After experiencing Tencent's newly released Marvis assistant, I've come to believe that the ultimate destination of Personal AI is the operating system itself.
In 1995, Microsoft embarked on a rather utopian experiment. Believing Windows at the time was too complicated, they attempted to turn the computer desktop into a "room" that ordinary people could intuitively underStand—CLIck on the notebook on the desk to type, click on the wall clock to check the time, eliminating the need to wrestle with counterintuitive navigation paths. The product was called Microsoft Bob, but it was pulled from shelves almost immediately after launch. The reasons were varied, but fundamentally, it never truly penetrated the system's. Whenever a user's needs became slightly more complex, it was powerless to help.
Thirty years later, many AI assistants still haven't fully escaped this dilemma. It was only recently, while testing Tencent's new AI Assistant, Marvis, that I finally saw something different. While others are fiercely competing to be the best AI chatbot, Marvis aims to transform your entire computer into a partner you can command at will. Its official positioning is equally serious: an OS-level personal AI assistant (official website: marvis.QQ.com). However, knowing that the team behind it is the Same one that has built the Yingyongbao APP store for 14 years, this ambition wasn't surprising. Their long-term experience working deep within the ecosystems of both PC and mobile platforms means they understand not just AI, but also devices and systems at their core.
Currently, Marvis is available on Windows PC and Android phones, with iOS and macOS veRSIons in development.
Six Agents: A private AI Team on Standby
The installation process is straightforward, but there are hardware requirements. The minimum for the Windows version is an 8-core CPU, 16GB of RAM, and a solid-state drive. The official word is that compatibility will gradually extend to 6-core CPUs in the future.
Unlike many agent products on the market that require users to build workflows and install plugins themselves, Marvis is ready to use right out of the box. Six agents are pre-configured and require zero setup, making it exceptionally user-friendly for the aveRAGe person. Upon entering the interface, you'll find six agents on standby 24/7. Each agent has its own "workstation." When not on a task, you might see one napping, wandering around the office, exercising, drinking coffee, or even heading to the restroom. Once you issue a command, the task is broken down and assigned to the corresponding team members.
This is a team with a clear division of labor: the PM agent is responsible for understanding your needs, breaking down tasks, and dispatching work; the File Agent handles file searching, reading, and format conversion; the Computer Agent tackles system configuration, hardware checks, and toggling settings like night mode; the APP Agent operates various software on your behalf, such as checking flight tickets or comparing e-commerce prices; the Search Agent quickly locates answers from public resources and provides key citations; and finally, the Browser Agent serves as a web interACTion specialist, handling web page interaction and data scraping. Watching their animations during long task chains adds a layer of unique chArm.
What's most interesting is that these agents can work in parallel. For example, I gave a single command: "My computer is booting up too slowly. Check which startup programs are unnecessary and disable them for me." Noticeably, the interface displays real-time character consumption, which is essentially token usage. The daily free allowance is 20 million tokens, which is Generally sufficient. The results included plain-language explanations for each program, advising on which third-party software was safe to disable as needed. Crucially, for any Operation involving core system configuration changes, it actively seeks confirmation before execution. This "AI won't act on its own" safety mechanism provides considerable peace of mind.
When you've used a computer for a long time, you inevitably encounter a few persistent, nagging issues. "Check battery health." "The internet feels slow; test my current upload and download speeds and latency." "I want to play a new game; check my hardware specs against the minimum requirements and tell me where I fall short." In the past, resolving these meant navigating to the Start menu settings or searching for tutorials online. Now, you can just say a single sentence. Marvis deeply integrates with system Information and configuration portals. It can view hardware parameters and directly modify system settings. It doesn't just open the Control Panel and leave you to find things yourself; it simply gets the job done. For those unfamiliar with computer settings, this is the most intuitive experience. Take a concrete example: you want to turn off ads on the Windows lock screen. Do you know what this setting is officially called? "Disable Windows Spotlight." How many users recognize that this terminology translates to disabling ads? Marvis can disable it with one voice command.
