find.
Status messages versus pixels
Status text and activity logs tell you what the agent says it is doing. The screen shows what it is actually doing. For short tasks the difference hardly matters. For a long shopping comparison or a form spread across several pages, it does. A user watching the Muse agent’s browser can spot a wrong size, an unexpected pop-up or a site that has changed its layout before the agent reaches the payment step. That is also the moment when a takeover control earns its place.
What the VM actually is
Meta explained the machine behind the tab in detail when Muse launched. “You and your Muse share your own dedicated computer in the cloud,” its research blog on Muse security says. “Each Virtual Machine (VM) is an isolated linux box with a browser and enough storage, CPU, and memory to do real work.” The agent’s harness, files and tools run in a restricted container with its own full Debian image, filtered system calls and limited kernel capabilities, while security services sit outside it.
The browser is “a real up-to-date Chromium based browser running behind a virtualization layer”. The new tab, in other words, is a view onto a real Linux machine assigned to each Muse agent user, not a simulation or a replay.
Taking Over the Muse Agent's Browser
The takeover clip is the most interesting of the three changes, because Meta described the capability in writing before Muse launched.
What Meta said on 8 September
Meta’s launch-day security post is unambiguous. “Users can see what Muse is doing in the browser and take over at any time,” it says. It also explains what happens next: “When the user takes over control of the browser, or while secure credential storage is filling a form, the agent is paused and can’t act at all.” The design treats takeover as a hard handover. While a person is driving, the Muse agent is not.
What the new build suggests
TestingCatalog’s post frames web takeover as something users “will soon be able” to do. There are two plausible readings. Either the web client lacked a takeover control that the mobile apps or the old browser widget already offered, or the capability existed and is being given a clearer, permanent control next to the new screen tab. The build does not say which, and Meta has not commented.
Either way, the direction matches the launch design. More of the Muse agent’s machine is becoming directly visible and directly controllable from the web client.
Why takeover is a security feature
Takeover is not only about convenience. Logins are the obvious case. Meta says Muse captures usernames and passwords through a custom interface and routes them “directly to authd”, a credential service outside the agent’s container, so the main agent never sees them. The browser sub-agent sees “an accessibility tree snapshot of the page — not the raw DOM”, cannot run JavaScript in the page, and has Chrome developer tools disabled.
Classifiers also watch for prompt injection in pages, images and downloaded files, for personal data leaving through the browser, and for “attempted submission of high risk forms”. Checkouts trigger a human approval with the exact purchase details every time. A visible screen and a takeover control give people a practical way to act on those prompts. It is basic cybersecurity practice applied to an agent: when something looks wrong, a person takes over.
The stakes are already visible
Muse has had a difficult first month on security. A researcher published a working hijack of the Mac app, covered in our report on the Muse security flaw, and Meta later added a clearer Muse safety warning after a bug bounty report. Meta prices the risk openly. Its public bug bounty pays “up to $300,000 for valid reports, including up to $130,000 for successful prompt injection attempts that affect one user”.
Muse Agent Voice Mode, From Connect Stage to Browser Tab
The voice-calling widget is the web version of a feature Meta put on stage at Connect. Our Connect roundup of the Muse updates covers the full list of announcements. This section follows voice specifically.
What Meta announced at Connect
Meta’s Connect recap frames voice as more than dictation. “Talking to an agent as capable as Muse is different from voice chat with an assistant,” it says. Users can also “design your Muse’s voice just by describing what you want: faster, slower, or even a specific accent.” Meta also introduced Muse Realtime Avatar, which gives the Muse agent an animated face and body for real-time video chat.
The web widget looks like a phone call
The design choice matters. A call keeps talking and working apart. The Muse agent can carry on with a task in its VM while the conversation continues in a compact widget, instead of the voice screen taking over the whole window. That fits the way Muse is meant to be used, with long tasks running in the background and short check-ins by voice.
It also matches Meta’s wider interest in calls. The Muse agent gained the ability to make phone calls on a user’s behalf in mid-September, so a call-style interface for talking to the agent itself is a consistent step.
The road from Meta AI voice to Muse voice
Voice has moved quickly. In May 2026, Meta launched voice conversations powered by Muse Spark in the Meta AI app, starting in the US and Canada. On 10 September, two days after Muse launched, TestingCatalog found a dedicated voice mode being prepared for Muse, with a voice button, a selector of several voices and the ability to generate a custom voice on request. Some internal labels referenced Microsoft voice technology, although that does not mean Meta will use it in production. Connect confirmed Voice Mode on 23 September, and the web widget appeared four days later.
