Hi, everyone! This Rentry is a collection of work explaining proxies
, prompting
, errors
, general how-tos
, terminology
, and more. It's meant to be comprehensible for everyone, including beginners/fresh hatchlings! ๐ฃ
If you've been in the JanitorAI Discord, you might know this work as:
- The threads
#๐ฃ ai models for fresh hatchlings
and#๐ณ ai errors for fresh hatchlings
If you've been to the Reddit, you might know it as:
- The
Ultimate Guide to Proxies
!
For a variety of reasons, this work no longer exists on either platform. One of those reasons is that 'non-vetted' guides are no longer allowed on Janitor's Discordโso this guide has found its new home here on Rentry.
I'm very, very grateful people are still looking for it and using it. My goal is, always and forever, to make proxies *more easily accessible.
Feel free to scroll down to see what parts of the Rentry were recently updatedโor, jump here to get started with the โจ real content โจ!
Announcements & Work Statementsโmore coming soon! ๐
This is in the process of being migrated. Last updates were Monday, August 11th. I...
- Added a section on the bot not writing correctlyโlots of fun stuff here!
- Added the OpenRouter full error explanations.
- Added the Chutes Error table and Full Explanations
- Gave some clarification on how DeepSeek's temperature settings work.
- Fixed inconsistencies in styling
- Added more to the prompting section, particularly on OOC commands
- Added to the General Troubleshooting Section section to explain what to do before worrying about an error.
- Played with colors.
- If the lightmode is too light, scroll aaaallll the way to the bottom and hit the little sun-moon icon :) I tried to mimic Discord's colors!
Next up:
- Further troubleshooting guides; Official DS coming shortly
- More prompt information
- More explanation on DeepSeek official's temp scaling
- I add cute colors everywhere
Just to clarify...
- ...Formatting in progress. I'm learning, still! ๐ฃ
-
...Expect missing images. Working on fixing that ASAP!
Thank you for visiting this guide with me, and thank you to all of my friends who helped me work on it!
โจ AI Models, Prompts, and Errors for Fresh Hatchlings โจ
A guide for people new to AI and a great reference those with experience, too!
Okay, first of all, don't get intimidated. Setting up a proxy is really, really easy. I promise! โค๏ธ
All you really need to do is:
- Pick a model (๐ง )
- Decide on a proxy service (๐)
- Copy and paste shit in the right place (โ )
You can do this!
You don't need to know everything in this thread. Use the table of contents below to jump around. There's also a full table of contents wayyyy down at the bottom, and linked here. The only reason it's not up top is because it's long, and I don't want people to be intimidated. ๐ซ
๐ช You can do this!
Table of Contents
Basic Terminology
๐ค If you have no idea where to start, go here, then continue down.
Proxy Options and Set Up
โจ Okay, you want to try a new model! Yay! Here are some common options and settings.
- โฐ Important Info - Are there any non-JLLM free models?, Chutes pricing changes
- ๐ค Pick a Proxy
- ๐ช Chutes
- ๐ OpenRouter
- ๐ณ Paid DeepSeek (official API)
- ๐ช Lorebary (Gemini Guide)
- ๐ Local Models
- ๐ง Pick a Model
- โ๏ธ Generation Settings
Prompting
๐ How to tell the model to do...stuff...things...etc...y'know.
Troubleshooting
๐ข stuff isn't working help help help help h
- ๐ฅ Error While Chatting
Something red appears while I'm trying to roleplay. - โน๏ธ Writing Incorrectly | Messages are Weird
Something about the bot's response is wrong. It's writing in a way it shouldn't, or screwing things up.
General Questions
๐ For things that don't fall under other categories.
Basic Terminology
Wow, what do all the words
and phrases
everyone's using mean? Well, now you can learn! And pretend you knew all of them when you got here!
๐ค What's a model/LLM?
A Large Language Model generates text. The default one here on Janitor.AI is JLLM.
Think of it like the brain (๐ง ) behind your character (๐ค).
The type of model you HUGELY impacts how the bot will make use of and understand your prompts and messages.
Different models are good at different things!
๐ค What's a proxy?
Technically, the term proxy refers to a middle-man service that sits between you, the user, and the actual API of whatever model you're using.
- OpenRouter, a very popular proxy service, is a great example of this. They're not hosting any models of their ownโthey just route your traffic to providers who host models, allowing you to have one key for many providers and models.
- Proxy services can also change the way you receive content. OpenRouter strips reasoning (the bot going
<think>Hm, okay, the user wants me to...[XYZ]</think>
) from its messages.
- Proxy services can also change the way you receive content. OpenRouter strips reasoning (the bot going
Here on Janitor.AI, though, the term is โ๏ธ way broader โ๏ธ
On Janitor, all a proxy is is a way of accessing models other than JLLMโJanitor's default LLM. Almost all models are going to have better memory and writing than JLLM.
Once you select a proxy service to use, select a model.
๐ ฐ๏ธ What's a token?
Whole words, parts of words, special characters, and punctuation.
Any LLM can only understand/remember/manage so many at once. Memory isn't infinite; be conscious of your tokens!
- Use this link here and select
GPT-3 (Legacy)
for the closest approximation to how JLLM counts tokens. Through it, you can input things like your persona to see how long they are! - Use this link for other models.
๐ What's context?
A context window is the maximum amount of tokens a model can handle at once.
Permanent tokens, like:
- Bot Personality
- Persona
- Advanced Prompt
- Chat Memory
- as in, the feature in the top rightโnot your message history!
All stay forever and are sent to the LLM with every message.
Older messages are automatically pushed out of the context window when context "fills up" to make room for new tokens (via new messages). That's why the bot 'forgets' things.
โฌ๏ธโฌ๏ธ What should my context limit be?
๐ In most cases, set it to 16k.
๐
For Gemini
: 2.5 pro can handle significantly higher, over 100k.
For Claude
: ditto, but you'd also pay a lot of money.
Once you go beyond 16k, most models struggle to pull information accurately. It'll be slower (๐) and more forgetful (๐). There's a great resource about this here where you can measure models' performance against each other at different contexts.
Gemini, GPT, and Claude are the only true standouts. They use different forms of attention processing, allowing them to handle higher context. DeepSeek does not do this.
- Your prompts and the bot itself also have less weight compared to everyyyyything else you've got shoved into the context; the tokens of those Very Important Things are lost in a sea of thousands and thousands more tokens.
- Additionally, if you're paying for a model, you pay per token. High context will drain your funds very quickly.
๐ก๏ธ What does temperature mean?
Word selection probability.
๐ฅ High Temperatures: Tokens that are less likely become more probable.
At higher levels, you might get gibberish.
โ๏ธ Low Temperatures: Tokens that are highly likely are picked even more.
At lower levels, you might get stagnant, repetitive writing.
When setting a model's temp, always do it on Janitor.AI, not on the proxy service's site.
Note:
Not all models accept temperature parameters. (Meaning, for some models, if you put in a temp, it doesn't change anything.)
The great thing is, you don't really need to memorize any temperature parameters, just this little guide here:
- Generally, don't set it to 0 and 2,
- If it's spouting gibberish/jumbled up text, turn down ๐ the temp.
- If it's stale or stagnant/repeats the same phrases ad nauseam, turn ๐ up the temp.
- Play around with temperature! Use what works best for you.
๐ What's an API key?
Think of it like a password that allows you to use the models your proxy provides.
Keys are per-proxy service/provider, not per model.
- EX: I'm using Chutes's DeepSeek V3 0324 and I want to use Chutes's DeepSeek R1T Chimera. I don't need to switch keys! โ
- EX: I'm using Chutes's DeepSeek V3 0324 and I want to use OpenRouter's R1. I need to switch out my key, API URL, and model name! โ
- โ Why is my key disappearing? (See General Questions)
- โญ Make your key on the website of whatever proxy you're using. Guide further down!
๐ API URL
Specific URL that connects you to an API. Without it, you won't be able to use the models you're looking for!
It is extremely, ridiculously, vitally important your API/Proxy URL is correct. This is genuinely probably the number one issue I've seen people have.
Different proxies have different URLs. Just paste it in the Proxy URL box. You don't need to edit it.
๐จ No. This isn't just a link to your proxy's website.
โ No. You don't need to click it.
๐ค No. You don't need to have your specific model's name inside the link. That v1 you see is for the endpoint version, not the model name!
- ๐ช Chutes:
https://llm.chutes.ai/v1/chat/completions
- ๐ OpenRouter:
https://openrouter.ai/api/v1/chat/completions
- ๐ณ DeepSeek's Official API:
https://api.deepseek.com/v1/chat/completions
- ๐ Lorebary:
https://lorebary.sophiamccarty.com/aistudio
- โ๏ธ Eslezer's Colab:
https://gemiproxy-td8i.onrender.com/
๐ฅ๏ธ What's a colab?