A New AI Gateway for the Local File Era
To test its capabilities, I dug up a long-standing annoyance: a screenshot I downloaded from a group chat two months prior. I only remembered it was related to the game Black Myth: Wukong and had completely forgotten the file name. I typed: "Find images or screenshots related to Black Myth: Wukong from about a month ago." Honestly, I didn't have high hopes. But after taking the task, Marvis first followed a Skill guide to filter 1,000 photos based on file name keywords and time range, then used Python to quickly sift through the results, ultimately finding 8 candidate images. The very first one was exactly what I was looking for.
Moreover, its search scope is remarkably broad, covering file names, document content, imagery within pictures, and text inside screenshots. Even with a vague description, it can retrieve the right item. This capability is indeed a significant leap beyond Windows' built-in search.
Of course, the mention of "scanning images" might raise privacy concerns. Upon first use, you can set the scanning scope yourself. If you see content in the generated knowledge graph that you don't want indexed, a simple right-click can block it. Furthermore, the on-device model itself includes basic filtering, ensuring it's not an indiscriminate scan. Beyond file searching, Marvis can deeply understand and analyze local documents, spreadsheets, and other file types—supporting content optimization, chart generation, copy polishing, and format conversion, making previously tedious and time-consuming tasks feel effortless.
For instance, if you've downloaded a batch of academic papers and don't know where to start reading, you can simply ask Marvis: "Scan these few PDFs and tell me the research question, methodology, and main conclusions of each." Within minutes, you'll have a clear idea of which papers are worth a detailed read and which can be set aside, eliminating the need to flip through abstracts one by one. As another example, after reading The Three-Body Problem, you might have several pages of notes to prepare for an upcoming class presentation. Manually organizing an outline, picking key content, and fitting it into a PPT template could easily take an hour or two. But by asking Marvis to generate a PPT based on your reading notes, it can understand the logic and structure within your notes. There's no need for manual copy-pasting and template fitting; it comprehends the content and generates the presentation directly.
Control Your PC from Your Phone, Be Productive Anywhere
The ability to rEMOtely control a computer from a phone is an extremely practical cross-device experience. After downloading the Windows client from the official website and installing the Marvis App on your phone, logging in with the same WeChat or QQ account seamlessly connects both ends. Specifically, once the phone app connects to the computer, you can see the computer desktop in real-time on your phone and control it directly with touch input. Even if the computer is locked, the phone can still take over, without the need for the computer to remain unlocked. For example, if I'm out and a friend mentions a newly released game, I can pull out my phone and type: "Help me install the game on my computer," so it's ready to play when I get home. Connecting via the WeChat ClawBot is also possible; just scan a QR code with WeChat to log in and start a conversation.
However, I believe the most practical use case is remote troubleshooting for elderly family members. When your parents say, "QQ won't open," previously you'd have to guess blindly over the phone or go over to fix it. Now, you can directly take control of their desktop, glance at the screen, and instantly see what the problem is. The prerequisite is that the Marvis app on your phone can connect to the Marvis client on your parents' computer—the mobile app supports connecting to multiple PC devices. If the computer is offline, the mobile app automatically switches to Cloud Marvis mode, supporting independent invocation of cloud capabilities, allowing most tasks to continue running. There's also a unique trick: directly operating apps on your phone from your computer. Apps like Flush and Feichangzhun are already being integrated through partnerships, a capability that benefits from the Yingyongbao team's years of accumulated cross-platform engine experience. Currently, Yingyongbao supports mobile, PC, and in-vehicle platforms, which means the team has ample cross-device R&D experience.
Like other desktop agents, Marvis also supports custom automated tasks. For example, you can set up a "Shenzhen Houde Pin Yuan SECond-hand housing average price alert," fill in a task description like "Query the latest average second-hand housing price and notify the user," and specify the execution time, such as 10:00 AM every Monday. After that, you don't have to do anything; the task runs automatically at the scheduled time and pushes the results directly to you.
