DEV Community

Abel Legesse Shiferaw
Abel Legesse Shiferaw

Posted on

I built a local-first tool that plans projects for AI coding agents, then checks the code they write

#agents #ai #programming #softwaredevelopment

AI coding agents are good. My prompts were the problem.

I kept starting sessions in Cursor and Claude Code with a vague idea, and ended up with code that worked but had no clear structure. Then I had a second problem: once the code existed, I had no quick way to tell how healthy it really was.

So I built Blueprint Studio, a free, MIT-licensed tool that handles both sides: plan before the agent builds, and check after.

Part 1: Plan before you build

You write a short brief, then answer a guided interview. Blueprint Studio turns that into:

  • An architecture map with a recommended stack
  • An editable handoff package you can drop into Cursor, VS Code, Claude Code, Codex, or JetBrains

The idea is that the agent starts from a real spec instead of a one-line prompt.

Greenfield architecture map

Part 2: Check what you already have

Point it at a local repo and it gives you:

  • A code quality report with a health score
  • Dependency and supply-chain checks, including known CVEs
  • Framework-specific findings for Next.js, Remix, Expo / React Native, and Node.js (Express, Fastify, Koa, Hono, Nest)
  • Project structure, scripts, and an overview, which helps a lot with unfamiliar repos
  • Click-through to the exact file and line in your editor (Cursor, VS Code, Windsurf, JetBrains, Zed, Sublime)
  • Optional AI review on only the files you choose

It works on code you wrote, code an agent wrote, or a repo you just cloned.

Codebase overview with framework findings

Local-first by design

Specs and scans stay on your machine, and there's no hosted backend for your projects. AI steps are optional. When you use them, requests go only to the provider you configure with your own key (OpenAI, Anthropic, Gemini, Ollama, OpenRouter, and others), and Blueprint Studio doesn't proxy them.

The README spells out exactly what gets sent: only the files you select for AI review, plus related static-finding notes and light project context. For the planning flow, only the brief and interview answers you type.

How it's built

React and Vite for the UI, with a Tauri 2 desktop shell in Rust for folder access, opening files in your IDE, and writing handoff files to disk. There's also a browser version, though the desktop app is better for scanning folders.

Honest limitations

  • The planning interview needs an AI provider configured.
  • Deep framework analysis covers Next.js, Remix, Expo, and Node today. Other stacks get the general scan.
  • macOS builds aren't Apple-notarized yet, so you may need a one-line Terminal command after installing (the README has it).

Try it and help shape it

I'd love feedback on two things:

  1. What should a code quality report include so it's useful, not noisy?
  2. Which framework should I add next?

I'm also looking for contributors. Adding a framework profile (Django, Rails, SvelteKit, Laravel, and so on) is a fairly self-contained first PR. There are good first issues and a CONTRIBUTING.md to get you started.

If you try it, tell me in the comments what worked and what didn't.

Top comments (0)