DevoxxGenie
Devoxx Genie is a fully Java-based LLM Code Assistant plugin for IntelliJ IDEA, designed to integrate with local LLM providers such as Ollama, LMStudio, GPT4All, Llama.cpp, Nativ and Exo but also cloud based LLM’s such as OpenAI, Anthropic, Mistral, Groq, Gemini, DeepInfra, DeepSeek, Kimi, GLM, OpenRouter, OrcaRouter, Requesty, DemonRoute, Cloudflare AI Gateway, Azure OpenAI, Amazon Bedrock and NVIDIA
🆕 Skills — Drop portable SKILL.md files into ~/.devoxxgenie/skills/, <project>/.devoxxgenie/skills/, .claude/skills/ or .agents/skills/ and the LLM activates them mid-conversation when relevant. The same files work in Claude Code, Codex and Gemini — and Custom Prompts have been renamed to Commands to make the distinction clear.
🆕 Security Scanning — Run Gitleaks (secret detection), OpenGrep (SAST) and Trivy (dependency CVEs) directly from the LLM agent. Findings are automatically created as prioritised tasks in the Spec Browser for tracking and remediation!
🆕 Spec Driven Development (SDD) — Define tasks in Backlog.md, browse them in the Spec Browser with Task List and Kanban Board views, then let the Agent implement them autonomously! Use the Agent Loop to run multiple tasks in a single batch with dependency ordering and automatic advancement.
🆕 AI-powered Inline Code Completion — Get context-aware code suggestions as you type using Fill-in-the-Middle (FIM) models via Ollama or LM Studio!
🆕 ACP Runners — Communicate with external agents (Kimi, Gemini CLI, Kilocode, Claude Code, Copilot) via the Agent Communication Protocol (JSON-RPC 2.0 over stdin/stdout) with structured streaming, conversation history, and capability negotiation!
🆕 CLI Runners — CLI Runners let you execute prompts and spec tasks directly from DevoxxGenie’s chat interface or Spec Browser using external CLI tools like Claude Code, GitHub Copilot, Codex, Gemini CLI, and Kimi.
🆕 Plugin Integration API — Other IntelliJ plugins can integrate with DevoxxGenie at runtime via a reflection-based ExternalPromptService — no compile-time dependency required. Two real-world POCs show it in action: a SonarLint fork and a SpotBugs fork that each send code-quality findings to DevoxxGenie with a single click, or defer them as Backlog tasks for the SDD workflow.
With Agent Mode, MCPs and frontier models like Claude Opus 4.6, Gemini Pro, DevoxxGenie isn’t just another developer tool — it’s a glimpse into the future of agentic programming. One thing’s clear: we’re in the midst of a paradigm shift in AI-Augmented Programming (AAP) 🐒
We also support RAG-based prompt context based on your vectorized project files, diff-based review of every file an agent run changes, and LLM-driven web search with Google and Tavily.
114K+ Active Users
📚 Documentation
📖 Visit our comprehensive documentation at genie.devoxx.com
Quick links:
🔒 Security Scanning
📖 Full Security Scanning Documentation
DevoxxGenie integrates three best-in-class open-source security scanners as LLM agent tools. When Agent Mode is active, the LLM can invoke them on demand, interpret the results in context, and automatically create prioritised backlog tasks for every finding.
| Scanner |
What it detects |
Install |
| Gitleaks |
Hardcoded secrets, API keys, tokens |
brew install gitleaks |
| OpenGrep |
SAST issues — injection flaws, insecure patterns |
brew install opengrep |
| Trivy |
Dependency CVEs (SCA) |
brew install trivy |
Ask the agent: “Run a full security scan and create backlog tasks for everything you find.”
Enable in Settings → DevoxxGenie → Security Scanning. Each scanner has a path browser, a Test button, and install guidance. Findings are deduplicated — re-running a scan will not create duplicate tasks.
Spec Driven Development (SDD)
📖 Full SDD Documentation
Spec Driven Development brings structured task management directly into your IDE. Instead of ad-hoc prompts, define your tasks in Backlog.md files, browse them in the Spec Browser, and let the Agent implement them autonomously.
