From Chat‑Only to Full‑Machine Control: Testing Marvis System‑Level AI Agent

Most AI assistants on the market remain confined to chat windows, offering only textual suggestions and lacking the ability to directly manipulate local files, system settings, or hardware. The article distinguishes two categories of AI assistants: (1) conversational AI that runs in browsers or standalone windows and can only answer questions or generate text, and (2) system‑level agents that deeply integrate with the operating system, gaining permission‑based access to files, processes, hardware, and can translate natural language into concrete actions.

Marvis belongs to the latter category. Its architecture combines three technical layers: the system layer (obtaining OS‑level permissions to read and modify resources), the chip layer (leveraging local CPU, GPU, and NPU for on‑device model inference, reducing cloud token consumption), and the application‑ecosystem layer (installing, configuring, scheduling software and enabling cross‑device interactions).

The core framework hosts multiple specialized agents: PM Agent (interprets user intent and distributes tasks), File Agent (handles document, image, and table extraction and bulk organization), Computer Agent (performs hardware checks, disk cleanup, system optimization, and diagnostics), App Agent (manages software installation and configuration), Browser Agent (scrapes web content and fills forms), and Search Agent (conducts web‑wide information retrieval and analysis). Users interact via a unified “Office” UI that displays task progress and logs, ensuring transparency.

Marvis offers two operating modes. The Efficiency Mode focuses on speed, using cloud resources for complex tasks while keeping file and system operations local. The Privacy Local Mode processes all data—including documents, images, and reasoning—on the device, never uploading to the cloud, making it suitable for sensitive information and offline use. High‑risk actions such as bulk deletions or system changes trigger explicit confirmation dialogs.

Four high‑frequency usage paths are demonstrated with exact natural‑language commands and observed outcomes:

Scenario 1 – Semantic file search and bulk archiving: The command scans a target folder, classifies files by content, proposes a folder structure, and after user confirmation moves files without manual dragging.

Scenario 2 – Disk space management: Scanning the C: drive, ranking files by size, labeling deletion risk, and presenting a cleanup plan before any deletion.

Scenario 3 – System fault diagnosis and performance health check: Collecting CPU, memory, disk, and process metrics, identifying startup items and damaged components, then recommending safe actions such as disabling autostart items or terminating high‑usage processes.

Scenario 4 – Cross‑device remote control: From a mobile device, locating a specific PPT on the PC, confirming the path, and transferring the file to the phone, illustrating seamless multi‑device collaboration.

The article then compares industry roadmaps, noting that Microsoft Copilot and Apple Intelligence rely heavily on cloud compute and specific hardware, whereas Marvis provides a dual‑mode design that works on a broader range of Windows and macOS machines and emphasizes local operation permissions.

Installation steps are outlined: download the Windows/macOS/mobile client, log in with a unified WeChat/QQ account, grant directory permissions, and optionally enable a low‑resource mode to reduce CPU load.

Marvis supplies a generous free quota of 10 million tokens per day; local‑mode tasks consume virtually no cloud tokens, making everyday office automation sustainable.

Safety recommendations include granting initial access only to a test folder, switching to privacy mode for sensitive documents, and confirming any high‑risk operation before execution.

The target audience comprises office workers with abundant documents, content creators needing frequent file organization, regular users seeking simplified PC maintenance, and remote workers requiring cross‑device file access.

In conclusion, AI assistants are shifting from pure conversation to genuine device‑side problem solving. Marvis demonstrates that system‑level capabilities can encapsulate complex OS actions behind natural language, lowering the barrier for non‑technical users to perform tasks that traditionally required command‑line expertise.

https://marvis.qq.com/