An LLM-based plugin for the IntelliJ product family, focused on staying out of the way and more reliable when used with self-hosted models.
The principle behind this extension is that the LLM doesn't advertise itself, and doesn't show up unless explicitly invoked by the user.
Alt-G keybinding to trigger code generation (or replacement generation for a given selection).
Alt-Q keybinding to ask the model questions about the current context. This can be a single
file, or across multiple files if the plugin is configured so.
Not published on the plugin marketplace.
You can build your copy locally with ./gradlew buildPlugin, or download the extension from the GitHub release, and
use the "install from disk" facility within the IDE Plugins view.
You can find a video of the plugin in action at the end of this blog post.
Project was built as part of a month-long experiment with GPT-5.5. All vibecoded content is isolated to commits: 925ae9b3b2a6bcd9269b817448ac130b7e0a0d8b
All content referenced in those commits is public domain (CC0). Everything else is Apache 2.0 licensed, keeping in line with the upstream forked template.
With self-hosted models, and the limited compute capabilities, there are certain paths within the code to provide extra resiliency to model failure. For example the plugin will retry requests when the model fails to provide a response (no content, or only reasoning), or will extract code from markdown fenced responses.
During development, the extension was mostly tested with GPT OSS 20B, Gemma4 26B A4B, and Qwen3.6 35B A3B.
There is no implied support or further development guarantee for this project. Feel free to open up a discussion topic if you'd like to share your opinion or experience with the plugin.
For any kind of improvements to the project, you are free to fork the project.