What Do You Build With a Brand-New MCP? We Built a "Launch War Room"

ToolJet has developed an AI-native internal tool builder called ToolJet MCP, which goes beyond generating an app from a description. Unlike previous tools that created a code dump, MCP can work directly inside ToolJet, creating and refining components, queries, data, and workflows that a human builder would use. This results in a proper ToolJet app that remains inspectable and maintainable, with a much larger scope for design, building, testing, and iteration.

To demonstrate the capabilities of MCP, ToolJet created a real-world example called Shiproom, an internal app to track launch checks, blockers, and approvals. The app addresses the common Thursday afternoon dilemma faced by software shipping teams. When presented with a prompt covering the entire application, including each page's function, data behavior, blocker and approval rules, and visual direction, MCP generated the design, built the underlying tables and logic, assembled the app, and checked the result against the original design.

Shiproom features a restrained design with cool greys, compact rows, a limited blue accent, and status colors only where necessary. It includes a launch list page showing upcoming and recent launches, with a countdown, readiness score, and the biggest blocker highlighted. The war room page is divided into four sections: critical checks, blockers, approvals, and the final ship decision.

This design ensures the workflow remains visible without turning it into a miniature project management suite. Additionally, Shiproom includes populated, empty, loading, and error states, ensuring a complete user experience.

The data model for Shiproom consists of three tables: launches, launch_checks, and launch_events. Instead of using a permanent status field or stored readiness percentage, the current state is derived from event history. This append-only model allows for easy explanation of decisions six weeks later without constantly overwriting status fields. The final result of MCP's work includes 94 components, 12 queries, and 2 repair cycles, with no issues encountered.