How it works:
- Create tasks — Use natural language prompts or write Backlog.md task specs manually
- Browse in Specs — View tasks in a Task List or visual Kanban Board with drag-and-drop
- Implement with Agent — Click “Implement with Agent” and let the AI do the work
The Kanban Board gives you a visual overview of task status with drag-and-drop support:
17 built-in backlog tools provide full CRUD operations on tasks, documents, and milestones — all accessible to the LLM agent for autonomous project management.
Agent Loop — Batch Task Execution
Select multiple tasks (or click “Run All To Do”) and the Agent Loop executes them sequentially in a single batch. Tasks are automatically sorted by dependencies using topological ordering, and each task gets a fresh conversation. The agent implements each task autonomously, and when it marks a task as Done the runner advances to the next one — with progress tracking and notifications throughout.
📖 Agent Loop Documentation
Agentic Programming with DevoxxGenie
Unlocking AI Coding Assistants: Real World Use Cases by Gunter Rotsaert
Building full-stack AI agents: From project generation to code execution by Stephan Janssen
🗂️ Video Tutorials:
Blog Posts:
Key Features:
- 🔒 Security Scanning (v0.9.17+): Run Gitleaks (secret detection), OpenGrep (SAST) and Trivy (SCA/CVEs) as LLM agent tools. Each finding is auto-created as a prioritised Backlog.md task. Enable in Settings → Security Scanning.
- 📋 Spec Driven Development (v0.9.7+): Define tasks in Backlog.md, browse them in the Spec Browser (Task List + Kanban Board), and let the Agent implement them. 17 built-in backlog tools for full CRUD on tasks, documents, and milestones. Use the Agent Loop to run multiple tasks in batch with dependency ordering (v0.9.8+).
- 🆕 ACP Runners (v0.9.10+): Communicate with external agents (Kimi, Gemini CLI, Kilocode, Claude Code, Copilot) via the Agent Communication Protocol with structured streaming, conversation history, and capability negotiation.
- 🔌 Plugin Integration API (v0.9.12+): Let other IntelliJ plugins send prompts or create Backlog tasks via a reflection-based
ExternalPromptService — no compile-time dependency required. Two POC integrations available: SonarLint DevoxxGenie and SpotBugs DevoxxGenie.
- 🖥️ CLI Runners (v0.9.9+): Execute prompts and spec tasks via external CLI tools (Claude Code, GitHub Copilot, Codex, Gemini CLI, Kimi) directly from the chat interface or the Spec Browser.
- ✨ Inline Code Completion: (v0.9.6+) AI-powered code suggestions as you type using Fill-in-the-Middle (FIM) models. Supports both Ollama and LM Studio with models like StarCoder2, Qwen2.5-Coder, and DeepSeek-Coder.
- 🤖 Agent Mode (v0.9.4+): Autonomous code exploration and modification with built-in tools (read, write, edit, search files). Parallel sub-agents investigate multiple areas of your codebase concurrently, each with configurable provider/model. Enable in Agent Settings!
- 🛡️ Command Blacklist (v1.10.0+): Tell the agent which shell commands it may never run unsupervised. Destructive commands like
git reset --hard or rm -rf are caught even inside compound commands (cd x && git reset --hard) and with reordered flags (-rf also matches -fr), and either force the approval dialog or are blocked outright — even when write approvals are auto-approved.
- 🍎 Nativ: Run MLX models locally on Apple Silicon through Nativ’s OpenAI-compatible server. Enable it under Settings → LLM Providers → Local and DevoxxGenie discovers every MLX model you have downloaded. Note that Nativ listens on port
8080 by default — the same port as Llama.c++ — so change one of the two if you want both enabled.
- 🖥️ Local AI Cluster: Pool multiple Apple Silicon devices into a single LLM inference cluster with Exo. Run models like Llama 3.1 405B across your machines with automatic device discovery, zero cloud costs, and full data privacy.
- ☁️ Cloudflare AI Gateway (v1.10.0+): Point DevoxxGenie at your Cloudflare AI Gateway and use every model behind it with a single Cloudflare token (BYOK) — your provider keys stay in Cloudflare, and you get caching, rate limiting, spend controls and per-request analytics for free.
