Let Your AI Agent Scaffold Apps With seed4j-mcp

If you've ever bootstrapped a Spring Boot + Vue project by hand, you know the routine: pick a build tool, glue in a frontend, add JPA, choose a database driver, wire Liquibase, remember the Maven wrapper, look up that one annotation for the seventh time this year. By the time you reach initial commit, half your motivation is gone.

seed4j already solves most of that. It's an open source application generator that applies modules to a project folder — click click click, project done.

But seed4j is still a tool a human drives. You read the docs, you pick the module slugs, you remember whether the property is packageName or basePackage this release. Fine, but not exactly thrilling.

seed4j-mcp is a Model Context Protocol server that hands seed4j to your AI agent instead. Plug it into Claude Code, Claude Desktop, or Cursor, and the agent can:

browse the seed4j catalog,
pick a coherent stack,
validate properties before applying anything,
and scaffold a project end to end.

You stay in the conversation. The agent does the orchestration.

Heads up — this is an alpha. seed4j-mcp is a fresh first release. The tool surface, defaults, and behavior may change as feedback comes in. It's stable enough to play with, but don't wire it into anything critical yet. Bug reports and suggestions are very welcome.

How the pieces fit together

seed4j-mcp is intentionally thin. Three components:

A running seed4j instance — the actual generator, reachable over HTTP (default http://localhost:1339).
seed4j-mcp — a Node.js MCP server that translates MCP tool calls into seed4j REST calls.
Your MCP client — Claude Code, Claude Desktop, Cursor, etc., speaking MCP over STDIO to seed4j-mcp.

[Claude Code] ⇄ STDIO ⇄ [seed4j-mcp] ⇄ HTTP ⇄ [seed4j server]

That separation matters: seed4j-mcp does not embed seed4j as a library. It's a translator. seed4j can evolve on its own and the MCP server keeps doing its job — no version-pinning gymnastics.

One implementation detail worth knowing: seed4j-mcp speaks MCP over STDIO. The MCP framing lives on stdout, which means anything else writing to stdout will corrupt the stream and your MCP client will hang. The server routes its startup errors to stderr for that reason. If you fork the project and add logging, use console.error or write to a file — never console.log.

Tutorial

Prerequisites

Three things on your machine:

Node.js 20+ — seed4j-mcp ships as an npm package and runs under Node (yes, even if you're a Java person).
A running seed4j instance — see the seed4j docs for how to start one. Default URL: http://localhost:1339.
An MCP-aware client — I'll use Claude Code, but Claude Desktop and Cursor work the same way.

Quick sanity checks:

node --version            # v20.x or newer
curl http://localhost:1339/api/modules | head -c 200

If the curl call returns JSON, you're ready.

Getting started

You don't need to install seed4j-mcp globally. The package is published on npm and the recommended pattern is to let your MCP client launch it on demand via npx. First run pulls it down, later runs use the cache.

If you'd rather install it globally:

npm install -g seed4j-mcp

Or build from source if you want to contribute:

git clone https://github.com/avdev4j/seed4j-mcp.git
cd seed4j-mcp
npm install
npm run build

Run it with Claude Code

Claude Code has an mcp subcommand built for exactly this. Pick a scope based on how widely you want the server available:

# Local: just you, this project (default scope)
claude mcp add seed4j -- npx -y seed4j-mcp

# Project: committed to .mcp.json, shared with collaborators
claude mcp add seed4j --scope project -- npx -y seed4j-mcp

# User: available across all your projects on this machine
claude mcp add seed4j --scope user -- npx -y seed4j-mcp

Running seed4j on a non-default URL? Pass it as an env var:

claude mcp add seed4j --env SEED4J_BASE_URL=http://localhost:7471 -- npx -y seed4j-mcp

Verify the server is wired up:

claude mcp list

Restart Claude Code and the tools become available to the agent automatically.

Run it with Claude Desktop or Cursor

Both clients read a JSON config. Add an entry pointing at the npx entrypoint:

{
  "mcpServers": {
    "seed4j": {
      "command": "npx",
      "args": ["-y", "seed4j-mcp"],
      "env": {
        "SEED4J_BASE_URL": "http://localhost:1339"
      }
    }
  }
}

Restart the client and you're done.

Example usages

The real test is what the agent does once the tools are available. A few prompts that exercise different flows.

1. "Scaffold a Spring Boot + Vue webapp"

The curated stack path. You name the vibe, the agent picks a preset. Under the hood it will:

Call list_presets to see what's on offer.
Pick the one that matches (e.g. "Webapp: Vue + Spring Boot").
Call get_preset_details for the ordered module list.
Call create_project to initialize the folder.
Call apply_preset to apply every module in order with a shared property map.

Your prompt is one line:

Scaffold a new project at /tmp/my-webapp using a Vue + Spring Boot preset. Use com.example.webapp as the Java package and mywebapp as the base name.

2. "I want a custom stack, not a preset"

The custom stack path. The agent will:

search_modules to find candidates matching your description.
get_module_dependencies to learn the prerequisite ordering and any feature choices (yes, you have to pick a datasource flavor — the agent won't guess for you).
validate_properties to dry-run the property map before touching the filesystem.
apply_modules to apply the full ordered list in one batch, stopping at the first failure.

Example prompt:

Create a Java library project at /tmp/my-lib with Maven, Jacoco, and SonarQube wired up. Package com.example.lib.

3. "What's already in this project?"

When the folder exists, get_project_status tells the agent which modules have been applied and what properties are in play — so it suggests sensible next steps instead of stomping over existing state:

Look at /tmp/my-webapp and tell me what's wired up. What would you suggest adding next?

4. "Show me the catalog"

You can also use the agent as a guided browser. It will reach for search_modules and get_module_details and summarize the catalog:

What persistence modules does seed4j support? Compare Postgres, MySQL, and MongoDB options.

Available tools

The agent gets a small but expressive toolbox covering module discovery, preset lookup, property validation, and project application. Rather than reproducing the full list here (and watching it drift the next time a tool is added), check the up-to-date catalog directly in the repo:

github.com/avdev4j/seed4j-mcp

The README lists every tool with the description the agent actually sees.

See it in action

If you'd like to see what a project scaffolded through seed4j-mcp actually looks like, here's one I generated:

avdev4j/seed4j-mcp-sample-app

It's the output of a real agent session — useful as a reference for the kind of structure, files, and module combination you can expect when you hand the wheel to your AI agent.

Wrapping up

Application generators have always traded flexibility for speed: faster than handcrafting, less flexible than writing it yourself. Putting an LLM in front of one tips the balance — the agent handles orchestration, and you stay at the level of "build me a Java library with Jacoco and SonarQube". No memorizing slugs, no fighting property names.

And the scaffolded project isn't a dead drop. seed4j writes a documentation/ folder alongside the code with the conventions, code style, and recommendations specific to the modules you applied. That folder is a goldmine for the next step: pointing your AI agent at it while you write the business code keeps the generated structure and the hand-written code coherent — same package layout, same patterns, same style — instead of the usual drift between bootstrapped scaffolding and the features you add on top.

seed4j-mcp is Apache 2.0. Source at github.com/avdev4j/seed4j-mcp, published on npm as seed4j-mcp. Issues, PRs, and creative stack requests welcome.

If you ship something with it, I'd love to see it.

推荐订阅源

DEV Community