How could I most easily setup an external Code Completion Server (NOT agentic and NOT chat) inside IntelliJ?
Currently I'm using GH copilot for chat completition. I've disabled agentic features (because they don't work very well, cost a fortune and I just simply feel like I'm much more effective writing my own code) and simple use the gpt 4.1 based code completition (which struggles with java due to the way imports work). I simply just want basic, inline next token prediction.
Now I'd like to use my old gaming computer as a simple external inline code completion server (dual 3090, should be plenty), running something Codestral or Qwen to provide my own inference for code completion, probably running something like llama.cpp as an inference server.
So basically the setup would be:
- My server, running in my network, serving a language model
- IntelliJ connecting to the server and using it for code completion ONLY, using a plugin
- no cloud services, subscriptions, telemetry or agentic features
- All data and code staying inside my network, without a subscirption or account being necessary
Now there used to be tools like tabby, tabnine and continue.dev, but they've all pivoted to agent and some even require a subscription now. Almost all of them are essentially wrappers for claude and openAI. But I want to run code completion, not agents. I don't want terminal integrations or code review or them running random code on my machine.
Ideally it'd be open source aswell
Are there other people with setups like this, and if yes what are they using?
What are my options here?
Please sign in to leave a comment.
Yes—there are definitely people running setups like this. The most common approach today is self‑hosting open‑source inference servers with models like Codestral, Qwen‑Coder, or StarCoder2, usually via llama.cpp, Ollama, or vLLM for efficient GPU serving. For IDE integration, options include Continue.dev (still open source, can be pointed at a local server), Cody (Sourcegraph, but heavier), or custom IntelliJ plugins that connect to an OpenAI‑style API endpoint exposed by your server. Many developers NC toll login simply run Ollama or llama.cpp with a REST API and configure Continue or a lightweight plugin to use it for inline completions—keeping everything local, subscription‑free, and agent‑free.
It looks like Continue.dev has ceased development, but I'll look into Cody. Thank you for the recommendations. I wish IntelliJ just let you use custom model servers with the built in code completion plugin thing