Skip to main content

Concepts

You don't need to understand the code to build safely — but a few concepts help you know why the system behaves the way it does, and where to look when something surprises you.

Dashboard id​

Every dashboard has a single dashboard id (for example 300). It ties the frontend and backend together and identifies your gateway route. You pass it once, at scaffold time, and everything downstream uses it.

The gateway​

Your dashboard never talks to raw services. It talks to the gateway, which exposes a known set of endpoints. Because the set is known and discoverable, the AI can query what's real instead of guessing.

MCP servers​

Model Context Protocol (MCP) servers are the AI's source of truth:

  • guide-mcp — React Task Recipes (step-by-step procedures), guardrails, theme tokens, component templates, and layout patterns.
  • guide-angular-mcp — the same for Angular dashboards (standalone components, signals, TanStack Angular Query).
  • api-mcp — gateway endpoints, OpenAPI specs, endpoint validation.
  • backend-mcp — backend recipes, auth/db/logging templates, endpoint discovery.
  • orchestrator-mcp (optional) — fronts the guide/api/backend servers behind a single guided workflow so the recipe steps run in order and the verify gate can't be skipped (see The orchestrator below).

When the AI needs a fact — an endpoint, a token, a template — it asks an MCP server rather than inventing an answer.

Two frontends, one model​

Dashboards come in two frontend flavours — React and Angular — guided by guide-mcp and guide-angular-mcp respectively. Everything on this page is identical for both; only the framework idioms inside the recipes and templates differ (for example React uses TanStack Query, Angular uses TanStack Angular Query). You pick the stack at scaffold time and the right guidance is wired in.

Task Recipes​

A Task Recipe is an ordered procedure for a specific job ("add a chart", "create a backend endpoint"). It lists the steps, the tools to call, the files to create, and a final verify gate.

The verify gate​

The verify gate is what makes "done" trustworthy. It runs a real check — a build, a request, a contract test — and only passes when the check passes. If verify fails, the work isn't done, no matter what the AI says.

The three layers, together​

  • Static rules set the boundaries before work starts.
  • Live MCP queries keep every fact grounded in reality.
  • Task Recipes + verify prove the result.

Hold all three and an AI assistant becomes a safe collaborator instead of a confident guesser. See Why for the full rationale.

The orchestrator​

By default the AI is trusted to follow the recipe steps and run the verify gate. The optional orchestrator turns that trust into enforcement: it picks the right recipe, walks the AI through the steps in order, runs the discovery tools itself, and refuses to mark a feature done until the verify gate passes. Projects that want the guardrails to be unskippable register the orchestrator instead of the individual servers ("rails-only" mode).