Where the Mac fits
Muse for Mac already exists and, according to Meta’s Connect recap, “Computer use is now available with Muse for Mac. With your permission, Muse can now drive any app on your Mac.” A Mac user therefore has two machines in play: the Muse agent’s cloud VM and their own computer. A calling widget that sits in a corner while the agent works on either one would suit the desktop well, but Meta has not said it is coming there.
How the Muse Agent Compares With ChatGPT Agent and Operator
Watching an agent’s screen and taking over its browser are not Meta inventions. OpenAI has offered both since early 2025, and Meta itself bought a company built on cloud agents.
OpenAI got there first
OpenAI’s Operator, launched in January 2025, let users “take over control of the remote browser at any point”, and was trained to ask users to take over for logins, payment details and CAPTCHAs. ChatGPT agent followed in July 2025. It works “using its own virtual computer”, and “an on-screen narration provides visibility into exactly what ChatGPT is doing”. Users can take over the browser from a menu.
OpenAI’s help centre adds that screenshots are not captured while the user controls the browser, and that deleted chats and their screenshots are removed from its systems within 90 days. OpenAI is also reported to be preparing an always-on assistant called ‘o’, which would compete with Muse directly.
Meta already owned the pattern
Meta acquired Manus at the end of December 2025. At the time, TestingCatalog reported that Manus had processed more than 147 trillion tokens and “supported the creation of more than 80 million virtual computers”. According to CNBC, Manus agents “previously operated exclusively in the cloud” and were typically used through a web interface until a desktop app arrived in March 2026. The Muse agent’s per-user cloud VM follows the same basic model, so a screen tab and takeover control are ideas Meta’s own teams know well.
Where the Muse agent differs
Meta’s design makes some distinctive choices. Approvals appear “directly within the client UI — not via their conversation with Muse”, so a web page cannot fake a consent message inside the chat. Users can grant permission once, for a session, for a task, for a set period or permanently. The browser sub-agent works from an accessibility tree rather than raw page code. And the Muse agent is paused completely during takeover.
| Feature | Muse agent | ChatGPT agent | Operator (2025) |
|---|---|---|---|
| Where the agent works | A dedicated cloud VM per user: an isolated Linux box with a Chromium browser | “Its own virtual computer” with text and visual browsers | A remote browser |
| How you watch | Activity snippet and log at launch; a dedicated screen tab in testing | On-screen narration and a view of the browser | A view of the remote browser |
| Taking over | “Take over at any time” (8 September); a web control shown as coming soon | “Take over browser” from a menu; prompted for logins | Takeover mode; asks for logins, payments and CAPTCHAs |
| During takeover | Agent paused and “can’t act at all” | Screenshots are not captured | Does not collect or screenshot what the user types |
| How it reads pages | Accessibility tree snapshot, not the raw DOM | Screenshots of the virtual browser plus a text browser | Screenshots of the browser |
| Approvals | Sentinel dialogs outside the chat; five grant types | Asks permission before actions of consequence | Asks before finalising significant actions |
The Muse Agent Timeline in Numbers
The pace of change is part of the story. Twenty days after launch, the Muse agent has gained or previewed a string of features, and the VM itself has become steadily more visible to the people who use it.
Twenty days, six visible steps
Launch on 8 September came with Meta’s security design and the takeover promise. Two days later, voice mode work appeared in testing. On 18 September, TestingCatalog found an unreleased Mail tab. Connect followed on 23 September. By 25 September, Muse was offering a full file browser into its own VM, as The Verge reported. The screen tab and web voice call surfaced on 27 September, and the takeover clip on 28 September.
How many machines the tab could reach
Meta says each Muse user gets a dedicated VM, so the screen tab is, in principle, a view onto millions of machines. TechCrunch reported install estimates from three trackers on 25 September. They differ widely, and an install is not the same as an active account, so treat these as a rough indication of scale rather than a count of VMs. Our look at whether Meta’s Muse bet is working examines the adoption data in more detail.
What Is Confirmed About the Muse Agent and What Is Not
Most of this story comes from pre-release builds, so the level of evidence behind each claim matters more than usual.
