Testing Tencent Marvis Reveals Claude Code‑Style AI Agent Architecture

After work, the author receives a request for a project budget report that resides on his Mac. Instead of returning to the office or asking a colleague, he opens the Marvis app on his phone, tells it to find the report on the desktop and send it to his phone, and the file appears within a minute.

Marvis as a Dual‑Device AI Agent

Marvis, developed by Tencent, consists of a resident agent on the computer and a companion app on the phone. The computer side is downloaded from marvis.qq.com (or via official Android/iOS releases) and logged in with a QR code, requiring no invitation code.

The author describes the core experience as turning the phone into a voice‑controlled remote for the computer, with the computer‑side agent executing the requested actions.

Agent‑Based Task Dispatch

When the author asks the phone to locate a file on the desktop, Marvis does not perform the search itself; it delegates the job to a File Agent . The File Agent receives a send file command, retrieves the file, and returns it, showing a download button in the chat.

This pattern mirrors the multi‑agent approach used by Claude Code and Codex, where a main agent interprets the user’s intent, breaks the work into subtasks, and assigns them to specialized sub‑agents such as file handlers or shell executors.

Connection setup is straightforward: the phone lists online devices under the same account, the user selects the target computer, and a live screen of the Mac appears on the phone, effectively turning the phone into a second monitor.

Security and Permissions

The author investigates the security model and finds that Marvis does not automatically access the entire disk. Access must be granted folder by folder in the permission settings.

A "low‑resource mode" switch reduces CPU usage at the cost of slower indexing. The local data directory is shown as ~/Library/Application Support/com.tencent.mac.marvis, confirming that all processed files remain on the local machine.

Offline Capability

Even with Wi‑Fi disabled, Marvis can operate. The author placed five sample contracts in a folder and instructed Marvis to rename and categorize them by content. Marvis correctly renamed the files (e.g., "Software Development Contract", "IT Operations Service Contract", "Cloud Computing Service Contract"), listed their paths and sizes, and generated a summary paragraph describing the contracts.

Automation and Knowledge Base

Beyond ad‑hoc commands, Marvis supports scheduled tasks. The author created a simple task named "test" that creates a folder on the desktop at a specified time. When the time arrived, Marvis dispatched the task to the Shell Executor , which checked that the folder already existed and reported its path instead of recreating it.

Marvis also includes a local knowledge base: adding a folder of documents enables cross‑file search. Combined with scheduled tasks, the system can automatically organize newly downloaded files each morning.

The desktop client offers a "Skill Plaza" where users can integrate additional Skills, MCP, CLI tools, and third‑party extensions.

Conclusion

The author finds the Android version smooth and notes that the iOS version is now available in the App Store. Even when the computer is offline, Marvis can switch to a cloud mode to continue operation from the phone.

Overall, the combination of always‑available remote access and a resident, task‑dispatching AI agent lowers the barrier for retrieving files, running commands, and automating repetitive tasks when the user and computer are separated.