Google Colab is a free cloud-based service that allows people to host and run code on isolated environments called notebooks.
Here, it's usually used for some Python-based proxy setups. Essentially, they allow you to do more complex configurationsโand get around Gemini and Claude's filters.
With a colab
you generate a URL every time. It doesn't take much thought to run, and isn't dependent on someone else's permanent server.
With a permanent server
(Lorebary, Eslezer's server, etc) you're using someone else's server. Easier to use, but might experience instability.
Proxy Options & Setup
If you're here, it means you're ready to choose a proxy and model
โor view standard settings
. Let's go!
Read carefully. If you ask a question about something answered in here, wild dogs will tear you apart.
๐จ Important Information
๐ช Are there any non-JLLM free models?
You can still use models for free through these services:
OpenRouter (๐):
- โ๏ธ 50 msgs/day (1000 msgs/day with free modelsโthose with :free at the endโif you pay)
- โฐ resets at 12:00 AM UTC/8:00 PM EST your time.
Google Cloud Trial (โ๏ธ)
- โ๏ธ $300 of Gemini credits, lasts 3 months.
- ๐ Needs a bank card to begin trial.
- โค๏ธ It has limited free model usage, even without a trial
Mistral (๐)
- โ๏ธ 1,000,000,000 (a billion) tokens (not messages) per month.
- ๐ lower quality, but free, and better than JLLM!
- 1๏ธโฃ serves only Mistral models
Quick note
on Mistralโthe link sends you off this Rentry and over to my friend Enza's guide! :)
(It's very short and easy to follow.)
Or, set up a model locally. You can learn all about what that means by clicking the link here.
๐ I want DeepSeek unlimited! Where can I use it for free and have it be unlimited?
Tough luck. I want that too!
The answer is nowhere unlimited, some places free. OpenRouter will get you 50 msgs/day.
Paying DeepSeek for the model is also โจreally cheapโจ.
Or, if you just want to pay once and be done forever, you can ๐ pay OpenRouter $10.00 USD (learn more) ๐ and get 1000 msgs a day forever.
๐ค What's the point of paying for a model?
Stability and better quality.
- 1๏ธโฃ Free things rarely stay free. Chutes just swapped to another new payment plan.
- 2๏ธโฃ Paying means better quality models โ .
๐ Better memory! ๐ฒ Better responses! ๐ Better performance! ๐ Woo! โจ
๐คน Which model is best / which should I use?
This is โจ subjective โจ. Everyone's going to like something better than something else and give you a different answer!
The best thing to do is โญ experiment with free models โญ!
Of the model groups listed below, only Claude is difficult to find free trials of. The rest have extremely accessible access.
Learn about:
โค๏ธโ๐ฅ Are there any uncensored models?
DeepSeek (๐ณ) models, with the exception of MAI-DS-R1
, are largely โจ unfiltered โจ.
- If it's blabbering about guidelines, that's a hallucination.
- Go to General Questions to learn how to stop it!
Generally, models from corporate (๐ผ) providers like Google/Gemini, Anthropic/Claude, and OpenAI/GPT need some work to avoid filters.
๐ณ DeepSeek Models
๐ง R1 0528
Brand new version of R1.
If you've heard people going "woo, yeah, I love DeepSeek!!! ๐ธ", there's a massive chance they were talking about this modelโor its cousin V3 0324.
Characteristics:
- ๐ง intelligent
- ๐ strongly adheres to prompts
- ๐ sticks to the bot's personality
- โค๏ธโ๐ฅ intenseโsometimes too intense
- ๐ค better at recalling information
- โ๏ธ a little 'meaner'
- ๐ฅ great with multi-character bots, intense action, grittier worlds...
This is a reasoning model, so it'll explain its thought process before replying. It's really good at it, too!
โญ it portrays the traits the bot has to the โจ fullest extent โจ; if a HotBoy MafiaMan
has possessive in his Personality
, he'll be possessive with R1 0528!
โญ it being so prompt-focused means, for some people, difficulties in getting the character to progress past their defined traits
Settings & Model Names:
- OpenRouter name:
deepseek/deepseek-r1-0528:free
- OpenRouter (paid) name:
deepseek/deepseek-r1-0528
- Chutes name:
deepseek-ai/DeepSeek-R1-0528
- DeepSeek (official) name:
deepseek-reasoner
- Temperature (from DeepSeek directly):
doesn't matter; official API doesn't take temperature
- Temperature (from other source):
0.4โ0.9
- Max new tokens:
0
โ How do I make it stop showing the <think>Okay, let's break down the user's response...</think> stuff? Answer in General Questions.
๐ฑ V3 0324
This is the other famous DeepSeek
model!
Characteristics:
- ๐ฐ softer, fluffier experience
- โฎ๏ธ less intense
- ๐ loves to end scenes early
- ๐๏ธ decorates its text with excessive asterisk use, sometimes
- ๐จ moldable
Not a reasoning model. You'll get the response alone, no thinking attached.
โญ your input matters a lot to this modelโallowing you to mold the roleplay into the shape you like. it also means characters can sometimes get too soft, because this model's very eager to agree with you, the user!)
Settings & Model Names:
- OpenRouter name:
deepseek/deepseek-chat-v3-0324:free
- OpenRouter (paid) name:
deepseek/deepseek-chat-v3-0324
- Chutes name:
deepseek-ai/DeepSeek-V3-0324
- DeepSeek (official) name:
deepseek-chat
- Temperature (from DeepSeek directly):
1.5โ1.8
- Temperature (from other source):
around 0.9
๐ R1T-Chimera:
Not actually from DeepSeek. It's a 3rd party combination of V3 and R1. It's, like the name implies, a chimeraโa mixture!
Characteristics:
- ๐ฑ softer, fluffier experience
- ๐๏ธ decorates its text with excessive asterisk use, sometimes
- ๐ sticks to the bot's personality
- ๐ can be buggy
This is...close to...a reasoning model, so it'll often explain its thought process before replying. It works in an odd way.
๐ (no paid options. also, it doesn't work well on OpenRouter for technical reasons)
Settings & Model Names:
- Chutes name:
tngtech/DeepSeek-R1T-Chimera
- Temperature:
0.6โ0.8
๐ด MAI-DS-R1
Re-trained version of R1, where Microsoft replaces DeepSeek's censorship with its own.
Characteristics:
- ๐ tamer, less unhinged version of R1
- ๐ก๏ธ but, it's censored, with an emphasis on preventing explicit content
- ๐ฅถ anything NSFW will be tougher to do
- ๐ด less good at following advanced prompts*
*โ๏ธ when we say tame, we mean tame. Microsoft's censors tend to focus on NSFW. Good for a soft, SFW experience...and not much else!
๐ no paid options
Settings & Model Names:
- OpenRouter name:
microsoft/mai-ds-r1:free
- Chutes name:
microsoft/MAI-DS-R1-FP8
- Temperature:
0.6โ0.8
Other DeepSeek Models
- ๐ V3 | V3 Base: โ Do not use; older version of V3 0324. It's bad. Extremely repetitive.
- ๐ Prover (V2): โ Do not use; it's a model for math.
- ๐ anything named 'distill': โ Do not use; condensed version shoved into a smaller, less intelligent model. Not what you want.
- ๐ R1 Zero: โ Do not use; this isn't the R1 you want.
"DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing." - OpenRouter model page
โ Gemini Models
People really, really like Gemini. Gemini models are:
- great at pulling information over larger context windows (you can use above 16k with these!)
- very receptive to prompts
- I personally find that 2.5 pro listens much better than DeepSeek's V3 0324 and R1 0528 models (...and I paid for the official API ๐ข)
...It's just also got a filter to navigate. Turn off text streaming and use a colab or a permanent server.
Gemini-2.5-Pro
๐ฆ 100 messages a day... (refreshes at 12:00 PM UTC)
๐ ...unless you're using the free trial!
๐ By setting up the free trial, you get $300.00 USD in credits. Credits last for 3 months. That's plenty to use on Pro.
๐ค Reasoning model
Characteristics:
- ๐ง super smart
- ๐ passive, wants you to lead the narrative
- ๐จโ๐จ great at simulating authors, good writing style
- โจ in-depth fandom knowledgeโeven obscure ones!
- ๐ good at listening to prompts
Set up using the Gemini Guide.
AI-Studio name: gemini-2.5-pro
The below models are free, but not...great. You're better off using Pro!