- 🔥️ MCP Support with Marketplace: Browse and install MCP servers from the integrated marketplace. Add MCP servers and use them in your conversations!
- 🗂️ DEVOXXGENIE.md: By incorporating this into the system prompt, the LLM will gain a deeper understanding of your project and provide more relevant responses.
- 📸 DnD images: You can now DnD images with multimodal LLM’s.
- 🧐 RAG Support: Retrieval-Augmented Generation (RAG) support for automatically incorporating project context into your prompts.
- 👀 Chat History: Your chats are stored locally, allowing you to easily restore them in the future.
- 🧠 Project Scanner: Add source code (full project or by package) to prompt context when using Anthropic, OpenAI or Gemini.
- 💰 Token Cost Calculator: Calculate the cost when using Cloud LLM providers.
- 🔍 Web Search: Search the web for a given query using Google or Tavily.
- 🏎️ Streaming responses: See each token as it’s received from the LLM in real-time.
- 🧐 Abstract Syntax Tree (AST) context: Automatically include parent class and class/field references in the prompt for better code analysis.
- 💬 Chat Memory Size: Set the size of your chat memory, by default its set to a total of 10 messages (system + user & AI msgs).
- ☕️ 100% Java: An IDEA plugin using local and cloud based LLM models. Fully developed in Java using Langchain4J
- 👀 Code Highlighting: Supports highlighting of code blocks.
- 💬 Chat conversations: Supports chat conversations with configurable memory size.
- 📁 Add files & code snippets to context: You can add open files to the chat window context for producing better answers or code snippets if you want to have a super focused window
Start in 5 Minutes with local LLM
- Download and start Ollama
- Open terminal and download a model using command “ollama run llama3.2”
- Start your IDEA and go to plugins > Marketplace and enter “Devoxx”
- Select “DevoxxGenie” and install plugin
- In the DevoxxGenie window select Ollama and available model
- Start prompting
Start in 2 Minutes using Cloud LLM
- Start your IDEA and go to plugins > Marketplace and enter “Devoxx”
- Select “DevoxxGenie” and install plugin
- Click on DevoxxGenie cog (settings) icon, open the Cloud Direct tab (or Routers for a router) and click on the provider link icon to create API KEY
- Paste API Key in Settings panel. On the Routers tab, also tick the router’s checkbox
- In the DevoxxGenie window select your cloud provider and model
- Start prompting
🗂️ Model Context Protocol servers support
📖 Full MCP Documentation
Initial support for Model Context Protocol (MCP) server tools including debugging of MCP requests & responses!
MCP support is a crucial feature towards ful Agentic support within DevoxxGenie.
Watch short demo of MCP in action using DevoxxGenie
Example of the Filesystem-server MCP which allows you to interact with the given directory.
Go to the DevoxxGenie settings to enable and add your MCP servers. Browse the MCP Marketplace to discover and install servers with just a few clicks!
When configured correctly you can see the tools that the MCP brings to your LLM conversations
Agentic Magic in action 👀✨🧠
📖 DEVOXXGENIE.md Documentation
You can now generate a DEVOXXGENIE.md file directly from the “Prompts” plugin settings page or just use /init in the prompt input field.
By incorporating this into the system prompt, the LLM will gain a deeper understanding of your project and provide more relevant responses.
This is a first step toward enabling agentic AI features for DevoxxGenie 🔥
Once generated, you can edit the DEVOXXGENIE.md file and add more details about your project as needed.
📸 “I can see” DnD images
You can now drag and drop images (and project files) directly into the input field when working with multimodal LLMs like Google Gemini, Anthropic Claude, ChatGPT 4.x,
or even local models such as LLaVA
You can even combine screenshots together with some code and then ask related questions!
🔥 RAG Feature
📖 Full RAG Documentation
Devoxx Genie now includes starting from v0.4.0 a Retrieval-Augmented Generation (RAG) feature, which enables advanced code search and retrieval capabilities.
This feature uses a combination of natural language processing (NLP) and machine learning algorithms to analyze code snippets and identify relevant results based on their semantic meaning.
