Skip to content
CubeDot CubeDot

MCP server & agent assets

The Model Context Protocol is the wire between your blueprint and your agent. cubedot init writes an .mcp.json that points your agent at the CubeDot MCP server for your project, and from then on the agent can read the spec directly.

What it reads is the committed blueprint: your brief, features, functionalities, stories, tech stack, UX, database schema, risks, and build order, plus queries against the project’s knowledge graph (“what depends on this functionality?”). It’s the difference between an agent working from the plan and an agent guessing at it.

Alongside .mcp.json, init writes assets that shape how the agent works:

  • CLAUDE.md / AGENTS.md, the project constitution it loads on start.
  • Feature Implementation Agent, a role tuned to build one functionality at a time against its spec.
  • Verifier, a role that checks the work against its acceptance criteria before you mark it done.

Your project is served as a pinned, read-only projection, so what the agent sees stays stable and consistent within a work session. And you never paste API keys into your agent, the CLI authenticates for you with the token from init.