Kounaisha's Ultimate OpenRouter Proxy Guide
Take a look at my other guides too:
How to install and use SillyTavern on Windows - Jailbreak list for LLMs - How to use Proxies on janitorAI using ChutesAI
Hello! I am kounaisha and in this guide I'll teach you guys how to use free proxies on JanitorAI by using OpenRouter.
As you may already know, JLLM has its slight problems, besides having a very short memory (9001 tokens). Thanks to our beloved greedy fucking ass Sam Altman, OpenAI is not open source (how ironic) and you need to pay to use their API. However, there are many other models like META's Llama models that can be used with third-party APIs.
In this guide, I'll teach you how to configure the new AI that was released and that made OpenAI cry: Deepseek. There are several different models besides Deepseek that are very good too and I will bring them here eventually. All models in this guide (so far) have 128k memory tokens, making them perfect models for those who want to build a long adventure with their bots.
ㅤ
Instructions
1 - Make an account on OpenRouter (https://openrouter.ai/) by clicking the "Sign In" button in the top-right.
2 - On the drop-down menu with your profile image on it, click on the "Keys" button, and once you're at the keys page, click "Create Key".
3 - In the key creation window, you will have the option to enter a name and credit limit. You can enter whatever name you prefer, while the value is the amount of credits you want to use for that key. Since our model is free, no credits will be used, so leave it blank or put a random value if you want.
4 - Once you create the key, the key will pop up. It should start with sk-. Click on the clipboard button to copy it, then go to JanitorAI
5 - Enter any chat of any bot you want to chat with so you have access to the API settings in the top right corner.
ㅤㅤㅤㅤㅤㅤㅤㅤ
6 - With the API settings window open, you will click on the "Proxy" tab. The window should look like this:
ㅤㅤㅤㅤㅤㅤㅤㅤ
7 - As you can see, we need to configure three things: the model, the proxy URL and the API key. If you followed the guide correctly, you have your API key in your clipboard, so just go to the API key text field and paste your key there.
8 - In the "Other API/Proxy URL" field, put the following link: https://openrouter.ai/api/v1/chat/completions
9 - In the model section, you will select the "Custom" option and place in the text field one of the models below:
deepseek/deepseek-chat:free (DeepSeek V3 model, the latest verson... at least while i'm writing this)
deepseek/deepseek-r1:free (DeepSeek R1 model, The first version released)
deepseek/deepseek-r1-distill-llama-70b:free (DeepSeek R1 bathed in some llama instructions. doesn't change much regarding roleplays.)
-- 02/12/2025--
nvidia/llama-3.1-nemotron-70b-instruct:free (Nvidia Llama model)
mistralai/mistral-nemo:free (Mistral model)
-- 02/13/2025--
qwen/qwen2.5-vl-72b-instruct:free (Qwen model)
-- 03/28/2025
deepseek/deepseek-chat-v3-0324:free (New Deepseek V3 model)
In case you're in doubt of which one to get, Here's a list of pros and cons of each model:
| Model | Pos | Cons |
|---|---|---|
| DeepSeek V3 | The most up-to-date version (for now) of DeepSeek. Generates decent responses quickly and has high token memory (128k) | Even though It has 2 providers, the connection is very unstable; Need to wait a few seconds before refreshing a message. |
| DeepSeek R1 | The first version of deepseek. Generates very immersive responses. Has three connection providers that have proven (so far) to be quite stable. Has 128k memory tokens. | Deepseek R1 has a habit of simply forgetting that it is in a roleplay and generating normal responses; Need to wait a few seconds before refreshing a message. |
| Deepseerk R1 Distill Llama | Yes, the Deepseerk R1 Distill Llama is back, even with its problems. After all, the Deepseerk R1 is, among all the models on this list, the most immersive model, so here is another option in case the first one doesn't work (this one probably won't too). | The same problems as R1, forgets that he's in a roleplay, loses connection every time, has to wait a few seconds before refreshing a response... sigh... yeah, that's the price of being poor... |
| Nvidia Llama | Nvidia Llama caught my attention because of its unusual way of demonstrating expressions, such as changing the font size, making it bold for emphasis, etc. Really impressive. So far I can say that it's the most stable model among all, I haven't had any errors during my tests.. Has 131k memory tokens. | Sometimes, it can generate statistical responses, making lists and pointing out a lot of details that (perhaps) you find boring; Apparently, nvidia model doesn't work very well with janitor's jailbreak and sometimes doesn't generate NSFW responses. Refreshing sometimes solves the problem, but still, a negative point. |
