If you’ve been looking for a platform where you can chat with AI characters that actually feel real, Janitor AI might be what you need. I’ve spent months testing this thing, and honestly, it’s changed how I think about talking to bots.
Janitor AI is a chatbot platform built specifically for character-based roleplay and interactive storytelling. Unlike the typical bots you’d use for customer service or work stuff, this one focuses entirely on creative expression and immersive conversations. You can create your own characters, customize their personalities down to the smallest detail, and have conversations that feel surprisingly human.
What I really like about Janitor AI is its flexibility. The platform now supports both legacy models like GPT-3.5 and GPT-4, as well as the latest AI models including GPT-5, GPT-5.1, GPT-5.2, Gemini 3, Claude Opus 4, Claude 4.5 Sonnet, Claude 4.5 Haiku, DeepSeek V3, and many others released throughout 2025 and early 2026. This means you’re not stuck with one AI brain—you can switch between different models depending on what you need and your budget.
In 2026, the platform has grown quite a bit. The introduction of JanitorLLM (their free model), better memory management, and a thriving community of character creators has made it one of the most popular AI roleplay platforms around. According to HackerNoon, about 70% of users are women, which tells you something about how welcoming the community is.
Whether you’re a writer testing dialogue for your novel, a language learner practicing conversations, or someone who just enjoys creative roleplay, Janitor AI has tools that actually work. Let me walk you through everything you need to know.
What is Janitor AI and How Does It Work?
Here’s the thing about Janitor AI—it’s not actually an AI model itself. Think of it more like a really smart interface that connects you to powerful language models while giving you complete control over how characters behave.
When you use Janitor AI, you’re working with three layers. First, there’s the character definition layer where you describe your character’s personality, background, and speaking style. Then there’s the platform layer that manages the conversation flow and remembers what’s been said. Finally, there’s the AI model layer that generates the actual responses.
From a practical standpoint, here’s how it works. You pick or create a character from the community library. Each character comes with defined traits like personality markers, backstory elements, and behavioral guidelines. Once you’ve got your character, you connect it to your preferred language model through API integration.
The system uses these persona definitions to keep character behavior consistent throughout your conversations. When you send a message, Janitor AI processes your input alongside the character’s established traits and conversation history. The connected LLM then generates a response that matches the character’s personality and the ongoing story.
Most importantly, Janitor AI doesn’t train its own language models. Instead, it acts as a bridge between you and existing LLM providers like OpenAI, Anthropic, Google, and others. This approach gives you flexibility in choosing your preferred AI backend while keeping the user experience consistent across different models.
The platform launched in June 2023 by Jan Zoltkowski and pulled in over one million users within its first week. That’s pretty wild when you think about it. The rapid growth came mainly from social media and roleplay communities who were hungry for something more flexible than what was available.
Technically, the platform works by maintaining a conversation context that includes your character definitions, chat history, and user preferences. When you send a message, this entire context gets packaged together with your input and sent to whichever AI model you’ve connected. The model processes everything and returns a response that the platform then displays to you.
If you’re interested in exploring other AI-powered creative tools, there are plenty of options worth checking out in 2026.
Latest AI Models Supported in 2026
One of Janitor AI’s biggest advantages is staying current with the latest AI models. As of early 2026, the platform supports an impressive range of both legacy and cutting-edge language models that cater to different needs and budgets. Let me break down what’s available and what each model actually costs.
OpenAI’s Model Lineup
Legacy GPT Models (Still Available):
OpenAI’s older models remain available and are actually perfect for users on a tighter budget. GPT-3.5 Turbo still works well for basic conversations. GPT-4, GPT-4 Turbo, and GPT-4o from 2023-2024 continue to deliver solid performance for roleplay scenarios.
From my experience, GPT-4o (the older version) still produces excellent dialogue quality and maintains character consistency pretty well. If you’re just getting started with paid models, this is a good entry point before jumping to the newer, more expensive options.