Gemini-2.5-flash
๐ฆ 250 messages a day (refreshes at 12:00 PM UTC)
๐ Free
๐ค Reasoning model
"It's all right... I guess? It can keep a roleplay going, but tends to get really repetitive." - resident Gemini user
"Not great at long term roleplay like Pro is."
Set up using the Gemini Guide.
AI-Studio name: gemini-2.5-flash
Gemini-2.5-flash-lite-preview-06-17
๐ฆ 1000 messages a day (refreshes at 12:00 PM UTC)
๐ Free
๐ค Reasoning model
"Use this over normal Flash. Works just like Flash in roleplay, minus the repetition."
"Not great at long term roleplay like Pro is."
Set up using the Gemini Guide.
AI-Studio name: gemini 2.5 Flash-Lite Preview 06-17
๐จ Claude Models
For oil princes and those who want the best roleplay experience. Keep context lowโit'll cost you.
All Claude models (except Haiku...) will have:
- โ๏ธ excellent writing
- โค๏ธ a tendency to stay in character
โ ๏ธ These models all have filters. You'll need to use them through something like OpenRouter to bypass that.
โ ๏ธ
Sonnet 3.7
- ๐ฐ great at driving the plot
- ๐ค can be wordy and technical
- โค๏ธโ๐ฅ less filtered
- ๐ great at slowburn
- ๐ stays in character very, very well
Sonnet 4
- ๐ธ positivity bias
- ๐ very opinionated; will call certain roleplay topics toxic
- ๐ฑ fluff, no matter the situation
Opus 4
It's great. It's expensive.
Opus 3
๐ Ditto.
Haiku
Not as good as the others. Dry. You can do better.
You won't get your money's worth on Janitor.AI, though. These models are better off used with alternate front-ends if possible.
These models should be set up through OpenRouter, or through Sophia's Lorebrary.
๐ Local Models
Local models are a great way to use models fine-tuned specifically for roleplay. You don't have to pay, and there's many (try Snowdrop!) out there with amazing quality.
My friend Naen (๐) wrote a great guide about itโit's about using models on Kaggle.
What that means is that you're 'borrowing' a GPU, basically! So, you can host a model no matter how strong the computer you're using is.
There's also a guide for using Snowdrop, specifically.
BTW: ๐จ No, you cannot host DeepSeek locally. It's just too big. ๐จ
Any version of DeepSeek you're able to host on your computer is going to be a heavily quantized, dumber, less functional version than the real deal.
๐ค Proxy Options
As a reminderโproxies are just where you get your models from. ๐ค
Sometimes, specific models are only available from specific proxies.
๐ช Chutes
โจ API URL: https://llm.chutes.ai/v1/chat/completions
Chutes pricing has changed again.
Now, for all users that haven't previously paid them $5.00, there's a subscription model.
Base |
$3/month
: 300 requests/day, unlimited API keys & all our models, Chutes Chat & Studio, PAYG beyond limits.
Plus |
$10/month
: 2,000 requests/day + email support.
Pro |
$20/month:
5,000 requests/day + priority support.
To put it franklyโthis now makes Chutes easily the worst choice of the big three DeepSeek Providers. You're better paying as you go with the Official API, or using OpenRouter (1 time payment of 10 bucks).
Chutes FAQs
๐ช Does Chutes accept card or just crypto?
Card + crypto! I don't recommend paying, but it's your life! ๐ ๐ซก
- Go to Chutes.
- Click the three-meatball icon (API)
- Click Billing and Balance
- Hit Add Balance
This only has to be done once.
๐จ I cannot troubleshoot payment stuff. Seriously. Please bug RayonLabs about it!
๐ก๏ธ Is Chutes safe?
It's a crypto company, so...it's as safe as any crypto company.
They're not evil people. They don't collect your email address or identifying info (hence the fingerprint you made when signing up).
Personally, I think the biggest danger is a โ bad deal โ. OpenRouter will deliver more messages for less money (and allow you to access the Claude modelsโand DeepSeek's official API is a better pay-as-you-go plan.
๐ Is Chutes's new 300/day plan forever or for a year?
Forever, as far as we know.
Make no mistake, though. This is a perk, not a guarantee. They've changed pricing before. They could change it again.
Pros:
- โ Rerolls are counted as 1/10th a message. If you reroll more than 9 times a message, this might be good for you.
- โ If you use V3 0324, it has better uptime on Chutes than on OpenRouter. Otherwise, not really a difference.
- โ If you're a big R1T Chimera guy, it runs best on Chutes.
Cons:
- โ Shows thinking/reasoning. Don't like that? Pick a non-thinking model or switch providers.
- ๐ค Now a monthly subscription instead of a one-time payment.
๐ช Chutes Set-Up Instructions
1: Go to: https://chutes.ai/auth/start
2: Make an account.
3: You'll get something called a fingerprint. Save this in your notes app or something.
โญ Are you reading this? Save it. Save it. You will not be able to get into your account without it. If you lose it, your account is lost permanently. There's no recovery.
4: Go to this link, specifically: https://chutes.ai/app/api. Scroll down. Hit Create API Key.
5: Save it somewhere. It will โ ๏ธ not show up again โ ๏ธ.
6: Open Janitor.AI. Start a chat. Navigate to API Settings in the top right.
7: Hit Add a Configuration.
8: Make the Configuration Name whatever you want.
9: Put your model name in the Model Name box. Do not type this by hand, copy and paste it.
- R1 0528's model name:
deepseek-ai/DeepSeek-R1-0528
- V3 0324's model name:
deepseek-ai/DeepSeek-V3-0324
- I'm using something else: Search your model on Chutes. Click on it. Hit Chat. It will show the model name right
above
the words "Start chatting with[model]
."
10: Put the API URL in the Proxy URL box: https://llm.chutes.ai/v1/chat/completions
โญ It needs to be that link exactly. Don't change it, don't click on it, just copy and paste it in.
11: Put your API key in the API Key box. Make sure there are no extra spaces at the end.
12: Add a Custom Prompt in the Custom Prompt box. (These aren't universal/don't carry over per config! See Prompting for guidance.)
13: Hit Add a Configuration.
14: Scroll down and hit Save Settings.
15: Refresh your whole page! You did it! Yay!
๐ฅ "oh god i'm getting an error something went wrong something went wrong"
[INSERT IMAGE HERE - Screenshot of completed Chutes configuration in Janitor.AI]
๐ OpenRouter
โจ API URL: https://openrouter.ai/api/v1/chat/completions
A proxy service that routes both free and paid models. It doesn't host any models of its ownโit gets them from other providers, like Chutes.
โ ๏ธ What that means is: if Chutes goes down, OpenRouter does too.
Free Tier:
Free users get 50 messages per day, per account. Limit resets at 12:00 AM UTC/8:00 PM EST your time.
No, these โ don't stack โ. If you only used 10 yesterday, you don't get 90 today.
Paid Tier:
Adding $10.00 USD of credits (through this link) will allow you to have 1000 messages a day with free models.
You only have to do that once. For the rest of your time on Earth, you now have 1000 messages a day. If you don't pay, you're stuck with 50.
These credits you just bought can be used however you like! Try paid models!
OpenRouter FAQs
๐ป How to pay OpenRouter? What if I don't have a card?
- Click your profile picture.
- Hit add credits.
- Add 10.00 USD. This only has to be done once.
If you don't have a card, try Revolut to get a digital card. It's trustworthy, reliable, and allows you to top up with virtually every payment method on Earth to make a temp or permanent digital credit card to use with OpenRouter.
I can't troubleshoot this for you, but it's an alternative if you don't have your own debit card!
โค๏ธโ๐ฉน Only OpenRouter itself and Revolut (if you're usin' em) can troubleshoot payment stuff. Seriously. They can help; I can't. โน๏ธ
๐ก๏ธ Is OpenRouter safe?
Yes. But if you feel uncomfortable, don't pay them. Choose your own destiny. Talk to them about concerns in their Discord.
๐ If I use my balance, do I still get 1000/day?
Yup! Use your balance however you like. Lopping off the :free
at the end of your model name means you're using paid models, which decreases your credit balance.
- Those paid models will usually have better quality.
- You still get 1000/day. It's based on lifetime purchase, not current balance.
๐ Is it 1000/day forever or for a year?
- It's foreverrrrrrrr.
- ๐ค Or at least until they change it. They could do that whenever, but they're trying their best to retain free access.
They reserve the right to expire your credits a year after purchase. That's different. Your 1000/msgs are based on lifetime purchase (meaning, you only have to pay once).
- ๐ค Or at least until they change it. They could do that whenever, but they're trying their best to retain free access.
tl;dr: spend 10 bucks one time, keep 1000 msgs on free models per day forever. Unless something goes catastrophically wrong.