| Mistral Nemo | A balanced model. It has proven to be stable as far as I could test, even with only 1 provider. The responses are quite natural, without technical details or anything like that. It has 128k memory tokens. | Researching further, I was able to discover that this model is not one of the most intelligent, so don't expect much creativity in dialogues. |
| Qwen 2.5 VL | Stable connection as far as I could test and does not suffer any type of problem with roleplay or NSFW messages. The messages are very natural and it seems to write more than the other models. Has 131k memory tokens. | The connection is stable but suffers constant pauses during response generation. This model can only generate 2k tokens per message, it is not a very worrying problem but you may encounter incomplete messages if the response is too long. You may prefer to set a max token limit with this one. |
| !NEW! DeepSeek V3 0324 | The successor to deepseek v3. It is smarter and more charismatic than the first one. I didn't have many problems with it. | This new version still doesn't escape the very rare cases where Deepseek simply forgets that he's in a roleplay, but it's still much better than R1 which has this problem ALL THE TIME. |
10 - Click on "Check API key/model" to verify that everything is correct and working. You will see a green popup from janitor saying that everything is ok.
11 - Custom prompt is recommended to be left as is. The current prompt is what janitor uses to break LLM limits. Just scroll down and click "save settings".
12 - Done! You're now chatting using proxy.
Bad and good news: Openrouter had a limit of 200 messages per day using the free models. A few days ago, that limit was reduced to 50. The "good news" is that if you deposit $10 into Openrouter, your daily messages go up to 1000 per day. You don't have to spend the $10, just leave it there.
Recommendations
The settings I personally use are:
- Temperature: 1. I had several problems like LLM repeating the same sentences all the time when the temperature was below 1. At the same time, it generates unreadable and very bizarre texts when it is above. LLM rarely presents problems at temperature 1 so that is the value I most recommend.
- Max tokens: 0. I've noticed that it's very rare for LLM to write more than 400 tokens, even with the limit set to 0. So if you'd rather make sure that LLM doesn't write too much in any given situation, I'd suggest setting the limit to 400 or less, depending on your taste, but it's possible that LLM will leave some text unfinished. If you're like me and don't care too much, leave it at 0.
- Context tokens: 128k tokens. As I said, the models that I recommend have a context memory of 128k tokens, but you can decrease this amount if you want in generation settings.
FAQ
- "I got a 'network error', how do I fix it?"
If you get a "network error", it probably means that the specific model provider is offline. However, be sure to refresh janitor's website and try again. if the error persists, try another model.
- "I get 'network error' even after testing all LLMs"
The network error can also occur when you paste some information in the wrong place. Your API settings should look exactly like the photo below:
ㅤㅤㅤㅤㅤㅤㅤㅤ
- "Which model works best for rp?"
The Deepseek R1 has proven to be the most immersive model of all so far. Unfortunately, the immersion is broken by its connection issues and its habit of forgetting that it is in a roleplay. That said, I recommend the Nvidia Llama and Mistral Nemo for stability because deepseek models are presenting problems very often...
- "Bot does not roleplay and only gives normal responses"
This problem seems to be present in some models like Deepseek R1, sometimes it juts forgets janitor's rules. Just refresh the bot's response and it will solve it, but I don't know how long the problem can persist. If it bothers you too much, I suggest changing it.
- "API key connects but I get random "unk" error when I try to chat"
This means that the LLM provider connection is unstable. This is a problem that seems to be common in Deepseek models. If you are having this problem, I suggest using Nvidia Llama or Mistral Nemo.
- "Confirmation popup does not appear when I test the connection"
This happens sometimes, especially with DeepSeek models. It is probably because DeepSeek always has response failures and this failure happened right at the time of confirmation. It does not mean that the model is not working, if you did not receive an error popup, it means that everything is fine.
- "Help! I'm getting a 'limit exceeded' error"
Unfortunately, OpenRouter imposes a limit of 200 50 messages that you can send per day. Welcome to capitalism... :c