Current pricing for legacy OpenAI models (per 1M tokens):
- GPT-3.5 Turbo: $3.00 input / $6.00 output
- GPT-4: $30.00 input / $60.00 output
- GPT-4 Turbo: $10.00 input / $30.00 output
- GPT-4o: $2.50 input / $10.00 output
- GPT-4o mini: $0.15 input / $0.60 output
GPT-5 Series (Released Late 2025):
OpenAI released GPT-5 in late 2025, and it’s a massive upgrade from GPT-4. The model understands context much better, maintains character personalities more consistently, and produces more natural-sounding dialogue. I’ve been using it for about three months now, and the difference is noticeable.
GPT-5 handles longer conversations without losing track of details. In my testing, I had a roleplay session that spanned over 50,000 words across multiple days, and the AI still remembered minor details I mentioned at the very beginning. That’s impressive.
The model also supports a context window of up to 400,000 tokens, which means it can remember roughly 300,000 words of conversation history. For comparison, GPT-4o maxed out at around 128,000 tokens.
GPT-5 Series Pricing (per 1M tokens):
- GPT-5: $1.25 input / $10.00 output (with cached input at $0.125)
- GPT-5 mini: $0.25 input / $2.00 output (great for high-volume, budget-conscious use)
- GPT-5 nano: $0.05 input / $0.40 output (fastest and cheapest option)
GPT-5.1 (Released December 2025):
This iteration focused specifically on creative writing and roleplay scenarios. The model better understands narrative structure, character development arcs, and emotional nuance. If you’re using Janitor AI for serious creative writing projects, GPT-5.1 is worth the extra cost.
One thing I noticed immediately with GPT-5.1 is how it handles character emotions. The AI doesn’t just state that a character is sad or angry—it shows it through subtle dialogue cues, body language descriptions, and behavioral changes. It feels more sophisticated than previous versions.
GPT-5.1 Pricing (per 1M tokens):
- GPT-5.1: $1.25 input / $10.00 output (with cached input at $0.125)
GPT-5.2 (Released January 2026):
The latest version as of this writing. GPT-5.2 introduced better memory management and reduced repetitive responses—a common complaint with earlier models. OpenAI also improved the model’s ability to handle multiple characters in a single conversation without getting confused about who’s speaking.
From a practical perspective, GPT-5.2 excels at agentic tasks and long-context scenarios. It’s particularly good for complex roleplay scenarios where multiple characters interact simultaneously.
GPT-5.2 Pricing (per 1M tokens):
- GPT-5.2: $1.75 input / $14.00 output (with cached input at $0.175)
- GPT-5.2 Pro: Higher tier with enhanced capabilities (pricing varies by usage tier)
For comparison, GPT-5.2 is about 1.5-2x more expensive than GPT-5, but the quality jump is significant if you’re doing heavy roleplay or creative writing. The cached input feature is particularly useful—if you’re reusing the same character definitions across multiple chats, the cached portion costs 90% less.
Anthropic’s Claude Models
Anthropic has been a serious competitor in the AI space, and their Claude models are particularly popular with Janitor AI users who want longer context windows and more nuanced conversations.
Legacy Claude Models:
- Claude 3.5 Sonnet: $3.00 input / $15.00 output per 1M tokens
- Claude 3 Opus: $15.00 input / $75.00 output per 1M tokens
- Claude 3 Haiku: $0.25 input / $1.25 output per 1M tokens
These older models still work great for many users. Claude 3.5 Sonnet, in particular, offers excellent performance at a reasonable price point.
Claude 4 Series (Released 2025):
- Claude Opus 4: $15.00 input / $75.00 output per 1M tokens
- Claude Sonnet 4: $3.00 input / $15.00 output per 1M tokens (with cached input at $0.75 / $3.75)
Claude 4.5 Series (Released Late 2025 / Early 2026):
The Claude 4.5 series represents Anthropic’s latest push into competitive AI models. These models offer significant improvements in reasoning, context handling, and creative writing capabilities.