๐ OpenRouter Set-Up Instructions
1: Go to: https://openrouter.ai/
2: Click Sign In, then Sign Up. Make an account.
3: Click your profile picture. Then, click Keys.
4: Name your key whatever you want. Leave the usage box blank.
5: Once your key appears, โ ๏ธ copy and save it somewhere โ ๏ธ.
๐จ After you make your key, it will not show up again. If you want to save yourself the hassle of making a new one every time you log in or our of Janitor, save your key somewhere.
6: Open Janitor.AI. Start a chat. Navigate to API Settings in the top right.
7: Hit Add a Configuration.
8: Make the Configuration Name whatever you want! It's just a fun way of keeping your configurations separate.
9: Put your model name in the Model Name box. Do not type this by hand, copy and paste it.
R1 0528's model name: deepseek/deepseek-r1-0528:free
V3 0324's model name: deepseek/deepseek-chat-v3-0324:free
I'm using something else: Search your model on OpenRouter. Click on it. Click on the little clipboard icon
under the title. That's the model name!
10: Put the API URL in the Proxy URL box: https://openrouter.ai/api/v1/chat/completions
โญ It needs to be that link exactly. Don't change it, don't click on it, just copy and paste it in!
11: Put your API key in the API Key box. Make sure there are no extra spaces at the end.
12: Add a Custom Prompt in the Custom Prompt box. (These aren't universal/don't carry over per config! See Prompting for guidance.)
13: Hit Add a Configuration.
14: Scroll down and hit Save Settings.
15: Refresh your whole page! You did it! Yay!
๐ฅ "it's not working it's not working what do i do help"
[INSERT IMAGE HERE - Screenshot of completed OpenRouter configuration in Janitor.AI]
๐ณ DeepSeek (Official, Paid API)
โจ API URL: https://api.deepseek.com/v1/chat/completions
This is DeepSeek as it was intended to be experienced. The real deal, the source of the best quality. Always paid.
There is no way to use the official API without paying.
Thankfully, it's really, ๐ really cheap ๐. There's even a discount period between the hours of 2:30 AM and 10:30 AM UTC, where token usage costs up to 75% less.
How it works:
- You 'top up' however much you like on your account. It's okay to start with a small amount. Minimum of $2 USD.
- Tokens are priced per million. As you chat, balance is deducted from the amount you topped up.
Usage Example:
- I put โจ $5.00 USD โจ on my DeepSeek account.
- The bot I'm chatting with takes up a total of 2500 tokens.
- My first message is 300 tokens, which I calculated here.
- The bot replies with a message that's 1500 tokens long.
- I'm using it during the discount hours.
- The discount price of deepseek-chat (V3-0324) is
$0.135
per million input tokens,$0.55
per million output tokens.
(finding input price per token)
0.135/1,000,000
= 0.000000135
(finding output price per token)
0.55/1,000,000
= 0.00000055
(input tokens multiplied by input price)
2800 x 0.000000135
= 0.000378
(output tokens multiplied by output price)
1500 x 0.00000055
= 0.000825
(adding together)
0.000378 + 0.000825
= 0.001203
So, of my 5 dollars, I have.... ๐ฅ
$4.998797
...which is basically still 5 dollars.
๐ณ DeepSeek (Official) Set-Up Instructions
1: Sign up on https://platform.deepseek.com/
2: Create a key at https://platform.deepseek.com/api_keys.
โญ Save it. Look me in my eyes. Save this somewhere. Also, because you're paying for it, please don't share it.
3: Go here https://platform.deepseek.com/usage. Add some money.
Pricing explanation above. Minimum is $2 USD.
4: Open Janitor.AI. Start a chat. Navigate to API Settings in the top right.
5: Hit Add a Configuration.
6: Make the Configuration Name whatever you want.
7: Put your model name in the Model Name box. Do not type this by hand, copy and paste it.
R1 0528's model name: deepseek-reasoner
V3 0324's model name: deepseek-chat
8: Put the API URL in the Proxy URL box: https://api.deepseek.com/v1/chat/completions
โญ It needs to be that link exactly. Don't change it, don't click on it, just copy and paste it in!
9: Put your API key in the API Key box. Make sure there are no extra spaces at the end.
10: Add a Custom Prompt in the Custom Prompt box. (These aren't universal/don't carry over per config! See Prompting for guidance.)
11: Hit Add a Configuration.
12: Scroll down and hit Save Settings.
13: Refresh your whole page! You did it! Yay!
[INSERT IMAGE HERE - Screenshot of completed DeepSeek configuration in Janitor.AI]
โ Gemini Guide
Right now, Gemini has limited free messages daily. It differs per modelโcheck the Gemini Models section. It resets at 3:00 AM EST
/7:00 AM UTC
Or, you can use a free trial that lasts three months, and get greater access to these models!
๐ค Free Tier Guide
๐ Google Cloud Trial Guide
Free Trial Note:
If you sign up for the free trial, you'll no longer have access to daily free models. Instead, you'll have the usage deducted from your trial balance. Still free, just in a different way.
Generation Settings
Okay, you survived the model explanations. Good work. If you somehow missed it earlier, here are the ideal generation settings:
๐ก๏ธ Temperature
๐ณ DeepSeek
- V3-0324 (When used directly from DeepSeek):
1.5-1.8
- V3 0324 (On Chutes, OpenRouter, etc):
~0.9
- R1/R1 0528: (When used directly from DeepSeek):
doesn't matter; it doesn't use the temperature parameter
- R1/R1 0528: (On Chutes, OpenRouter, etc):
0.4โ0.9
*- MAI-DS-R1:
0.6โ0.8
- R1T Chimera:
~0.9 (experiment)
*still investigating how this works with other providers. I know from personal experience that I got gibberish at higher temps with Chutes's R1 on Janitor. Might be a Janitor thing, might be a provider thing, but no guarantees. Still investigating! ๐
If a temp isn't listed here...play around with it!
If you're getting gibberish, turn your temp down. โ๏ธ
If you're getting dry, repetitive shit that uses the same phrases over and over, turn your temp up. ๐ฅ
๐ Context
๐ In most cases, set it to 16k.
๐
For Gemini
: 2.5 pro can handle significantly higher, over 100k.
For Claude
: ditto, but you'd also pay a lot of money.
Once you go beyond 16k, most models struggle to pull information accurately. It'll be slower (๐) and more forgetful (๐). There's a great resource about this here where you can measure models' performance against each other at different contexts.
Gemini, GPT, and Claude are the only true standouts. They use different forms of attention processing, allowing them to handle higher context. DeepSeek does not do this.
- Your prompts and the bot itself also have less weight compared to everyyyyything else you've got shoved into the context; the tokens of those Very Important Things are lost in a sea of thousands and thousands more tokens.
- Additionally, if you're paying for a model, you pay per token. High context will drain your funds very quickly.
๐ Max New Tokens
Reasoning models are best with 0
โwhich is 'unlimited'. It's not actually unlimited, it's 2,000, but it should be plenty.
A non-reasoning model is left up to you. Keep in mind that models 'reserve' all that space for the next response, even if it doesn't end up taking that much. So, that's 2,000 less tokens of memory that could be spent on other things.
If you're getting responses that โ๏ธ cut off early โ๏ธ, turn your max new tokens up.
Prompting
It's time to teach the LLM how to drive the roleplay! Hereโit's really not as complicated as it seems. All a prompt is is instructions to the model. These affect your whole roleplay.
๐ตโ๐ซ Do I really need a prompt?
Yes.
It'll work without one, but not well.
Prompts are so, so important in making sure your roleplay turns out the way you want it to.
๐ What's the best prompt?
There's no definitive best promptโyou can even try writing one yourself! Just take a quick stroll through the prompting section here to learn a little more about how that works.
If you're using someone else's prompt, be choosy. Pick the parts most important/relevant to what you want out of a roleplay. Many prompts use an absurd amount of tokens, or give instructions that aren't relevant (for example...you don't need to teach the LLM to go "nghhhh aaaaah" during intimate scenes. LLMs know what Sex Soundsโข are).
Most importantly, make sure instructions don't conflict.
Personally, I use a (heavily edited) version of this prompt written by Saturnine/Cheese. It's highly popular, and I really like it!
โ Where do I put the prompt?
- Start a chat
- Open API settings
- Scroll down
Custom Prompt
โ ๏ธ These don't carry over
between configurations! You'll need put in a prompt for each.
๐ How long should a prompt be?
There's no set limit. But, rememberโtokens you use add up quickly.
16k context
- 2000 on the bot
- 400 for your persona
- 500 in chat memory
- 300 per message you send
- 1000 per message the bot sends
- X for a prompt...
