[HN Gopher] Show HN: Use Third Party LLM API in JetBrains AI Ass...
       ___________________________________________________________________
        
       Show HN: Use Third Party LLM API in JetBrains AI Assistant
        
       Author : Stream
       Score  : 75 points
       Date   : 2025-05-03 11:52 UTC (11 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | yekanchi wrote:
       | does not gpt4free [g4all] do the same?
        
         | Stream wrote:
         | The problem is that JetBrains AI Assistant only support Ollama
         | and LM Studio as model provider for now.
        
       | aghilmort wrote:
       | 1. this is cool / innovative
       | 
       | 2. had lots of fun with similar we should chain / tunnel this to
       | this it can't be done easily projects
       | 
       | 3. fond memories reverse engineering JNI 1.0 to let Java only
       | devs use C libs, get C# 1.0 to pipe via C++ to call C serial port
       | driver for early drone microcontroller etc
       | 
       | all to say / nice work / keep building / like that it solves your
       | very specific AI anywhere problem in JetBrains even tho it seems
       | like wait where to even start! It can't be done! Like Microsoft
       | once told me re: C# 1.0
        
       | ulrikrasmussen wrote:
       | I tried using the AI assistant when it came out but seemingly was
       | too stupid to figure out how to use it correctly. I tried to get
       | it to write single functions or short blocks of code for me, but
       | it would always start rewriting the whole file from scratch which
       | was way too slow.
       | 
       | Has it gotten better since, or did I just do it wrong from the
       | beginning?
        
         | glzone1 wrote:
         | The jetbrains ai assistant was truly terrible when it came out.
         | They blocked the use of the major LLM's and had supposedly a
         | better European (?) LLM for coding, but I could not get it to
         | work in any reasonable way. Now though they do allow you to use
         | the major LLMs I believe and so it's immediately much more
         | useful. There is also Junie, which I'm not sure how that
         | compares it's a bit confusing but also seems better.
        
         | rickette wrote:
         | It has gotten a lot better IMHO in the latest 2025 release.
         | Ability to enable/disable per project. Ability to favor local
         | models over online models (online as fallback), support for
         | OpenAI, Claude, Gemini, etc. Beter integration overall.
        
         | ta988 wrote:
         | I have used Junie their new system quite a lot recently it is
         | much much better. It solved problems that Claude code was
         | struggling with (but couldn't solve some that Claude code did
         | solve with max thinking).
         | 
         | It is still slowish, but the quality is definitely getting
         | there.
        
         | surgical_fire wrote:
         | I have started to use it recently, and I am finding it useful
         | as a code assistant.
         | 
         | 1) It has been doing a good job reviewing my code, and could
         | suggest some interesting performance improvements in the past.
         | 
         | 2) It is good to generate things such as simple REST endooints,
         | that sort of thing. Saves you a bunch of typing, even if you
         | have to do some finishing touches yourself.
         | 
         | 3) I think it generally does a decent job writing tests as
         | well.
         | 
         | 4) Current models have gotten really good at helping me explore
         | libraries I am not very familiar with.
         | 
         | The AI hype may be very annoying, and I am skeptical that in
         | the long run it makes financial sense, but it is a productivity
         | boost once you get the hang of it.
        
         | jasonjmcghee wrote:
         | I'm not a user, but I have accidentally hit cmd+/ and I believe
         | that is what you're looking for.
         | 
         | It pops up a little input inline, rather than a chat or
         | autocomplete.
        
       | roandepoan wrote:
       | There's also https://github.com/xsharov/enchanted-ollama-
       | openrouter-proxy
        
       | omneity wrote:
       | Congrats for shipping!
       | 
       | Have you seen LiteLLM Gateway before? As a bonus it supports a
       | lot of different providers.
       | 
       | https://docs.litellm.ai/docs/simple_proxy
        
         | Stream wrote:
         | Yes. Because my original purpose is to use third party LLMs in
         | JetBrains AI Assistant, it doesn't provide Ollama or LM Studio
         | endpoints as required. So I decide to develop one on myself.
        
       | bambax wrote:
       | Neat!
       | 
       | But instead of hitting each provider's API separately, managing
       | their keys, etc., it may be easier to go through OpenRouter,
       | which is a single endpoint, has hundreds of models, and doesn't
       | add any cost on top of public prices from the providers.
        
         | Lorean1 wrote:
         | If they don't add any cost I wonder what exactly is their
         | business model?
        
           | joshstrange wrote:
           | They charge you a service fee (5% + $0.35) when you reload
           | credits. This is very well hidden and not publicly stated (or
           | at least I couldn't find it would logging in and trying to
           | purchase).
        
           | bambax wrote:
           | They say they buy in bulk and get better prices than those
           | available to the general public... IDK if that's the whole
           | story, but could be.
        
         | joshstrange wrote:
         | It adds 5% + $0.35 every time you reload, so it absolutely does
         | add a cost on top of public prices.
        
