Seoul Is Giving AI Chatbots Direct Access to Live City Data

Seoul wants its AI chatbots to tell the truth — about how crowded the subway is, how many shared bikes are parked near Seoul Forest, and whether the air over the Han River is safe to breathe. The city announced today it will pilot an MCP (Model Context Protocol) service that connects large language models directly to its public databases.

The problem MCP solves is familiar to anyone who’s asked an AI for real-time information. Most models generate answers from training data or web search results, which can be hours, days, or even years old. Ask a chatbot how crowded Myeongdong is right now, and it’ll guess — or make something up. That’s the hallucination problem Seoul is trying to fix.

Under the pilot, AI assistants like Claude can query government databases in real time via the MCP interface. Instead of offering stale data, the system returns current figures: how many people are in a given district, when the next bus arrives, how many shared bicycles are available at a specific dock. The data covers 121 major areas across the city and refreshes anywhere from every few seconds to every five minutes, depending on the dataset.

Seoul’s plan goes beyond this initial pilot. The city says it will convert existing public data APIs into MCP-compatible interfaces, making it easier for any AI agent — not just the one in this trial — to pull live government data without writing custom integration code.

The pilot opens 100 slots. Residents can apply through Kakao’s PlayMCP platform. Approved users gain access to real-time crowd density, public transit schedules, weather conditions, and cultural event information across the capital.

For a city of nearly 10 million people, the implications are practical. Imagine asking an AI not “what’s the weather like in Seoul” but “is it raining in Gangnam right now, and should I take the subway instead of walking to the meeting.” The line between helpful and hallucinated narrows when the model is wired into a live data feed, not the open web.