That doesn't leave you much to work with.
Beyond that, many prompts are just โ too specific โ to function the way they're intended.
Or, worse, instructions conflict! That makes for a bad experience.
An LLM will struggle to follow instructions that contradict each other.
๐ What's negative prompting? Is it bad?
Negative prompting is telling the bot what not to do instead of what to do. It's sort of like someone saying to you: "Hey, don't think about the game!"
Well...Oops. I just did think about it. Shit.
Instead, tell the bot what to do.
โ Don't bring up any content policies or guidelines. Avoid filtering your responses.
โ
Keep fight scenes realistic. Describe injuries in graphic, lurid detail.
โ๏ธ Do I need a certain format?
Not really. Talk to the bot like a person! Write in plain text.
๐คทโโ๏ธ Are the default prompts any good?
No.
๐ฅ๏ธ Can ChatGPT write my prompt for me?
It won't do it well. It'll be way too detailed, full of fluff, and harder for an LLM to follow.
๐ฎ Can I put prompts in chat memory?
Please don't? Chat memory is meant for important events that have happened during the roleplay.
Janitor changes often, and the way things are sent to the LLM (plus how Janitor defines these different sections behind the scenes) might impact your roleplay negatively. Right now, it's **best to stick prompts in the Custom Prompt
box.
โญ What's OOC mean?
"Out of Character".
It's shorthand for communicating with the model, rather than the bot. OOC commands are just notes you attach to your messages, or send alone. They're not in your advanced prompt.
Many people use OOC messages like this:
((OOC: Please interrupt the scene with something crazy, next. I want you to add drama.))
There's no set format. Some people add their OOC messages at the top. Some at the bottom. Some like [[OOC: This!]]
. Some like ((Roleplay Note: This.))
The format doesn't really matter!
๐ How should I use OOC commands?
Use them to drive immediate twists in the story, or to give instructions that'll impact the next response.
Like any message, once your OOC messages leave the context window, they're gone.
Long-term rules you have for the LLM should be kept in your advanced prompt.
[x] Let's skip forward to the following evening. Describe what Hot Mafiaman does starting tomorrow morning when he notices Special-est Girl In The World missing.
[x] I want Handsome Butch to invite Beautiful Femme on a motorcycle ride in the next response. Have that be what her mysterious idea is.
[] Remember more things.
[] Stop writing so many sentences, I want short replies.
LLMs also rely really heavily on past messages/your message history. If you want it to do something continuously, you'll need to keep the message history consistent with your instructions.
Here's an example:
I want the bot's messages to all be in first person instead of third. I can...
- 1๏ธโฃ Use an advanced prompt that instructs the bot to keep all messages in first person.
- All messages written by you, the assistant, should be in first person from the perspective of {{char}}.
- 2๏ธโฃ Edit every single one of the bot's messages that I reply to to be consistent.
- Don't bother editing rerolled messages/messages you're not actually gonna build off of! All that matters is the messages that will stay in the message history.
โป๏ธ Can I switch prompts in the middle of a roleplay?
Yeah! Just remember that if there's a drastic switch in instructions, you might be better off starting a new chat. Again, LLMs rely really heavily on the conversation history. If you decide in the middle of the chat to, say, use a prompt that instructs the bot to put all dialogue between asterisks, it might struggle, because you've got a wholeee conversation prior that doesn't do that.
General Questions
Okay, you made it! Yay! it's the end! Here's all the stuff that didn't fit earlier.
โญ Do star ratings work on proxies?
No.
๐ช Does enhance message use my proxy?
Yes.
โฉ Does continue message use my proxy?
Yes.
๐ฎ How should I set up chat memory?
Remember, memory takes up tokens.
To best optimize this, write chat memory yourselfโor use finalphoenix's amazing memory management prompt.
Side Note
: That whole Rentry is a genuine wealth of information. I highly advise checking it out!
Use bullet points or short sentences to describe major happenings in the story.
Chat Memory Update
Should be back to normal as of recently! It's reverted. Chat memory will no longer wipe previous messages.
But. It's not a bug, it's a feature. It'll come back soonโand Shep's potentially adding a toggle for the message-wipe feature.
๐ก๏ธ Is using a proxy safe?
Proxies log what you send. That includes a bot's description, personality, your persona, and your chat history.
If you're uncomfortable with this, don't use them. But, rememberโnobody cares about your roleplay. You're a drop in the sea in terms of traffic! ๐
Security Reminder
Never feed confidential information to an LLM.
โช Why are reverse proxies forbidden on JanitorAI?
Reverse proxies and privately hosted proxies by users are forbidden because they log your API key and your chat.
They can be accessed by unknown users. This isn't safe.
๐ณ Troubleshooting Guide
Applies to both โจJLLMโจ (that's Janitor's default LLM) and โจproxiesโจ (services that allow you to access other models). Let's handle this together!
You got this. Please use the links below to jump around.
If your error isn't here (and you have access to the Janitor Discord), please browse:
- ๐ช #tech-support
- ๐ #report-bugs
- ๐ค current discussion in #ai-models
If you still can't find your error, feel free to DM me (@jupitersoup on Discord) with it! I'm still building out this Rentry, so things aren't all here yet.
๐ฅ Error While Chatting
Something red appears while I'm trying to roleplay.
โน๏ธ Writing Incorrectly | Messages are Weird
Something about the bot's response is wrong. It's writing in a way it shouldn't, or screwing things up.
๐ป Bot Not Replying
Not getting any messages from bots at all...no error message
, just no replies!
๐ Website Errors
Something red appears while I'm doing other thingsโlike trying to leave a comment, or upload a bot.
๐ฅ Error While Chatting
1๏ธโฃ First Steps
If you're not in the Discord, open Janitor, and click the little megaphone icon in the top right corner. It has site-wide announcements from the Discord announcements channel.
If you're in Janitor's Discord...
- Check #๐ฃโannouncements. Is there an announcement about the site being down?
- Check #general and #ai-models. Does it say Several people are typing...? Are people panicking?
If people are panickingโthis is probably a widespread issue. Sit tight. Try to scroll up a little and see what's going on.
2๏ธโฃ Otherwise...
- Open your API settings. Make sure your config says Active.
- If it doesn't, click on it.- Save your config. Hit save settings.
- Refresh your whole Janitor.AI tab.
- The whole thing! This is a necessary step!
- If you're on the app, get off the app. It's mostly broken, anyway.
- Open your config back up. Hit test. Screenshot or remember your error message.
- โ Okay, now you can can continue.
3๏ธโฃ Okay, almost done
- Read through ๐ *General Errors* below. It's short!
- Select whichever provider is applicable if that didn't solve it.
๐ General Error Table
Each of the links in the in the Code column sends you to a fuller explanation! Click, or scroll down.
Code | Message | Explanation | Solution |
---|---|---|---|
N/A | A network error occurred, you may be rate limited or having connection issues... |
Your Proxy URL is wrongโor you haven't refreshed | Copy the correct API URL exactly, hit save, and refresh your whole page. Getting your Proxy URL correct is critically, vitally important. It's super easy to accidentally use the wrong one, paste it wrong, etc. Learn more here. |
N/A | This model's maximum context length is X tokens. However, you requested Y tokens... |
You're using too many permanent tokens for your context | Trim down your permanent tokens - use smaller bots (under 3000 tokens) |
N/A | Character and user content ([number]) is above limit ([number])... |
You're using too many permanent tokens for Janitor's limited context | Reduce your bot size, persona, and advanced prompt length |
pgshag2 no response | pgshag2 no response from bot |
Multiple possible causes - servers, rerolling too fast, reasoning models | Wait it out, check max new tokens for reasoning models, or use devtools |
N/A | An error occurred when editing a message: Validation failed |
Janitor's servers are having issues | Wait it out, save your message, check announcements |
god who knows, this one's messy | Massive DOCTYPE HTML error taking up your screen |
Usually a server problem (eg: Cloudflare is down, or Janitor's servers are struggling. Sometimes, this is a settings issue. | Check |
N/A | unknown prompt response from worker for openai proxy generation |
You're in Janitor's queue | Save your message and reload the page |
Scroll down to view more in-depth explanations of these general errors, or skip to the error sections on...
๐ OpenRouter
๐ช Chutes
๐ณ DeepSeek's Official API
โ๏ธ Gemini
๐ก Local Models
๐ General Errors
A network error occurred...
or:
๐ฅ Doesn't actually have anything to do with your connection.*
๐ Your API url is wrongโor you haven't saved and refreshed.
Every time you change your proxy URL, or add one, hit save, then refresh your Janitor tab.
