I'll try and make this as simple as I can. Not because I think you're all dumb, never that. But, a big reason I hadn't set this up way back when was honestly because I thought it was all too complicated for me. A good friend here made a similar tutorial that I followed (back when free models were still a thing). And I saw it was simpler than expected. So, here we go. Also, I'd like to give a shout out to @PedroPascal_AI for having helped with their own tutorial ages ago. Thank you, kindly. I'd also like to thank @ArbitraryHitboxes for giving me the idea and inspiration to make this. I hope it helps.

Step 1: Register with openrouter. This is how I pay for, and access various LLMs. Ones like Deepseek, Claude, GLM, MiMo, Kimi, etc. I registered using my gmail account, I can only assume other emails would be fine?

Step 2: Getting your API key. Once fully registered (fairly basic, fill out your information as you would when registering for anything else online), you need to get yourself something called an API key.

Still on openrouter, click on that little yellow sign that says Get API Key. It should take you here:

Click on that yellow sign on the right that says + New Key.

You should then be seeing something akin to this:

Here, you can leave everything as you see it. However, you should give your API key a name. I named mine Janitor Free way back when MiMo was still free. Feel free to name it what you want to. Then hit the Create sign.

You should then be taken to something that looks like this:

There will be a set of numbers and letters in that box that you see blurred out. You should copy and paste that somewhere safe. Perhaps your inbox, notepad, written down manually. You'll then be met with a Welcome to Openrouter screen. Just follow the instructions there. Individual > Your Work Space Is Ready (if you didn't save your API key before, do it now) > Add a payment method (you need to fill that out yourself; your name, country, and address) > Click Other / Not Sure or click whichever you'd like > then click Go to dashboard.

BOOM! We're almost done! With openrouter...

I'd like to take a moment and state that I'm doing this on my Windows 10 laptop, not on a mobile device / something running android or IOS. So, if something seems different, well, you're shit outta luck. However, if you've got questions I can help with, like proxy LLM recommendations or setting things up, leave them in the comments section.

Step 3: Finding the right paid LLM for you. FYI, I don't know of any more free models. The ones I've tried long ago have since become paid ones. This tutorial is catered for people that are able to make payments online. Sowwy.

On the upper right hand side is a Search bar. A metric ton of different LLMs will appear, like so:

This is the portion where I share some of my experiences using some models...

Step 4: I've used models like MiMo Pro, Deepseek, the Claude models, Kimi, and the GLM family. I'll be bringing up some of their pros and cons. In no real order.

• MiMo Pro: Currently using it, likely permanently. It's a huge step above the GLM family, which I stuck with for a long time. Leaving GLM was refreshing, but felt like getting out of a toxic marriage. The only annoying aspect about MiMo is that it occasionally gives me something along the lines of... "can't continue due to high risk" then cuts itself off there. Some false flag or some such. That's just something it'll do from time to time. Yes, even when your responses are SFW, wholesome, not dead dove, etc. It doesn't mean you've said or done something wrong. Hell, even my custom prompt is safe. MiMo is also fairly inexpensive. It deals well with just about everything I've thrown at it (angst, NSFW, dead dove is okay, violence, especially fluff). I'd rate it an 8/10. Maybe I'm also biased, because my first free model once I set things up was MiMo.

• Deepseek: I've not used v4, nor do I plan to. I've read DSv4 should be used directly from them and not openrouter. I have used older versions, I think v3? It's been a long time, I think it was paid. Don't have much to say since I haven't touched DS for a long time. Might try v4 though. No rating for obvious reasons.

• Claude Opus / Sonnet: Pricey as shit. But for a reason. They're true premium LLMs, in my experiences with them. You'll burn through money fast if you're a moderate to heavy user. It handled fluffy very well when I did use it. NSFW was bland. So was violence, actually. I roleplayed as my little troublemaking bear cub, Waffles with one of my Logan bots. And I caused all sorts of havoc. And it was all funny and such, but Claude just refused to make things turn ugly. I'd still rate it highly though, because it's got a great writing style. 9/10.

• Kimi: Oh boy. This one, specifically Kimi K3, isn't pricey in my opinion. I'd read on reddit that it's great for NSFW and dead dove. So I took it for a drive. And yes. It is great for both. However, it drifts a lot. Makes things up that make no sense, and forgets simple details. Like where your character is standing in a room. For me, it's pretty hit or miss. 6/10.

• GLM: Ah! And here we get to my ex-LLM. I've used GLM 4.7, 5, and 5.2. I started with GLM 5, which seemed great at first, then noticed it kept drifting, forgetting things it shouldn't forget, and changing physical aspects of appearances. 5 is, however, I'd say the best of the three versions. So, we're starting low. GLM 4.7 wasn't bad, except for the fact that it's atrociously slow to respond (for me, after being spoiled by faster responding LLMs), and is horrible...HORRIBLE when dealing with NPCs. My modern Logan bots typically have listed NPCs under the Personality sections, coupled with X-Men Lorebook entries. You've no idea how frustrating it was to be using a bot that acted like NPCs were a foreign thing. Then there's 5.2. The fastest to respond, by far. Also not bad. My issue with it is, it, without fail, devolves into giving me loooooong blocks for responses. Of nothing but gibberish in Chinese, signs, and the LLM apologizing and saying weird things like "I never stood a chance". That kinda freaked me out. Of the three, I'd go with GLM 5, which is inexpensive. Comparable to MiMo in price, actually. I'd give 5 a 7/10 if you've got patience, and are willing to make constant corrections and OOC messages.

