Would it be unreasonable to suggest that the Intel...

# random

[pambrose]

Would it be unreasonable to suggest that the IntelliJ MCP server should provide all the tools necessary for Claude Code (or any other coding agent) to work with Kotlin code and obviate the need for grep and the other Unix utilities? This would be a win for users, as their coding agents' performance and token burn would improve significantly, and JB would have a compelling reason for users to keep IntelliJ integral to their CC workflow. In my humble opinion, I think this should be a "hair on fire" priority for JB.

[Youssef Shoaib [MOD]]

Unix utilities are used very commonly, so LLMs are likely to know how to use them. Compare that to a bespoke MCP server

[pambrose]

Yes, they absolutely know how to use them. The question is: how much more efficient and capable would CC be if it had the code knowledge contained within IntelliJ? Is refactoring better done with Unix utilities or with IntelliJ's magic? IntelliJ has a treasure trove of information that users would benefit from knowing while in CC. If you make the information more accessible and enhance CC's capabilities, JB will have an additional answer to the question: "Do I still need an IDE?"

[Stefan Oltmann]

Good point. I’d actually prefer AI to use IntelliJ commands rather than a CLI tool, because executing commands on CLI is unsafe compared to using tools within IntelliJ.

[Stefan Oltmann]

Right now I'm using that brave mode, because I hate AI to stop all the time to ask for permission to execute a „ls -l .“ command. But I admit that I'm also I'm fear to become another victim of the „AI wiped my disk and all my backups“ stories. 👀

[Stefan Oltmann]

So giving AI the knowledge/skill to use safe built-in tooling and make the execution of CLI commands a rare execration sounds good to me.

[pambrose]

If you are worried about rogue actions, something like this might be of interest:
https://agent-safehouse.dev

[Stefan Oltmann]

Reducing token usage by accessing context IntelliJ already has rather than reading in all files again also sounds like a great idea.

[Stefan Oltmann]

So I fully agree that some API to access insights and execute tooling would be very beneficial.

[Stefan Oltmann]

Agent Safehouse will help me on Mac. I also develop & test on Windows. I wonder if there is also a solution.

[CLOVIS]

Note that IntelliJ does already provide some information to agents. It's just… not that much

https://www.jetbrains.com/help/idea/mcp-server.html#supported-tools

[CLOVIS]

it does include the 'rename_refactoring' for example

[Stefan Oltmann]

It's a start. I guess I need to invest more time figuring out how to tell local agents to actually use that information.
Could be made easier out of the box.

[CLOVIS]

The TL;DR: in the settings there is a 'MCP server' category that you can enable. After that, you can configure IntelliJ itself as an MCP server for whatever agent you use.

The JetBrains assistant and Junie already use it. When you install an agent from the ACP registry there's a tick to decide if it has access to it or not.

[pambrose]

Yes, Junie CLI does exactly what I was asking for. It runs everything through the MCP server and leverages all the static analysis goodness in IJ. I am not sure why CC can't do the same thing.

[CLOVIS]

I think Junie has some more hidden powers that aren't in the MCP server. For exemple, it immediately knows when code it wrote doesn't compile, whereas all other agents need to actually run the compilation

[Stefan Oltmann]

I haven’t yet found a reliable way to run a local model (on 16 GB VRAM).
They always seem to struggle with using the tooling (like using wrong commands).

I installed OpenCode from the ACP registry so I can use it as an agent in an AI assistant.
Running Qwen 3.5 35B a3b through LM Studio sort of works, but tool usage remains a constant struggle.

It would be great if someone who’s figured this out could put together a guide showing how to use MCP and all the other components to get this working properly. It seems like there isn’t any out-of-the-box solution.

[Stefan Oltmann]

Junie CLI seems to be cloud model usage only, if I read https://youtrack.jetbrains.com/projects/JUNIE/issues/JUNIE-47/Add-ability-to-use-Junie-locally correctly.

[Stefan Oltmann]

Maybe local models, even if trained for tool usage, are just too stupid to properly use the tools.
In AI assistant I often get the "invalid tool" warning, when the model doesn't respect the available tools and just tries to call tools that don't exist.