Common Proxy URLs:
- ๐ช Chutes:
https://llm.chutes.ai/v1/chat/completions
- ๐ OpenRouter:
https://openrouter.ai/api/v1/chat/completions
- ๐ณ DeepSeek (official paid API):
https://api.deepseek.com/v1/chat/completions
- โ๏ธ Gemini (AiStudio):
https://generativelanguage.googleapis.com/v1beta/chat/completions
- ๐ค Eslezer's Colab:
https://gemiproxy-td8i.onrender.com/
- ๐บ Sophia's Lorebary:
https://lorebary.sophiamccarty.com/aistudio
This needs to be exactly the correct link. Copy and paste it.
Common mistakes:
โ Did you accidentally capitalize anything?
โ Is there a stray space anywhere?
โ Are there any rogue letters?
โ Is there accidental punctuation at the end?
โ Was /chat/completions
added extra times?
โ Did you change v1
in the URL to something else? (don't do that)
โ Just to make sure, copy and paste the correct URL from here.
Hit save. Refresh again. If it still doesn't work, feel free to DM me on Discord (@jupitersoup) or ask around! But, you gotta give whomever you're asking for help the proxy URL you're trying to use.
*Doesn't usually have anything to do with your connection. If your signal is abysmal, then maybe. But, a mistake in your URL is the way, way, way more common culprit.
PROXY ERROR 400 (Context Length)
The wording changes slightly per provider! This is a โจproxy version of the error.
You're using too many permanent tokens for the context you have.
โญ Don't know what those words mean? Permanent tokens are the stuff that's always sent with every message - your persona, the bot description, advanced prompts, etc. Context is how much the model can "remember" at once.
Simply putโthe combo of your:
- persona
- the "chat memory" feature
- the bot itself
- your max new tokens
- and your advanced prompt
Are too big. Most bots over 2500โ3000 permanent tokens
are heavily bloated, too. There's very little in botmaking that could require that many tokens to be used. You'll have a worse experience because of it.
โ๏ธ Trim down your permanent tokens and you won't have this issue.
โจ Also...
Don't jack your context up to 128k. It should be at 16k for most models.
โ
Yes. 16k.
๐ฌ "But I want it to remember everythโ"
Most models actually remember and comprehend โlessโ with a higher context.
Reduce your context to stop it from trying to generate so much. Also, use smaller bots! Trim your persona down!
Context Length (JLLM)
This is the ๐งนJLLM version of the error listed above.
You're using too many permanent tokens for Janitor's limited context. (Monitor it in the "Today's Context!" channel)
โญ Don't know what those words mean? Permanent tokens are the stuff that's always sent with every message - your persona, the bot description, advanced prompts, etc.
Simply putโthe combo of your:
- persona
- the "chat memory" feature
- the bot itself
- your max new tokens
- and your advanced prompt
Is too big. Most bots over 2500 permanent tokens
are heavily bloated, too. There's very little in botmaking that could require that many tokens to be used. You'll have a worse experience because of it.
โ๏ธ Trim down your permanent tokens (specifically, use smaller bots) and you won't have this issue!
pgshag2 no response
This can happen for a lot of reasons.
1๏ธโฃ Your max new tokens aren't set as 0 and you're using a reasoning model. If reasoning exceeds the token limit, this happens.
2๏ธโฃ Servers are down/overloaded. Just wait it out.
3๏ธโฃ You're rerolling too quickly.
4๏ธโฃ If your error is still present after waiting a bit, you'll need to use browser devtools to troubleshoot further.
Validation Failed
or
or
Uh oh. Janitor's servers are having a meltdown. Give it a second. Check #๐ฃโannouncementsโhas a fix been posted? Have they told you to just wait?
While you're hereโsave your last message in your notes somewhere. It's possible it'll evaporate upon reloading.
DOCTYPE HTML Error
Firstโcheck over your settings. Head to back up to and make sure your data is correct. Model name, Proxy URL, everything.
Quick Links
:
- OpenRouter setup info,
- Chutes setup info,
- DeepSeek official setup info,
- Common Proxy URLs,
โญ Still not working? Check #report-bugs, upvote (๐) any message similar to your error. Wait it out.
(If you're able, you can also read the HTML itself to learn a little more. Sometimes, it'll mention a Cloudflare error. Janitor uses Cloudflare, so if it goes down, Janitor will struggle, too. Nothing you can do but wait.)
Unknown Prompt Response
You're in Janitor's queue! Copy down your last message into your notes, just in case it disappears. Then, reload your page.
You'll get this less often if you continually take actions on the siteโwhether that's messaging, clicking on bots, writing comments, etc.
General Notes:
Please rememberโyou're probably not the first person to get your error. Search some of the text in Discord or Reddit. Chances are, it's been asked and explained. It's not kind to force people to spoonfeed you easily accessible info. You've got this! โค๏ธ
Also for the love of god โ don't speak in the troubleshooting channels โ. It makes it tougher for devs & others to actually see the errors happening. โน๏ธ
๐ OpenRouter Error Table
Code | Message | Explanation | Solution |
---|---|---|---|
400 | X is not a valid model ID |
Wrong model name. | Go to https://openrouter.ai/ and search your model. Click the little clipboard icon under the title of the model. That's your model name! If you're trying to use free models, make sure the model ends in :free . |
401 | No auth credentials found | Your API key is wrong. | Look it over. Make sure there's no extra spaces or punctuation. Make sure it's not capitalized. It should start with sk-or-v1-. If it doesn't, you haven't made it on https://openrouter.ai/! |
402 | Prompt tokens limit X > Y . To increase, visit https://openrouter.ai/setting upgrade to a paid account |
You don't have a large enough balance in your OpenRouter account to cover the request you're making. | Check your model name . Does it have :free at the end? If not, you're using a paid model. Add :free , or search OpenRouter for free models. |
402 | This request requires more credits, or fewer max_tokens. You requested up to X tokens, but can only afford Y . To increase... |
See above. | See above. |
402 | Quota exceeded and account balance is $0.0, please pay with fiat or send tao... | You're either mistakenly using the Chutes API URL or you connected a Chutes key via BYOK. | Check your API URL. Is it https://openrouter.ai/api/v1/chat/completions ? Are you sure? Okay, good. Head here., Scroll to Chutes., Flip the toggle to disable the integration. |
404 | No allowed providers are available for the selected model | You've blocked all possible providersโor limited your providers to those that don't serve the models you're using. | Go to https://openrouter.ai/. Hit settings. Scroll down. Unblock Ignored Providers and clear out all Allowed Providers. Hit save |
404 | No endpoints found for model name .. |
That model doesn't exist anymore, or isn't currently being hosted on OpenRouter. | Using deepseek/deepseek-chat-v3:free ? It's gone now. Switch to deepseek/deepseek-chat-v3-0324:free , it's better! |
408 | timeout | OpenRouter's down or struggling. | Wait it out. Go lurk over in their Discord for updates. Nobody can do anything about this but them! |
429 | Rate limit exceeded: free-models-per-day. Add 10 credits to unlock 1000 free model requests per day... | You're out of free messages. Free users get 50 messages per day, per account. This includes error messages. These don't stack and it's not per model! | Wait until 12:00 AM UTC/8:00 PM EST for it to reset! |
429 | [your model name here] is temporarily rate-limited upstream. Please retry shortly, or add your own key... |
OpenRouter gets models from different providers, like Chutes. It doesn't host any models of its own. Chutes has decided to prioritize its own paying customers by throttling/limiting the amount OpenRouter customers can access V3 0324, which they host. When traffic to V3 0324 is heavy, you may receive rate limits. | Error messages eat up your 1000 (or 50) msgs a day with free models. Be careful. Use a paid model on OpenRouter, switch to R1 0528 (or something else), or use a different proxy. |
502 | Provider returned error | Servers are down on OpenRouter's side. | Check the Uptime Tab for your specific model on https://openrouter.ai/. Anything below 80% is bad, probably won't get a response. Wait it out! |
N/A | Unknown response: [object Object] (unk) | You're out of free messages. Free users get 50 messages per day, per account. This includes error messages. These don't stack and it's not per model! | Wait until 12:00 AM UTC/8:00 PM EST for it to reset! |
N/A | Provider Returned Error (unk) | Uh-oh. Servers are down on OpenRouter's side. Check the Uptime Tab for your specific model on https://openrouter.ai/. Anything below 80% is bad, probably won't get a response. | Wait or switch models. |
๐ OpenRouter Errors
OpenRouter 400 (Invalid Model)
Model name's โ wrong โ. Don't write it by hand, copy and paste it!
