Tencent Marvis Deep Test: 5 Useful Features + 3 Areas Where Kaihe AIBOX Wins
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Summary: Tencent Marvis launched May 20 as a system-level AI assistant with 1 supervisor + 5 specialist agent architecture and 10 million free tokens daily. After two weeks of testing, 5 features genuinely impress — but 3 areas show clear gaps versus Kaihe AIBOX. Not a hit piece — a guide to choosing the right tool.
5 Genuinely Useful Features
1. Deep system-level control. Not surface-level "open browser" stuff. "Clean C: drive redundant files" — it analyzed disk usage, listed temp files/cache/old update packages, cleaned by partition, freed 23GB. Most AI assistants stop at "search files." Marvis directly modifies system settings.
2. Semantic file search. My most-used feature. "Find last month's PDF with user retention data" — it understands "user retention" as semantics, not filename. Searches across hundreds of files and pinpoints the right one. Also searches text inside images — "find that screenshot with Python exception handling" actually works.
3. Multi-agent collaboration. The 1 supervisor + 5 specialist architecture isn't marketing fluff. "Analyze student grades in desktop folder, generate individual reports and class summary" — supervisor decomposed the task, dispatched to file processing agent (read data), analysis agent (statistics), document agent (output). 40+ documents processed in 8 minutes. Manually? An afternoon, maybe.
4. Cross-device control. Sending commands from phone to control PC is incredibly practical. In a meeting, I told Marvis on my phone: "send the project proposal from my desktop to Feishu." It executed. Works even when the PC is locked.
5. Privacy mode. All inference runs locally, works offline. Essential for processing contracts and financial data. Uses Qwen on-device model (7B-class), sufficient for routine file processing.
3 Areas Where Kaihe AIBOX Wins
1. 24/7 operation. Marvis is desktop software — turn off the PC, it stops. No scheduled tasks, no midnight data scraping. Kaihe AIBOX is standalone hardware that runs continuously. OpenClaw's heartbeat mechanism enables 50+ day uptime without restart. If you need "AI that never sleeps," Marvis can't; AIBOX can. This is an architectural gap — software vs. hardware continuity.
2. Message channels. Marvis cross-device works through Tencent's own channel — phone app controls PC. But what if you want to chat with AI directly in WeChat? Marvis doesn't support it. AIBOX natively integrates WeChat, Feishu, Enterprise WeChat via OpenClaw. Your WeChat IS the AI console — no extra app needed. Small difference on paper, big in practice: WeChat is already open; another app is another switch cost.
3. Data ownership and customizability. Marvis privacy mode exists, but efficiency mode routes data through Tencent Cloud. Its agent architecture is closed — you can't write custom Skills. AIBOX runs open-source OpenClaw with a fully open Skill ecosystem. Want a feature? Write it or install from community. Data sits on your local drive — physical ownership is yours. For users who care about data sovereignty, this gap is decisive.
The official store has the full product lineup with compute specs.
How to Choose
Not either-or. Different positioning:
Marvis: daily office work, deep Windows control, Tencent ecosystem dependency, primarily used at the desk. Large free tier, zero cost to try.
Kaihe AIBOX: 24/7 automated workflows, data localization, custom agent capabilities, WeChat/Feishu-native AI control. One-time hardware investment, minimal long-term cost.
For personal office users, Marvis is genuinely good — free tier is sufficient, system-level depth has no current rival. But for automated workflows, always-on AI, or hard data privacy requirements, AIBOX's hardware form plus open-source ecosystem fits better.
Further Reading
- Does AI Still Work After You Turn Off Your Computer? — 24/7 use cases explained
- Full Product Selection Guide — Choose by compute needs