In this guide bot, I’m gonna show you how to easily set up a proxy. I’ve included screenshots with instructions for every step to make it super simple.

I figure you probably got here via a link or whatever, but I’ll explain what a proxy is real quick (just in case): basically, it connects a third-party, "smarter" LLM (AI) to your chat - like ChatGPT or DeepSeek. Yeah, you can actually take the exact models used on ChatGPT.com or DeepSeek.com etc. and hook them up to Janitor using a proxy.

If you run into any errors (that aren't mentioned in the guide or seem like special cases), feel free to ask me in the comments. I'll do my best to help out.

Here's the guide:

Step 1: Head over to OpenRouter.AI and sign up

Step 2: Once you've logged in, click the Get API Key button.

Step 3: Click Create API Key.

Step 4: In the pop-up window, give your key a name. You don't need to touch the other fields—just hit Create.

Step 5: COPY your new API key from the pop-up window. Make sure to save it somewhere safe (like in your notes). It’s a good idea to write down which email address it's linked to, just for future reference.

Step 6: Go to JanitorAI and open a chat with any bot (make sure Proxy is allowed for that bot). Click the three lines in the top right corner, then select API Settings.

Step 7: Switch to the Proxy tab and click Add Configuration.

Step 8: Fill in the fields exactly as shown in the screenshot

1. Model: z-ai/glm-4.5-air:free (outdated now, so try a new one: openrouter/owl-alpha)

2. Proxy URL: https://openrouter.ai/api/v1/chat/completions

Where it says API Key (optional), paste the key you copied earlier.

Step 9: Click save on everything, then refresh the page. Make sure you have Proxy selected (not JanitorLLM).

Step 10: Go to Generation Settings.

Step 11: Set all the values exactly as shown in the screenshot. (I’ll explain what each setting does and how to tweak them in the Advanced Proxy Guide further down).

Step 12: Finally, send a message to the bot. If it replies, everything is working properly.

If you did get an error, here is an explanation of all the common errors and how to fix them (Thx to Sprout):

Alright, so if everything’s up and running (meaning the bot replied), you can just leave it as is and go start roleplaying right now. OR, you can KEEP READING for the ADVANCED PROXY LEVEL (don't worry, it’s still easy, just takes a little bit longer).

ADVANCED PROXY GUIDE.

As I mentioned in the common errors section, the free tier on OpenRouter caps out at 50 messages a day. To bump that limit up to 1000, you’d usually have to drop $10 USD. But, there are a bunch of free ways to get around that limit (check the guide further down).

Now, let’s talk about which model (LLM) you should pick.

Here’s a list of the best free models currently on OpenRouter: google/gemma-4-31b-it:free, openrouter/owl-alpha (owl alpha seems to work well with context, so feel free to turn on 128k), poolside/laguna-m.1:free, nvidia/nemotron-3-ultra-550b-a55b:free

The biggest ones (by parameter count): openrouter/owl-alpha, poolside/laguna-m.1:free, nvidia/nemotron-3-ultra-550b-a55b:free

The most stable ones: openrouter/owl-alpha, poolside/laguna-m.1:free, nvidia/nemotron-3-ultra-550b-a55b:free

In my opinion, the top models rn are DeepSeek R1 0528 and Kimi K2. However, they can be pretty unstable. Because of that, I recommend sticking with GLM 4.5 AIR. It’s super stable and has Reasoning enabled (basically, the model "thinks" before it answers—you can toggle the reasoning text display off in the settings if you want).

The main perk of a lot of these free OpenRouter models is the context size. You get roughly 128k tokens, which is more than enough for the model to remember tiny details in your RP, even if the chat goes on for 150-200 messages. It’ll remember stuff from the very first message and bring it up later.

Now, let's briefly go over the Generation Settings.

Simply put, every model has its own recommended TEMPERATURE. You can find guides online for specific temps for specific models if you really want to experiment with different proxy models.

Max New Tokens: For ALL REASONING models, this needs to be set to 0 (which means unlimited). Otherwise, the model won't be able to generate an answer. For other models, you can set it to whatever you want.

Context Size: I recommend setting this to 64k-128k.

Some models struggle with large context and work best at 32k. Also, models like DeepSeek absolutely NEED a good Custom Prompt; at high context sizes, they might ignore the prompt otherwise. If you are TOTALLY confident in your model (like Gemini 2.5 Pro or 3 Pro, though those need their own guide), you can go up to 128k.

For GLM 4.5 AIR, you can probably do 64k–128k, I think it can handle it (haven't tested it myself though).

CUSTOM PROMPT.

Ideally, every model needs a prompt tailored specifically to it, but a generic one works fine for most things and won't change too much. (Exception: DeepSeek models. They need a strict, clear custom prompt or they start hallucinating).

There are tons of forums and sites for custom prompts. Here’s my favorite: (link to Sprout prompts + link to another custom prompt. If you made it this far in the guide, you can figure it out—just copy the custom prompt and the recommended temperature for your LLM.

Step 1: Go to API Settings => Proxy and hit COPY on your proxy configuration.

Step 2: Open the copy and just swap the old API key with the new one linked to your new email.

That’s it. Now you have 100 free messages a day. When you run out of your 50 (20 now? i dont really know..) free messages on one config and get the error Rate limit exceeded: free-models-per-day, just switch to your second proxy config and you're good (don't forget to refresh the page).

You can keep doing this indefinitely until you run out of email addresses. That’s the whole guide.

Thanks again to everyone who used my guide! If you have any questions, ask me in the comments

-Updates log-

02/25/2026: OpenRouter removed all the DeepSeek models from the free tier. The ones I recommend using right now are: qwen/qwen3-235b-a22b-thinking-2507 ; z-ai/glm-4.5-air:free

11/06/2026: OpenRouter changed up their free models again... I tried to find some decent ones. I also added another link to a custom prompt.

19/06/202: Swapped out an old model and added a new one. Heads up, nex-agi/nex-n2-pro:free is switching to paid on June 22nd (yeah, figured... it was super censored). But all the models currently listed, I tested them myself and they all do NSFW fine, just not the outright banned stuff.