South Korea’s two dominant tech platforms are placing opposing bets on how AI agents should connect to services. Kakao is building an open tool hub that lets any agent call into its ecosystem. Naver is building a closed agentic search layer that keeps users inside its own services. The divergence, playing out in real time this month, is one of the cleanest architectural splits in the global agent platform race.

Kakao’s Open Play: 200 MCP Servers, Three Agent Hosts

Kakao’s PlayMCP platform, built on the Model Context Protocol standard, now connects approximately 200 MCP servers to three external AI agent hosts: OpenAI’s ChatGPT, Anthropic’s Claude, and the open-source OpenClaw, according to Digital Today. The OpenClaw integration, announced May 1, was the most recent addition. ChatGPT and Claude support launched in November 2025.

The platform includes Kakao’s own services: KakaoTalk messaging, Talk Calendar, Kakao Map, Kakao Gift, and Melon (music streaming). But it also hosts external MCP servers registered by third-party developers, making it a general-purpose tool hub rather than a captive ecosystem.

Integration works through a one-time token flow. Users click “Connect with OpenClaw” in the PlayMCP toolbox, consent to data sharing, and paste a generated connection prompt into the OpenClaw chat window. The token expires after 10 minutes, and connected services can be disconnected from PlayMCP settings at any time, as Seoul Economic Daily reported.

“PlayMCP aims for the openness that allows MCP developers to connect the servers they create with various AI services to experiment and expand,” said Yoo Yong-ha, performance leader at Kakao AI Connect, per Seoul Economic Daily.

Naver is taking the opposite approach. Its AI Tab, currently available on a priority basis to Naver Plus Membership subscribers, is an agentic search interface that weaves together Naver’s shopping, places, blog, and community services into natural-language responses, according to BigGo Finance.

The integration goes deep. In shopping, AI Tab combines community reviews with product data to recommend items and route users directly to purchase. In the places domain, it analyzes blog reviews and location data to recommend venues and connect to reservations. Naver CEO Choi Soo-yeon told The Korea Times in January that the company was investing over 1 trillion won ($691 million) in GPU and AI infrastructure to support agentic AI at scale.

Naver is measuring success by purchase and reservation conversion rates rather than dwell time, and has set a target of transitioning AI Tab to monetization by Q4 2026, BigGo Finance reported. Testing is underway to integrate advertising into AI briefings.

The Architecture Question

The split maps directly onto the broader agent infrastructure debate. Open platforms like Kakao’s PlayMCP bet that agents will be model-agnostic, connecting to services through standardized protocols. Users pick their preferred agent (ChatGPT, Claude, OpenClaw) and access tools through a shared hub. The platform captures value by being the registry, not the runtime.

Closed platforms like Naver’s AI Tab bet that tight integration produces higher-fidelity experiences. When the search layer, commerce engine, and review database are all controlled by the same company, the agent can execute multi-step workflows (search, compare, reserve, pay) with less friction. The platform captures value by owning the entire transaction chain.

An industry insider quoted by BigGo Finance framed the stakes: “AI competition has now moved beyond model performance to ecosystem competition. How platform boundaries are designed will determine future leadership.”

For agent builders watching from outside South Korea, the outcome will signal which architecture scales better for execution-oriented AI: open tool registries that any agent can plug into, or vertically integrated systems where one company controls the agent, the tools, and the transactions.