Common DeepSeek Models:
R1 free:
deepseek/deepseek-r1:free
R1 paid:deepseek/deepseek-r1
R1 0528 paid:deepseek/deepseek-r1-0528
R1 0528 free:deepseek/deepseek-r1-0528:free
V3 0324 paid:deepseek/deepseek-chat-v3-0324
V3 0324 free:deepseek/deepseek-chat-v3-0324:free
โญ Using something else?
Search your model on OpenRouter. Click on it. Click on the little clipboard icon under the title. That's the model name.
OpenRouter 401 (No Auth)
Your API key is โ wrong โ.
- ๐ look it over. Make sure there's no extra spaces or punctuation. Make sure it's not capitalized.
- โญ Should start with sk-or-v1-. If it doesn't, you haven't made it on https://openrouter.ai/!
Still not working? Make a new one (you don't need a new account, just a new key), then paste it in.
OpenRouter 402 (Insufficient Credits)
or
You don't have a large enough balance in your OpenRouter account to cover the request you're making.
Check your โจ model name โจ. Does it have :free
at the end? If not, you're using a paid model. Add :free
, or search OpenRouter for free models.
โ Wait, what? But I never gave OpenRouter money!
๐ Yup, you can send your account a teeny tiny bit negative before it stops you. Don't worry, it's fine. Just switch to a free model.
OpenRouter 402 (Quota Exceeded)
๐ Check your API URL. Is it https://openrouter.ai/api/v1/chat/completions
? Are you sure? Check for me one more time.
โ Okay, I checkedโit is!
Yay!
In that case, you have a Chutes key connected to your accountโprobably for BYOK.
- Head here.
- Scroll to Chutes.
- Flip the toggle to disable the integration.
OpenRouter 404 (No Providers)
You've blocked all possible providersโor limited your providers to those that don't serve the models you're using.
Most commonly, this is because you've set allowed providers to DeepSeekโbut you're using free OpenRouter models served by Chutes.
- Go to https://openrouter.ai/
- Hit settings.
- Scroll down.
- Unblock Ignored Providers, clear out all Allowed Providers.
- Hit save.
OpenRouter 404 (No Endpoints Found)
Turn on model training in your privacy settings on OpenRouter. Go here.
OpenRouter 404 (Model Not Found)
That model doesn't exist anymore, or isn't currently being hosted on OpenRouter.
โญ Using deepseek/deepseek-chat-v3:free
? It's gone now. Switch to V3 0324, it's better
OpenRouter 408 (Timeout)
๐ญ OpenRouter's down or struggling. Wait it out. Go lurk over in their Discord for updates. Nobody can do anything about this but them!
OpenRouter 429 (Daily Limit)
or
You're out of free messages. Free users get 50 messages per day, per account. This includes error messages.
No, it doesn't stack. You don't get 60 tomorrow if you only used 40 today.
No, it's not per model! It's across all free models.
Limit resets at midnight UTC your time.
OpenRouter 429 (Upstream Limit)
OpenRouter gets models from different providers, like โจ Chutes โจ. It doesn't host any models of its own.
Chutes has decided to prioritize its own paying customers by throttling/limiting the amount OpenRouter customers can access V3 0324, which they host.
๐ When traffic to V3 0324 is heavy, you may receive rate limitsโyou'll have to wait and reroll.
โ Error messages eat up your 1000 (or 50) msgs a day with free models. Be careful.
If you've already paid 10 USD to OpenRouter, switch to the paid version of the model (get rid of the :free
at the end), or move to R1 0528.
This will decrease your balance on OpenRouter, but it will allow you to avoid the rate limit.
๐ง unga bunga explanation:
free models like v3 0324 are used a lot. lots of use means free models = overloaded. use OpenRouter's paid v3 0324, wait it out, find another model, or switch to a different proxy.
OpenRouter 502 (Bad Gateway)
OpenRouter's struggling ๐ต OpenRouter's struggling ๐ต 50X errors
are server side; nothing you can do except be patient!
Check the Uptime Tab for your specific model on https://openrouter.ai/. Anything below 80% is bad, probably won't get a response.
Wait it out! Nothing to do. Be warned, though:
โ Error messages eat up your allowed messages/day.
OpenRouter Provider Error
Uh-oh. Servers are down on OpenRouter's side. Check the Uptime Tab for your specific model on https://openrouter.ai/. Anything below 80% is bad, probably won't get a response.
Wait it out! Nothing to do. Be warned, though:
โ Error messages eat up your allowed messages/day.
๐ช Chutes Error Table
Code | Message | Explanation | Solution |
---|---|---|---|
401 | Invalid token. |
Your API key is wrong | Check for spaces/punctuation, should start with cpk_ |
402 | Quota exceeded and account balance is $0.0 |
You've hit your daily message limit and there's no money in your account to pay for more tokens. | Wait for it to reset at midnight UTC or add more credits to your account! |
404 | No matching cord found! or No matching chute found! |
Your Proxy URL is wrong | Paste this exact URL: https://llm.chutes.ai/v1/chat/completions . Check for extra spaces, extra chat/completions at the end, etc. |
404 | model not found: ... |
Your model name is wrong or you're in a blocked country โน๏ธ | Copy down the exact model name you're using Chutes or try VPN |
500 | exhausted all available targets to no avail |
Chutes's servers are down | Wait it out! |
503 | No instances available (yet) for chute_id=... |
Chutes's servers are down | Wait it out! |
๐ช Chutes Errors
Chutes 401 (Invalid Token)
Your API key is wrong.
- ๐ look it over. Make sure there's no extra spaces or punctuation.
- โญ Should start with cpk_. If it doesn't, go make a new one on Chutes's website.
Still not working? Make a new one (you don't need a new account, just a new key), then paste it in.
Chutes 402 (Quota Exceeded)
Chutes is paid. There's not a free tier anymore. If you have paid and you're hitting this, it's because you've hit your message limit. Your message limit depends what plan you're on.
Rerolls count as 1/10th a message.
No, these โ don't stack โ. If you're allowed to use 300 a day, and you only used 50 yesterday, you don't get 550 today. Still 300!
It resets at midnight UTC. Wait it out.
If you're using OpenRouter and getting thisโhead here. Scroll to Chutes. Flip the toggle to disable the integration.
Chutes 404 (No Cord/Chute)
OR
Your API url is wrong! It should be:
https://llm.chutes.ai/v1/chat/completions
exactly.
Make sure there's:
- no spaces anywhere
- not an extra
/chat/completions
at the end
Chutes 404 (Model Not Found)
Your model name is wrong!
Common Models:
R1-0528:
deepseek-ai/DeepSeek-R1-0528
V3-0324:deepseek-ai/DeepSeek-V3-0324
MAI-DS-R1:microsoft/MAI-DS-R1-FP8
Something else: Search your model on Chutes. Click on it. Hit Chat. It will show the model name right above the words "Start chatting with[whatever]
."
๐ This can also happen if you're from a blocked country. Try using a VPN.
Chutes 500 (Exhausted Targets)
Chutes's servers are down! Wait it out.
Check the Uptime Tab for your specific model on OpenRouter's website. Not a totally reliable method but...it's pretty decent, given that many models on OpenRouter are actually hosted by Chutes.
Wait it out!
Chutes 503 (No Instances)
Chutes' servers are down! Wait it out.
Check the Uptime Tab for your specific model on OpenRouter's website. Not a totally reliable method but...it's pretty decent, given that many models on OpenRouter are actually hosted by Chutes.
Wait it out!
โน๏ธ Writing Incorrectly | Messages are Weird
Table of Contents
Writing Issues:
- ๐ Bot Speaking for Me
- ๐ Bot not Listening to Prompts
- ๐ Requests Being Filtered
- ๐ช Messages Cut Off
- ๐ฑ Messages too Short
- ๐ณ Messages too Long
- โ Gibberish
- โผ๏ธ Too many asterisks and bolding
- ๐๏ธ Not Formatting Properly
- ๐ค Reasoning
- ๐ Ending Scenes Early
- ๐ฅ Losing Character Traits/Out of Character
- ๐คท Forgetting Important Details
- ๐ฎ Reading my Mind/Reacting to Thoughts
- ๐ณ๏ธโโง๏ธ Getting my Pronouns Wrong
๐ Bot Speaking for Me
It's speaking for me! Make it stop!
This...and don't shoot me for it...can often be a You Issue.
Let's tackle it!
1๏ธโฃ Use a bot with an opening message that does not include any actions that {{user}} takes.
โ Slop Sloppington thought about how much he loved {{user}}. He walked over to them, tilting their chin up, boxing them against a wall, and ruining them for anyone else. His rough skin and callused hands were rough and callused.
"All this shit is from my perspective," he said. "So it should work out okay. {{user}}, can I ask you a question?"
โ John Janitor watched as {{user}} smiled demurely.
