Claude is the reasoning layer of 409 Labs. It plans each batch, calls our engines as tools, reads reviews and product images, writes localized copy, builds pages through MCP, and stops for human approval before anything is published.
Plans the batch, chooses tools, extracts structured data, writes copy, checks rules, and asks a person to approve.
Architecture of the 409 Agent. The engines on the left are in use today. The Claude orchestration layer is in active development.
Pull ingredients, volume, claims and key features from product pages and detail-page images into a fixed JSON schema.
Label every review with a consistent codebook, such as adhesion, moisture, irritation and price, with positive or negative polarity.
Write natural Japanese and English titles, bullets and descriptions, not literal translations of Korean marketing copy.
Map each product to the correct marketplace category and brand code, replacing brittle keyword rules.
Flag wording that marketplace policies or cosmetics advertising rules do not allow, such as medical or exaggerated claims.
Use tool calling and MCP to move a batch from sourcing to Sheets to Figma to the marketplace, retrying failed steps.
An agent that publishes to real marketplaces has to be predictable. These are the rules we build by.
We are onboarding a small group of Qoo10 Japan and Joom sellers. Tell us what you sell.