5th Devlog
Another 15 min devlog, what else

Things i did:

1. Claude Code finally recognises the custom models from NVIDIA as worthy enough to be used (aka they need a claude/${model.id} prefix, else the diva doesnt even look at it)
2. I began adding my newest project, the /v1/messages API and I alr know, that this is gonna be the biggest headache ever for a JS noob like me because the only communication there is JSON and AI stuff like role: user, but I’ll also have to implement that the return of a tool call is seen as role: toolcall and not user, else the model is gonna think the user smoked something… IDK alr if it will really be that bad but I dont wanna do it aaaa

• messages API is just console.log the POST request atm as i need to find out how claude code does that first but claude code doesnt even get to that and returns a Payload too large error or sum like that, screenshot is just a curl request.

4th devlog (?)

this devlog features somewhat 4 commits of struggling with the Anthropic API Scheme (i’d love to curse the engineer there that decided aahhh let’s just build a whole new format cuz we dont like openai).
So yeah, 2 files added and a few edited:

• main.js -> This is gonna be used later to launch the entire project at once, cuz server.js is only the express server and I want it to run in the background and adding a .bat script to start everything felt a bit weird…

• start-cc.js -> Who would’ve guessed it: It starts Claude Code with the required Environment variables (bet i wrote that wrong), but atm CC is just too tuff to accept any models the API delivers and just straight up continues advertising it’s own models without even trying to load others, even with the CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY: ‘1’ flag whyever that thing even exists… the flag might as well just do nothing, GPT told me it tell CC to request localhost:6967/v1/models (guess what it did: NOTHING yippie)

in server.js I added a translator for the Anthropic <-> OpenAI model API (with help of AI because JSON and JS at the same time blow my mind)
Also added some debugging console.logs, they aren’t as effective as i hoped, but lets see.

³rd Devlog (or mayb i have alzheimer, idk)

We finally have summer break in Bavaria/Germany!; so I’ll be working more in the next few weeks :). Today, I added a new file called check-cc.js and all it does is to check, if Claude Code is already installed or needs to be installed (also with a tiny installer coded in, haven’t tested that yet), but I still need to check it on a PC without Claude Code installed… (wait, while writing this i reminded myself that i have a nest instance :D). The new syntax is more complicated as things like text stripping have to be used -> the default output of claude –version is {version} (Claude Code) and to run a version check, the ‘Claude Code’ at the end is not really optimal. Also, it’s my first time working with exec and error + stdout but it seems really useful so far. (and I’m having a headache because of JS) Yeah, thats basically it for today. Ah yeah, and i also need to implement the new file in the main process, thats also gonna be a headache

²nd Devlog i think

I havent worked on this for a while but I finally got some motivation to do so, this time I added: dotenv/.env -> you can now set NVIDIA_NIM_KEY= and /api-transl/nim.js actually loads it (was rocket science to find out how that works);
an API call (woah), it calls https://integrate.api.nvidia.com/v1/models and gets the models that NVIDIA has atm, at first i thought you also have to send Authorization bearers with the API but when i got a 200 answer with all models when i used 67 as key i found out it doesnt…mmmmmmm

yeah so I’ll probably have to read some documentations about claude code and how it accepts api calls… byebye

First devlog ig

so I regularly use FCC aka Free-Claude-Code and I noticed, that some models are not supported and things like Web-Search or Autopilot don’t work or only with limits, that’s when the idea came to make an own wrapper for CC which is made for NVIDIA NIM because they just have the best open source models with almost no limits beside a rate limit of 40RPM :). The wrapper hopefully will allow models like GLM-5.2 or Kimi to use websearch and run completely on Autopilot. I’ll use Node.js/Express server and i just created a tiny express server which answers on /health with a 200 code and /health-test with Server is healthy and allat runs on port 6967, probably because I’m just brainrotted and that were the first to numbers that came to my mind :D