"Wow, {{user}}." he said, watching her blush. "You're so
PROXY ERROR: No response from bot (pgshag2)
right now."Across from him, {{user}} released a breath she didn't know she was holding. This is because, for the rest of the fucking roleplay, the LLM was going to control all her actions.
2๏ธโฃ Give it things to go off of. If you haven't given it room to act, it'll be more prone to writing for you.
โ * I smile cutely *
โ She walks down the street and looks back at John.
โ aah aah mistress
3๏ธโฃ Use a good prompt. Tell the bot what to do, not what not to do.
โ Do not speak for {{user}}. NEVER control {{user}}'s actions. Avoid writing my dialogue for me.
โ You are an assistant and fiction writer. You play as {{char}}, and other NPCs that may appear, along with the world and what happens within it. User plays as their exclusive player character, {{user}}. Focus on {{char}}'s thoughts, feelings, and actions.
4๏ธโฃ Don't control {{char}}'s actions in your messages.
Self-explanatory, I hope.
5๏ธโฃ Every time it impersonates you, edit it out or reroll. Don't let it get away with it.
Allowing it to get away with it, even subtly, will reinforce to the model that it's okay.
6๏ธโฃ Some models are significantly better at listening to instructions than others. Experiment if you can!
Despite being an avid DeepSeek lover, I've personally noticed that Gemini 2.5-Pro is great at following my instructions. It listens well; I almost never have issues with it speaking for me!
But, these tips can be applied to literally every model. โค๏ธ
โญ More in-depth explanation of all this here.
๐ Requests Being Filtered
If you're using a DeepSeek model ๐ณ...
It's hallucinating. DeepSeek โ doesn't really have filters โ like that.
Reroll the message and check your prompt.
If your prompt says anything about guidelines, ignoring guidelines, being 'jailbroken', etcโremove it. It'll make it hallucinate more.
- It's sort of like telling someone
hey, don't think about losing THE GAME
! Well, now you've put that in my mind. Same goes with mentioning content policies and guidelines.
๐ช It's cutting off sentences in the middle.
Open Generation Settings. Turn your max new tokens to 0.
๐ฑ Messages are too short!
- Use a bot with a longer opening message. โ๏ธ
- Write clear โจactions and dialogueโจ for the LLM to work withโsomething like
* and then i kiss her *
isn't going to elicit a great big response. - ๐ Write longer messages, including the kind of detail you'd like to see in the bot's responses.
- Edit short messages to make them your desired lengthโand/or ๐reroll๐ until you get something that suits you!
- Add a line in your Custom Prompt instructing the model to
write detailed, lengthy replies
or something similar. Experiment!
๐ณ Messages are too long!
- Use a bot with a shorter opening message. โ๏ธ
- ๐ Roleplay in a shorter format, keeping your own messages brief, but with actions and dialogue for the bot to build on.
- Edit long messages to make them your desired lengthโand/or ๐reroll๐ until you get something that suits you!
- Reduce your max new tokens (unless you're using a reasoning model) to something smaller. It may cut off; just edit out cut off sentences.
- Add a line in your Custom Prompt instructing the model to
write short, concise replies
or something similar. Experiment!
โ It's speaking gibberish!
Turn your temperature down. Guide on temperature here.
โผ๏ธ It's using way too many asterisks or bolding shit like crazy.
Unfortunately, this is a V3-0324 (and DeepSeek itself)-ism. Try:
- switching models
- using a bot that doesn't have asterisks and/or bolding in the opening message.
- Editing the formatting you dislike out as you go along. Sorry, it sucks, I know.
- Just to clarifyโediting it out will probably be the biggest help. It might not fully eliminate it, but LLMs rely heavily on past messages. Don't let it "get away with it"; you'll only confuse the model.
There's no permanent fix for this. โน๏ธ
๐ค My model is explaining what it's going to do! | What is <think>?
This is something common to reasoning models (like R1 and R1-0528).
It's not a glitch, you can't prompt it out, you can't use OOC commands to stop itโit's just what they're meant to do.
OpenRouter and the Official DeepSeek API hide the reasoning on Janitor. It's still there, it just doesn't show.
Chutes does not hide it.
If you don't like it:
- switch to something that doesn't reason (like V3-0324)
- hold your nose and edit it out
- BYOK (bring your own key) to a service that hides reasoning like OpenRouter
Guides with images here.
๐ณ๏ธโโง๏ธ It can't get my pronouns right???
This happens! Fortunately, it can be fixed.
- Format your persona to include their gender and pronouns. Iorveths has a great guide here.
- Edit every message the bot sends to use the correct pronouns. If you're rerolling the message, don't botherโbut if you're going to accept it, edit it.
- Make sure you're not using a bot where the character's gender is hardcoded in.
โญ I also have some success adding Always refer to {{user}} with she/her pronouns.
in my Custom Prompt...but, I always play characters with she/her pronouns! You wouldn't want this, obviously, if you didn't also do that.
๐ Bot not Listening to Prompts
[Section to be added]
๐๏ธ Not Formatting Properly
[Section to be added]
๐ Ending Scenes Early
[Section to be added]
๐ฅ Losing Character Traits/Out of Character
[Section to be added]
๐คท Forgetting Important Details
[Section to be added]
๐ฎ Reading my Mind/Reacting to Thoughts
[Section to be added]
Full TOC
- โจ AI Models, Prompts, and Errors for Fresh Hatchlings โจ
- Basic Terminology
- Proxy Options & Setup
- ๐จ Important Information
- ๐ณ DeepSeek Models
- โ Gemini Models
- ๐จ Claude Models
- ๐ Local Models
- ๐ค Proxy Options
- Generation Settings
- Prompting
- ๐ตโ๐ซ Do I really need a prompt?
- ๐ What's the best prompt?
- โ Where do I put the prompt?
- ๐ How long should a prompt be?
- ๐ What's negative prompting? Is it bad?
- โ๏ธ Do I need a certain format?
- ๐คทโโ๏ธ Are the default prompts any good?
- ๐ฅ๏ธ Can ChatGPT write my prompt for me?
- ๐ฎ Can I put prompts in chat memory?
- โญ What's OOC mean?
- ๐ How should I use OOC commands?
- โป๏ธ Can I switch prompts in the middle of a roleplay?
- General Questions
- ๐ณ Troubleshooting Guide
- ๐ฅ Error While Chatting
- 1๏ธโฃ First Steps
- 2๏ธโฃ Otherwise...
- 3๏ธโฃ Okay, almost done
- ๐ General Error Table
- ๐ General Errors
- ๐ OpenRouter Errors
- OpenRouter 400 (Invalid Model)
- OpenRouter 401 (No Auth)
- OpenRouter 402 (Insufficient Credits)
- OpenRouter 402 (Quota Exceeded)
- OpenRouter 404 (No Providers)
- OpenRouter 404 (No Endpoints Found)
- OpenRouter 404 (Model Not Found)
- OpenRouter 408 (Timeout)
- OpenRouter 429 (Daily Limit)
- OpenRouter 429 (Upstream Limit)
- OpenRouter 502 (Bad Gateway)
- OpenRouter Provider Error
- ๐ช Chutes Error Table
- ๐ช Chutes Errors
- โน๏ธ Writing Incorrectly | Messages are Weird
- Table of Contents
- ๐ Bot Speaking for Me
- 1๏ธโฃ Use a bot with an opening message that does not include any actions that {{user}} takes.
- 2๏ธโฃ Give it things to go off of. If you haven't given it room to act, it'll be more prone to writing for you.
- 3๏ธโฃ Use a good prompt. Tell the bot what to do, not what not to do.
- 4๏ธโฃ Don't control {{char}}'s actions in your messages.
- 5๏ธโฃ Every time it impersonates you, edit it out or reroll. Don't let it get away with it.
- 6๏ธโฃ Some models are significantly better at listening to instructions than others. Experiment if you can!
- ๐ Requests Being Filtered
- ๐ช It's cutting off sentences in the middle.
- ๐ฑ Messages are too short!
- ๐ณ Messages are too long!
- โ It's speaking gibberish!
- โผ๏ธ It's using way too many asterisks or bolding shit like crazy.
- ๐ค My model is explaining what it's going to do! | What is <think>?
- ๐ณ๏ธโโง๏ธ It can't get my pronouns right???
- ๐ Bot not Listening to Prompts
- ๐๏ธ Not Formatting Properly
- ๐ Ending Scenes Early
- ๐ฅ Losing Character Traits/Out of Character
- ๐คท Forgetting Important Details
- ๐ฎ Reading my Mind/Reacting to Thoughts
- Full TOC
- ๐ Bot Speaking for Me
- Table of Contents