Tencent's Marvis Bets the AI Agent Race Will Be Won on Your PC, Not in the Cloud

Tencent's newest AI product, Marvis, does not live in a chat window. Built by the same team that runs Tencent's app store, it sits on the user's PC, reads local files, navigates browsers, and manages device permissions. The company is publicly betting that is where the durable AI agent gets built, not in the cloud.

The product launched on May 20 from Tencent's Yingyongbao (应用宝) team, the same group responsible for Tencent's Android app store, and is being positioned by Tencent executives as a "system-level" agent rather than a general-purpose assistant. The framing is a deliberate counter to the cloud-chatbot model that dominates the current AI agent race. Tencent vice president Lin Songtao laid out the thesis on stage at the World Artificial Intelligence Conference (WAIC), hosted by 36kr, alongside Yingyongbao cross-device lead Cai Jiantao.

The numbers Tencent has disclosed, all self-reported at WAIC and not independently verified, give the bet a proof point. According to Cai, Marvis crossed 300,000 daily active users within two days of launch and held roughly 54% seven-day retention. The team's own breakdown of user scenarios: local files 44%, PC hardware management 28%, browser tasks 18%, app-related 16%, and search only about 6%. The search share is the giveaway. A chatbot-shaped product would invert that ordering.

The mechanism under the framing is what Tencent calls device-level perception. A general-purpose agent typically round-trips screenshots to a cloud model and clicks through interfaces by sight. Marvis, by Tencent's account, reads the file system, tracks browser cookies and permission state, and parses install logs so it can act on the PC the way a local user would, by moving files, opening settings, or recovering from a failed install. The team describes the design as having gone through three throw-out-redo cycles during 2025, with an internal name of "设备AI助手" (device AI assistant) before launch. Third-party hands-on coverage from Kunpeng-ai and ChinaBizInsider tracks the same system-level pitch without independently testing it.

The strategic bet extends beyond Marvis. Lin said Yingyongbao is repositioning for the agent era across three products: Marvis as the device-side system agent, WorkBuddy (DAU #1 in domestic efficiency agents by Lin's count) as the consumer-facing agent, and a supply-side mobile coding tool called 吐司 (Toast) for app creation. The QClaw team has been folded into the WorkBuddy department, though QClaw itself continues. A collaboration with Tencent's Hunyuan team is underway on an on-device model cluster. A separate 36kr English hands-on and Chinese product profiles on aitop100 and yxsoft describe the same positioning.

Mobile is the gap in Tencent's own framing. Cai explicitly de-scoped phones, citing memory and battery constraints, and cast the phone as a controller with the PC or a cloud container as the executor. The choice rules out the consumer surface where most of China's chatbot traffic already lives. It also makes the system-level bet a narrower one than the cloud-chatbot crowd, and that is the trade Lin is willing to make: a smaller addressable device footprint, in exchange for owning the working surface of the devices that remain.

A vision-based general agent that already works at the OS level through permissions and cloud screenshot round-tripping can, in principle, do the same work without owning the OS layer. Whether the system-level framing is a genuine technical moat or a marketing line for a screenshot agent with broader permissions is the falsifier that determines whether the bet pays off. Independent voices, analysts, competitor product leads, or enterprise users, have not yet weighed in on record.

According to Lin and Cai, Tencent's own north star for Marvis is the number of real tasks completed, not DAU or session count. The next marker to watch is whether the Hunyuan on-device model cluster lands in time for Marvis to make the system-level claim about its own inference stack, or whether the model still has to round-trip to the cloud.