Another feature worth singling out is the custom extension of Skills. Marvis supports one-click Installation of dedicated skill packs and offers numerous third-party Skills. Naturally, you can also import your commonly used Skill files, shaping it to better fit your personal workflow. Users with specific industry needs—like content creation, financial Investment, or academic research—don't have to rely on official iteration cycles; they can install the corresponding knowledge bases and operational logic themselves.
Safety: Efficiency Mode and Privacy Mode
Finally, any agent that handles live tasks inevitably encounters the most critical question: security. Marvis provides two modes: Efficiency Mode and Privacy Mode. When you switch to "Privacy Mode" in the settings, Marvis loads a local, on-device model. All operations, including document and image retrieval, parsing, and reCognition, run entirely locally. Your files never leave your computer, and you can even use this mode offline. The official promise of "data never leaving your domain" finds its truest manifestation when you unplug the network cable.
However, the hardware threshold for experiencing "Privacy Mode" is quite high: it requires at least a 16-core processor, 32GB of RAM, 16GB of VRAM, and 35GB of free SSD space. Therefore, for everyday use, switching back to "Efficiency Mode" is advisable. In this mode, understanding and planning are handled by cloud-based large models (Hunyuan and DeepSeek V4), while execution happens locally, resulting in faster speeds. You can switch between modes as needed.
Making the PC Truly "Understand" People
The way we use computers today is still based on a logic from the 1980s: one task corresponds to One Piece of software. You must find it, open it, and learn its rules to complete the task. Over forty years, there’s been more software and increasingly complex settings, but this fundamental logic hasn't changed. The distance between the average user and their computer hasn't truly shortened. Marvis's logic is reversed: it first understands what's inside this computer, then decides what it can do for you. Take my grandfather, for example. After using a computer for years, he still has no idea where the "Control Panel" is. But he knows his computer boots up slowly, and he knows he wants to find an old photo. In the past, the solution was to call his children for help. Now, even if he can't type, he can simply speak his need directly to the WeChat Clawbot or the Marvis mobile app, and the computer will do it for him.
File search, system settings, cross-device control, automated tasks, privacy mode, and Skill extensions—these appear to be several distinct features, but they all point in the same direction: A personal AI should not be merely a Q&A portal; it should become the task scheduling layer within your device.
I should add that I find Marvis has a sense of Aesthetics. Details like the virtual office animations and task progress presentations make it feel less like a cold AI tool and more like a small, orderly, collaborative studio with a touch of life. It is particularly well-suited for several types of users: heavy local users with numerous and disorganized files; Professionals in finance, law, and healthcare who prioritize data isolation; and efficiency enthusiasts who enjoy customizing their workflows. Unlike some agents that need to be "nurtured" to be useful, Marvis's initial built-in agents and natural language commands are sufficient to cover most daily computer operations. For many people with AI anxiety who don't know where to start, Marvis feels more egalitarian, bARRier-free, and ready to use straight out of the box.
The personal AI of the future will naturally become a part of the operating system, just like search and screenshots have. Thirty years ago, Microsoft Bob tried to make it easier for users to understand computers. Thirty years later, Marvis is trying to make computers understand users in return. That Marvis has reached this stage is intrinsically linked to this team's 14 years of app store ecosystem accumulation and four years of cross-platform, cross-system technical experience. A massive user base has given them deep insight into what users need on different devices and how they use them. This kind of accumulation is difficult for a pure AI company to replicate in the short term—they can quickly train a smarter model, but they can't quickly establish control over the of PCs and mobile ecosystems. With this, a truly usable, system-level AI assistant driven by natural language has finally arrived as a concrete sample within the Chinese-language ecosystem. The focus of so-called "personal AI" has never been solely on the AI, but on the "personal." Marvis has finally allowed us to see that when Technology chooses to walk side-by-side with every concrete user need, the power it releases can be both quietly pervasive and Earth-shaking. This is the moment personal AI begins to be truly useful.
Comments & Questions (0)
No comments yet
Be the first to comment!