ToolJet has published a worked example of an agent building an entire internal application, not a screen. In Build a Geofence Compliance Register with ToolJet MCP, Athulya R hands ToolJet’s MCP server a single consolidated prompt for a fleet compliance app, six tables named in the prompt, depots, zones, vehicles, breaches, app_users, and settings, and gets back a data model, 10 queries, and a page carrying 47 components, with one repair cycle. The line that carries the piece is the architectural one: “ToolJet MCP works through the whole application, not just the screen, so the data model, pages, components, queries, and the wiring between them land as one structured ToolJet application.” The result stays editable inside the platform rather than arriving as code someone has to reconcile.
The number in the headline is a vendor number. “You can easily build a 4-page app with ToolJet MCP for as little as $0.17 with a lean model (Luna medium), or pay more for the latest flagship reasoning (Sol).” Seventeen cents is a model-inference cost on ToolJet’s cheapest tier for one specific prompt, not a benchmark, and the 67% developer cost savings and 50% time freed up further down are promotional copy that the article does not measure. What survives is the shape of the claim: the MCP server generates the application object itself, tables included, and role-based access for compliance officers and depot managers is configured after generation rather than prompted into existence. That is a real difference from tools that emit a component and leave the data layer to you.
The catalog knows ToolJet through the other door. The ToolJet provider page lists 4 API pages, all of them the External API for managing the platform rather than building on it. The application the story generates is the object the ToolJet Applications API manages, the workspace it lands in is the ToolJet Workspaces API, and the compliance-officer and depot-manager roles are the ToolJet Groups API. The MCP server dimension is lit, so the catalog agrees the thing the post describes exists. The agentic access profile maps 23 operations, 15 of them acting, which is the management surface an agent could drive once the app exists.
The Kin Score is 46.5, developing band. Discoverability carries it at 68.5 and contract quality at 62.1. Developer ergonomics is 23.8, the lowest facet on the record, and the catalog description explains why: the External API authenticates with a static access token set in an environment variable and passed as a Basic authorization header. The Agent Readiness score is 23.2, agent-aware. Error semantics, OpenAPI examples, and agent skills are unlit, and so is every identity dimension. The post shows an agent building a governed application, with roles and audit, for less than a coffee. The API that governs those applications from outside has no described errors, no examples, and one shared static token. ToolJet has made the agent a builder. The record has not yet made it a citizen.