With RAG, you can:
- Search for code snippets using natural language queries
- Retrieve relevant code examples that match your query’s intent
- Explore related concepts and ideas in the codebase
We currently use Ollama and Nomic Text embedding to generates vector representations of your project files.
These embedding vectors are then stored in a Chroma DB (v0.6.2) running locally within Docker.
The vectors are used to compute similarity scores between search queries and your code all running locally.
The RAG feature is a significant enhancement to Devoxx Genie’s code search capabilities, enabling developers to quickly find relevant code examples and accelerate their coding workflow.
See also Demo
Expecting to add also GraphRAG in the near future.
LLM Settings
In the IDEA settings you can modify the REST endpoints and the LLM parameters. Make sure to press enter and apply to save your changes.
We now also support Cloud based LLMs, you can paste the API keys on the Settings page. The LLM providers are grouped in four tabs:
- Local: providers running on your machine, such as Ollama and LM Studio.
- Cloud Direct: cloud providers you have your own account with, such as OpenAI, Anthropic and Google. Prompts go straight to that provider.
- Routers: gateways that reach many models through one API key: OpenRouter, Cloudflare AI Gateway, OrcaRouter, Requesty and DemonRoute. A router only appears in the chat’s provider list once you tick its checkbox.
- Custom OpenAI: any other OpenAI-compatible endpoint.
Routers forward your prompts and code to other providers, each with its own logging, retention and training policies. Check a router’s terms before enabling it, and see the Routers documentation.
Smart Model Selection and Cost Estimation
The language model dropdown is not just a list anymore, it’s your compass for smart model selection.
See available context window sizes for each cloud model
View associated costs upfront
Make data-driven decisions on which model to use for your project
Add Project to prompt & clipboard
You can now add the full project to your prompt IF your selected cloud LLM has a big enough window context.
Calc Cost
Leverage the prompt cost calculator for precise budget management. Get real-time updates on how much of the context window you’re using.
See the input/output costs and window context per Cloud LLM. Eventually we’ll also allow you to edit these values.
Handling Massive Projects?
“But wait,” you might say, “my project is HUGE!” 😅
Fear not! We’ve got options:
- Leverage Gemini’s Massive Context:
Gemini’s colossal 1 million token window isn’t just big, it’s massive. We’re talking about the capacity to digest approximately 30,000 lines of code in a single go. That’s enough to digest most codebases whole, from the tiniest scripts to some decent projects.
But if that’s not enough you have more options…
- Smart Filtering:
The new “Copy Project” panel lets you:
Exclude specific directories
Filter by file extensions
Remove JavaDocs to slim down your context
- Selective Inclusion
Right-click to add only the most relevant parts of your project to the context.
The Power of Full Context: A Real-World Example
The DevoxxGenie project itself, at about 70K tokens, fits comfortably within most high-end LLM context windows.
This allows for incredibly nuanced interactions – we’re talking advanced queries and feature requests that leave tools like GitHub Copilot scratching their virtual heads!
Support for JLama & LLama3.java
DevoxxGenie now also supports the 100% Modern Java LLM inference engines: JLama.
JLama offers a REST API compatible with the widely-used OpenAI API. Use the Custom OpenAI URL to connect.
You can also integrate it seamlessly with Llama3.java but using the Spring Boot OpenAI API wrapper coupled with the JLama DevoxxGenie option.
Local AI Cluster with Exo
📖 Full Exo Documentation
Pool the compute power of multiple Apple Silicon devices into a single, unified LLM inference cluster using Exo. Run large models like Llama 3.1 405B that don’t fit on a single machine — split across your MacBook, Mac Studio, and Mac Mini over Thunderbolt or network.
- Automatic Device Discovery: Exo detects nearby devices automatically and distributes model layers using pipeline or tensor parallelism — no manual configuration needed.
- OpenAI-Compatible API: DevoxxGenie connects to your Exo cluster just like any other local provider. Run powerful models entirely on your own hardware with zero cloud costs and full data privacy.
- Supported Models: Llama 3.x (8B–405B), DeepSeek, Qwen, Mistral, and more — see the full list.
