Today, DeepMind announced the release of Gemma 4, a very impressive and powerful new version of the Gemma family of models. As I’ve been contributing to ADK Java a fair bit recently, I was curious to see how I would configure ADK Java agents to work with Gemma 4.

In this article, we’ll explore two paths:

• Calling the AI Studio API surface directly,
• Calling Gemma 4 hosted via a vLLM instance thanks to the LangChain4j bridge.

With the appropriate model weights format, we’ll also be able to run Gemma 4 locally via Ollama. But that’s for another day.

Since I started using Model Context Protocol (MCP) to equip my AI agents with useful tools, I’ve been looking for ways to quickly build and iterate on local servers. A few weeks ago, I shared how to easily build a local MCP server in Java with a custom skill in Gemini CLI. Today, I wanted to put that skill to the test by creating a Wikipedia MCP server.

What’s impressive is that I didn’t even have to leave my terminal or read documentation. The entire process was a conversation with Gemini CLI, leveraging its ability to search the web, find libraries, and even check migration guides!

I’m happy to echo here the release of ADK for Java v1.0, Google’s Agent Development Kit framework to build AI agents in Java. I spent a lot of time on this project. I also wrote the announcement blog post on the Google for Developers blog. And I’ve recorded this YouTube video highlighting some of the new features of the framework, in which I’m demonstrating some of them via an app I built: my Comic Trip agent (pun intended).

Generative AI isn’t just about text or images (with Nano Banana) but it’s also great at generating videos (with Veo 3). And now with the recently released Lyria 3 model from DeepMind, you can create some engaging and creative music with lyrics (generated, or your own) or invent a calming instrumental track to loop in the background of your online TikTok or YouTube Shorts.

And of course, if you’re a Java developer like me, you can do all that in Java!

LLMs are great at generating structured data, in particularly those which support a strict JSON output mode (sometimes also called structured decoding), but sometimes they give you a bit more than the requested JSON. You get a Markdown code block wrapped in “Here’s the data you asked for:” and “Hope this helps!”. If you’re lucky, the JSON is valid. If you’re not, it has trailing commas or comments that break standard parsers.

We all recognize AI writing when we see it. As language models get used for everything, their specific tics are everywhere. People call this “AI slop”. The grammar is fine, but the text is boring, repetitive, and lacks any real voice.

A site called tropes.fyi tracks these patterns. It lists the exact words and structures that give AI away. You’ll see callouts for overused adverbs like “deeply” and “arguably,” the dreaded “delve” family, and structural crutches like “It’s not X, it’s Y.”

Recently, I’ve been exploring the Model Context Protocol (MCP) and how to easily create custom servers to extend the capabilities of AI assistants like Gemini CLI which I use daily.

I wanted a way to build these servers in Java without the heavy boilerplate of a traditional Maven or Gradle project, or with a complex framework. The solution? Combining JBang, LangChain4j, and… 🥁… a custom Gemini CLI skill!

In this post, I’ll walk you through how I streamlined the creation of MCP STDIO servers, by creating an agent SKILL.md to replicate what I had learned in my previous article.

By now, you’re certainly all familiar with the Model Context Protocol (MCP)? It’s the standard for connecting Large Language Models (LLMs) to tools and data. But if you look at the current ecosystem, you’ll see a lot of Python and TypeScript…

As a Java developer, you might be wondering: How can I easily and quickly run my own MCP servers?

On this blog, I’ve explained how to develop MCP servers with Quarkus and Micronaut. But thanks to a recent community contribution to LangChain4j, and the simplicity of JBang, building a local MCP server in Java is even easier and with zero boilerplate.

Today, let’s come back to one of my favorite generative AI topics: Retrieval Augmented Generation, or RAG for short.

In RAG, the quality of your generation (when an LLM crafts its answer based on search results) is only as good as your retrieval (the actually retrieved search results).

While vector search (semantic) and keyword search (BM25) each have their strengths, combining them often yields the best results. That’s what we often call Hybrid Search: combining two search techniques or the results of different searches with slight variations.

If you’ve recently vibe-coded and exported a Gemini-powered app from Google AI Studio to host it online on Google Cloud Run, you might have noticed a server/ directory containing a Node.js application. This isn’t just a simple file server; it’s a clever “transparent proxy” designed to solve a classic problem in frontend AI development:

How do I use my API key without leaking it to the browser?

In this post (although vibe-coding is supposed to be all about not looking at the code at all) we’ll dissect exactly how this architecture works, why it’s safer than a client-side key, and where its security limits lie.