           | tecleandor wrote:
           | What does 'reload' mean in this context? (haven't used
           | OpenRouter)
        
             | joshstrange wrote:
             | You buy credits on OpenRouter, or rather you load money
             | onto your account and then as you make API calls your
             | account balance is deducted by the cost displayed on
             | OpenRouter's website (market rates).
             | 
             | So once money is loaded on your account you pay the same as
             | if you used the LLM provider directly but you pay a premium
             | to load money onto your account, it's effectively adding a
             | 5% fee to the base LLM API prices (and $0.35/per reload,
             | depending on reload amount this might be negligible but on
             | a $10 reload it means you are effectively paying a 9% fee).
        
       | nullchan wrote:
       | I literally was working on something similar. Emphasis on `was`
       | since you beat me to it! :D
        
       | bionhoward wrote:
       | One thing I thought was funny was remembering these AI services
       | all have the customer noncompete clause and this morning
       | realizing they can actually sue you (not just rug pull you,
       | actually bankrupt you with legal fees) if you use em to develop
       | competing stuff
       | 
       | Does that imply using OpenAI/Anthropic/Gemini API/Grok/Copilot
       | for anything but the most trivial jokish personal uses ought to
       | be considered a serious breach of fiduciary duty because it
       | exposes your business to potentially massive legal costs?
        
         | ta988 wrote:
         | Can you direct me to where you see that in the Jetbrains
         | agreements? Same for the other services.
        
       | NoelJacob wrote:
       | Wouldn't an extension be more suitable that runs on Jetbrains
       | boot?
        
       | linotype wrote:
       | JetBrains really needs to support alternatives fast. Cursor,
       | CoPilot, etc all need much better integration. VS Code is
       | becoming almost mandatory for companies that force the use of LLM
       | tools.
        
         | ivape wrote:
         | Really? This stuff is mandatory for those who are dependent on
         | it. Many people still just look at the code in a separate
         | browser tab and make a meaningful decision of if they want to
         | pull something in or not. Why even use an IDE is the better
         | question. What exactly is an IDE for if the AI will search and
         | determine where to add/remove code? I don't think Jetbrains
         | needs to lose focus on their product to compete with products
         | that are masquerading as an IDE but have no intention of
         | fulfilling the role of an IDE. They want to re-invent it, so
         | it's better they simply stay in their own lane as they will be
         | cannibalized by their own genre as there are no rules in that
         | wild-west. For example, why does my automatic code writing
         | genie look like a fucking IDE? Is there a reason? Cursor and
         | the like will have their own competition and they'll struggle
         | to keep up with a blank screen that just magically outputs
         | software with no visibility into the code. Then they'll be
         | stuck trying to sell people why it's better to let magic be
         | written inside of an IDE, both competing with new magic and old
         | hard tools (like Jetbrains). Cursor is not the future, it's
         | dead on arrival.
         | 
         | Jetbrains and existing IDEs are much more focused for exactly
         | what it is, manual code writing/inspection. How they integrate
         | AI should be done meaningfully to improve the IDE experience.
         | If they want to be in the business of reinventing code-writing,
         | then you have to scrap the IDE altogether.
        
           | kurtis_reed wrote:
           | Huh?
        
           | sillysaurusx wrote:
           | That seems like a false dichotomy. AI doesn't decide what
           | kind of code to write. That's still up to the human (for
           | now). Jetbrains also has to have some kind of AI offering, if
           | only because otherwise users will migrate to competitors that
           | do.
           | 
           | It's an interesting question whether we'll need an IDE at all
           | in the future. I used to program in Vim with minimal help.
           | Maybe one day I'll program in Vim with maximal help from an
           | AI. It would be nice to say "find where authentication
           | happens" and it pulls up all instances of auth code. But
           | that's also another example of a feature that can go into an
           | IDE: case sensitive search, regex search, natural language
           | search. So I'm skeptical we won't be using IDEs in the
           | future. And if all of this is a part of an IDE, then that's
           | squarely in Jetbrains' wheelhouse.
        
           | linotype wrote:
           | Some companies are mandating the use of Cursor or other tools
           | like it.
        
         | KronisLV wrote:
         | > JetBrains really needs to support alternatives fast.
         | 
         | They do support offline models:
         | https://www.jetbrains.com/help/ai-assistant/switching-to-off...
         | 
         | I also used the Continue.dev plugin for a bit which worked
         | nicely with Ollama but was buggy when it comes to the UI, their
         | AI Assistant feels better in that regard:
         | https://plugins.jetbrains.com/plugin/22707-continue
         | 
         | Sadly seems like their Junie tool doesn't support that yet,
         | would be pretty cool to have that convenience instead of
         | needing something like Aider or other tools for more complex
         | multi-step edits: https://www.jetbrains.com/junie/
        
       | christkv wrote:
       | What's considered the best small local model for coding?
       | Something that can run on an m1 max 32gb ram?
        
       ___________________________________________________________________
       (page generated 2025-05-03 23:01 UTC)