Confirmed by Meta
Meta has confirmed the dedicated cloud VM, the Chromium browser, the ability to see and take over the browser, Sentinel approvals, Voice Mode, custom voices, Realtime Avatar, Muse Mail and computer use on the Mac. These come from Meta’s research blog and its Connect recap.
Seen only in builds
The dedicated screen tab, the voice-calling widget on the web and the new web takeover control have only been seen in a test build and in short clips shared by TestingCatalog. Meta has not announced them, so the Muse agent that users see when they ship could look different.
Still unknown
It is not known when any of the three changes will ship, or whether they will reach iOS, Android and the Mac at the same time. It is also unclear whether the screen tab will show a live stream or periodic snapshots, and whether voice sessions will count against Muse’s token allowance.
| Claim | Source | Evidence level |
|---|---|---|
| Each user gets a dedicated cloud VM with a Chromium browser | Meta AI Research, 8 September | Confirmed by Meta |
| Users can see the browser and take over at any time | Meta AI Research, 8 September | Confirmed by Meta |
| Voice Mode, with voices designed by description | Meta Connect recap, 23 September | Confirmed by Meta |
| Computer use on the Mac | Meta Connect recap, 23 September | Confirmed by Meta |
| Dedicated tab for the VM screen | TestingCatalog, 27 and 28 September | Seen in a test build |
| Voice-calling widget on the web | TestingCatalog, 27 and 28 September | Seen in a test build |
| Web takeover control “soon” | TestingCatalog video, 28 September | Seen in a test build |
| Release dates, platforms and usage cost | None | Unknown |
What Muse Agent Users and IT Teams Should Do Now
None of these changes needs action today, but they are a good prompt to review how the Muse agent is being used at home and at work.
For people already using Muse
Watch the screen during long or unfamiliar tasks, especially the first few times you ask the Muse agent to use a new site. Take over for logins and anything involving payment details, as Meta’s own design intends. Prefer one-time, session or task-scoped approvals to permanent ones until you trust how the agent behaves on a given site.
For IT and security teams
A consumer agent with its own browser and access to email is effectively a new endpoint. If staff connect a Muse agent to work accounts, company data can end up on a Meta-hosted VM. Include personal AI agents in acceptable use policies and in any vulnerability assessment of how staff reach company systems. For businesses that want agents to do real work, a managed approach to autonomous AI agents keeps accounts, logs and approvals inside the organisation.
For people outside the US
Muse is still unavailable in many markets, and Meta has not published an international rollout schedule. TestingCatalog notes that Meta has opened an early-access route for upcoming capabilities, and that users can ask Muse itself to tell Meta they want to join.
Frequently Asked Questions About the Muse Agent Update
What is the new VM tab in the Muse agent?
It is a dedicated tab in a test build of the Muse web app where users can watch the screen of the agent’s cloud computer while it works. Meta has not announced it.
Is the Muse agent voice-calling widget available now?
Not publicly. Muse Voice Mode was announced at Connect on 23 September. The web version styled as a call has only been seen in a test build.
Can I already take over the Muse agent’s browser?
Meta’s launch-day security post says users can see what Muse is doing in the browser and take over at any time. A new web control for this was shown on 28 September as coming soon.
Does the screen tab show my own computer?
No. It shows the browser on the agent’s cloud VM. Computer use on your own Mac is a separate feature that needs your permission.
Does the Muse agent work outside the US?
Availability is still limited, and Meta has not announced a wider rollout date.
References
TestingCatalog: Meta prepares screen viewing and web voice calls for Muse
TestingCatalog on Threads: Muse agent will get a dedicated tab for its VM screen
TestingCatalog on Threads: Muse users will soon be able to take over the web browser
Meta AI Research: How We Built Safety Into Muse
Meta: Everything We Announced at Meta Connect 2026
TechCrunch: Everything new coming to Meta’s AI agent Muse
TechCrunch: Meta is putting its muscle behind Muse as the AI app takes off
TestingCatalog: Meta’s Muse agent enters closed alpha with invite codes
TestingCatalog: Meta prepares custom voices and model upgrades for Muse
The Verge: Meta makes the Muse filesystem even more accessible
OpenAI: Introducing ChatGPT agent
OpenAI Help Center: ChatGPT agent
TestingCatalog: Meta acquires Manus to expand its AI agents platform
CNBC: Meta’s Manus launches desktop app to bring its AI agent onto personal devices
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