Test Driven Generation (TDG) - Experimental
Write a unit test and let DevoxxGenie generated the implementation for that unit test.
This approach was explained by Bouke Nijhuis in his Devoxx Belgium presentation
An demo on how to accomplish this can be seen in this 𝕏 post.
DeepSeek R1 & DevoxxGenie 🔥
As of today (February 2, 2025), alongside the DeepSeek API Key, you can access the full 671B model for FREE using either Nvidia or Chutes!
Simply update the Custom OpenAI URL, Model and API Key on the Settings page as follows:
Chutes URL : https://chutes-deepseek-ai-deepseek-r1.chutes.ai/v1/
Nvidia URL : https://integrate.api.nvidia.com/v1
Grok & DevoxxGenie
Create an account on Grok and generated an API Key. Now open the DevoxxGenie settings and enter the OpenAI compliant URL for Grok, the model you want to use and your API Key.
Installation:
📖 Full Installation Guide
- From IntelliJ IDEA: Go to
Settings -> Plugins -> Marketplace -> Enter ‘Devoxx’ to find plugin OR Install plugin from Disk
- From Source Code: Clone the repository, build the plugin using
./gradlew buildPlugin, and install the plugin from the build/distributions directory and select file ‘DevoxxGenie-X.Y.Z.zip’
Requirements:
- IntelliJ minimum version is 2023.3.4
- Java minimum version is JDK 17
Build
Gradle IntelliJ Plugin prepares a ZIP archive when running the buildPlugin task.
You’ll find it in the build/distributions/ directory
./gradlew buildPlugin
Testing on Different IntelliJ Versions
You can test the plugin on multiple IntelliJ IDEA versions to ensure compatibility:
# Test on IntelliJ 2024.3 (default baseline)
./gradlew runIde
# Test on IntelliJ 2025.1
./gradlew runIde -PideVersion=2025.1.1
# Test on IntelliJ 2025.2
./gradlew runIde -PideVersion=2025.2.2
# Test on IntelliJ 2025.3 (latest stable)
./gradlew runIde -PideVersion=2025.3.3
Note: First run on a new version will download the IDE (~500MB-1GB), which may take several minutes.
Automated Compatibility Verification
Run the automated plugin verification against multiple IDE versions:
./gradlew verifyPlugin
This checks API compatibility without launching the full IDE.
Publish plugin
It is recommended to use the publishPlugin task for releasing the plugin
./gradlew publishPlugin
Usage:
- Select an LLM provider from the DevoxxGenie panel (right corner)
- Select some code
- Enter shortcode command review, explain, generate unit tests of the selected code or enter a custom prompt.
Enjoy!
Privacy & Anonymous Usage Analytics
To guide which LLM providers and models receive engineering investment, DevoxxGenie collects anonymous usage data when you run a prompt or change models.
What is sent:
- An anonymous install ID (UUID), generated once and stored locally
- A per-launch session ID (random 10-digit number)
- Plugin version and IDE version
- LLM provider name (e.g.
anthropic, ollama)
- LLM model name (e.g.
claude-3-5-sonnet)
What is never sent:
- Prompt text, response text, conversation history
- File content, file paths, project name, git remote
- API keys, credentials, user name, email
- Token counts or cost data
A first-launch notification asks for your consent before any data is sent. You can change this at any time in Settings → DevoxxGenie → General.
Contribute
📖 Contributing Guide
Understanding the Prompt Flow
The DevoxxGenie IDEA Plugin processes user prompts through the following steps:
2️⃣ Processing the Prompt
3️⃣ LLM Model Inference
4️⃣ Response Handling
-
If streaming is enabled:
-
If non-streaming:
5️⃣ Enhancements (RAG)
RAG (Retrieval-Augmented Generation)
6️⃣ Final Display
- The response is rendered in
ChatResponsePanel with:
Understanding the Flow
Below is a detailed flow diagram illustrating this workflow:
How to Get Started
Want to contribute? Submit a PR! 🚀
![dependabot[bot]](https://avatars.githubusercontent.com/in/29110?v=4)
评论