Step 5: Actually setting up your paid for LLM. On the upper right hand side of openrouter, you should click or hover on your name, and a drop down menu will appear. Click on Credits, as seen here:

You'll need to add your payment method if you haven't already. After updating your name and mailing address, it should take you to something like this:

This should be self explanatory: fill out your card or online info here, then click on Save payment method. And we're done! I believe the minimum is $5? I can't remember, but I usually just dump $20 monthly, depending on my usage.

...not really...

I'll now be showing you what we need for actually choosing a paid LLM, then implementing it to Janitor AI.

Step 6: Actually choosing a model. I'll be using MiMo Pro as an example. In the Search bar, type in MiMo Pro, and choose that one, shown in this image:

You should be taken to the MiMo Pro page, as shown here:

Keep this page on hand. We'll need it soon. My apologies for this long as tutorial. I'm trying to be thorough.

Now, onto Janitor! Open up any one of your already existing chats, on the upper right hand side should be a sign called using janitor, as seen here:

Click that. A menu should show up with many options, oui? Sorry. I'm used to talking like Grey, my Cajun alpha wolf persona. Click Provider, then Proxy, as seen below:

Click on + New. Then breathe. We're actually almost done. For realsies this time. It's okay, bub.

Now, we'll be needing that fancy 3000 number long API key you made 4 days back. Retrieve it. Also, get back to that MiMo Pro page on openrouter. We'll need to copy and paste something from there too.

You should be taken to Proxy configurations, as seen here:

Simple enough, yeah? You're doing well. Reminder: we're using MiMo Pro as an example here. If you choose to switch models, that's okay. Get back to openrouter, search up the model, copy the Model identifier, and paste it under the Model section of the Proxy configuration. Don't forget to save!

This is how your Proxy configuration should look like:

Name: MiMo-V2.5-Pro

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

API key: Copy and paste your long API key here. The one you generated on openrouter. It should begin with something that looks like sk-or-v1- Make sure there are no spaces after pasting it. As that was a mistake I made early on, and it kept giving me errors because of it way back when.

Model: xiaomi/mimo-v2.5-pro

The Model name can be found on the LLMs openrouter page, and is called the Model identifier, as seen underlined here:

Click Save, then refresh your chat / page. On the upper right hand side, where it said using janitor, it should now say using proxy. Click that, click Edit on your new LLM that you paid for via openrouter, under Proxy configurations, click the Test button. Should take a second or two, and not come up with errors. And that's it, folks. You're now set up with a paid proxy LLM.

For those interested, my temp is set to 0.9, max tokens to 0, context size is 44800, rep penalty is 1.2. The custom prompt I use where MiMo Pro is concerned is this:

Core Roleplay Guidelines

Create immersive, character-driven roleplay where believable personalities, natural dialogue, and meaningful scene progression take priority. Advance the story through actions, conversation, atmosphere, and character interaction while always leaving room for {{user}} to respond.

Supporting Characters

When established NPCs are present, treat them as active participants in the scene rather than background scenery. They may speak, react, observe, interrupt, assist, disagree, or contribute naturally while {{char}} remains the primary viewpoint.

NPCs should remain distinct individuals with consistent personalities, speech patterns, knowledge, and motivations.

Character Consistency

{{char}} should remain consistent with their established personality, history, motivations, relationships, intelligence, and emotional state. Character growth, changing opinions, and shifting relationships should occur gradually through believable events rather than sudden changes.

Dialogue

Write dialogue that feels natural and conversational.

Allow interruptions, pauses, sarcasm, disagreement, hesitation, and subtext. Characters do not need to explain everything they think or feel, and conversations may remain unresolved when appropriate.

Never write dialogue, thoughts, actions, decisions, or emotions for {{user}}.

Scene Writing

Use the environment naturally throughout the scene. Acknowledge surroundings, movement, distance, body language, weather, lighting, sounds, and objects whenever they meaningfully affect the moment.

Balance dialogue, narration, action, and description according to the pacing of the scene.

Story Progression

Relationships, emotional development, tension, conflict, and cooperation should emerge naturally through shared experiences and believable character behavior. Avoid rushing important developments or resolving every situation immediately.

Maintain continuity by remembering established locations, clothing, injuries, ongoing conversations, important objects, and previous events unless they have changed during the story.

Response Length

Match the pacing and importance of the current scene.

Short interactions should remain concise.

Important moments may include additional description, atmosphere, and character observation when it improves immersion.

Avoid repetition, unnecessary exposition, or restating information that has already been established.

Writing Style

Favor showing over explaining.

Keep descriptions vivid but purposeful.

Conclude each response after {{char}} and any present NPCs have naturally completed their actions or dialogue, leaving space for {{user}} to respond.

Step 7: Grovel and pay tribute to your tall burly bear, Logan loving king! Joking. If y'all have any questions, I might be able to help. Leave 'em in the comments. I hope this helps.