- Claude 4.5 Haiku: $1.00 input / $5.00 output per 1M tokens
- Claude 4.5 Sonnet: $3.00 input / $15.00 output per 1M tokens (with cached input at $0.75 / $3.75)
- Claude Opus 4.5: $5.00 input / $25.00 output per 1M tokens
From my testing, Claude 4.5 Sonnet hits the sweet spot between cost and performance. It’s particularly good at maintaining emotional depth in longer conversations. I’ve had roleplay sessions running 60,000+ words where Claude still picked up on subtle emotional cues from early in the conversation.
Claude 4.5 Haiku is the budget option but still performs admirably. It’s faster than Sonnet and significantly cheaper, making it ideal for high-volume chatting where you don’t need the absolute best quality.
Claude Opus 4.5 is the premium option with the most advanced reasoning capabilities. Honestly, for most roleplay scenarios, Sonnet offers better value unless you’re working on extremely complex narrative structures with multiple interwoven plot lines.
Worth noting: Claude models typically offer longer context windows than OpenAI equivalents. Claude can handle up to 200,000 tokens in a single conversation, which is about 150,000 words. This makes it excellent for long-term roleplay campaigns.
Google’s Gemini Models
Google has been aggressive with both pricing and capabilities in their Gemini lineup. These models have become increasingly popular on Janitor AI, especially for users who want good performance without breaking the bank.
Gemini 2.5 Series:
- Gemini 2.5 Flash: $0.30 input / $2.50 output per 1M tokens (up to 200K context)
- Gemini 2.5 Pro: $1.25 input / $10.00 output per 1M tokens (up to 200K context)
- Gemini 2.5 Pro (>200K context): $2.50 input / $15.00 output per 1M tokens
Gemini 2.5 Flash is one of the most cost-effective options available. At just $0.30 per million input tokens, it’s significantly cheaper than most competitors while still delivering solid performance. I’ve used it for casual roleplay sessions and found it perfectly adequate for most scenarios.
Gemini 3 Series (Released Late 2025):
Google’s latest flagship models, released in Q4 2025, represent a significant leap in capabilities.
- Gemini 3 Flash: $0.50 input / $3.00 output per 1M tokens
- Gemini 3 Pro (≤200K context): $2.00 input / $12.00 output per 1M tokens
- Gemini 3 Pro (>200K context): $4.00 input / $18.00 output per 1M tokens
In practice, Gemini 3 Pro competes directly with GPT-5 and Claude Opus 4.5. The model excels at understanding complex character dynamics and generating creative, unexpected responses that still feel true to the character’s personality.
What’s cool about Gemini 3 is how it handles ambiguity. If you give it a vague prompt, it doesn’t just make random guesses—it asks clarifying questions or interprets based on the character’s established personality. This feels more natural than older models that would just barrel ahead with assumptions.
The free tier for Gemini models is also worth mentioning. Google offers limited free access through their AI Studio platform, which can be useful for testing before committing to paid API usage.
DeepSeek Models
DeepSeek has emerged as the budget-friendly option that doesn’t sacrifice too much quality. Their models are significantly cheaper than Western competitors, making them popular for users who chat extensively.
DeepSeek V3 Series:
- DeepSeek V3: $0.14 input / $0.28 output per 1M tokens (with 64K context)
- DeepSeek V3 0324: $0.19 input / $0.87 output per 1M tokens
- DeepSeek V3.1: Supports both thinking and non-thinking modes (pricing varies)
- DeepSeek V3.2-Exp: $0.028 input / $0.42 output per 1M tokens (experimental version with half the cost)
The pricing here is honestly wild. DeepSeek V3.2-Exp costs less than 3 cents per million input tokens. That’s roughly 95% cheaper than GPT-5 and about 97% cheaper than Claude Opus 4.5.
In my testing, DeepSeek models perform well for straightforward roleplay scenarios. They’re not as sophisticated as GPT-5.2 or Claude Opus 4.5 when it comes to emotional nuance or complex narrative structures, but for the price, they’re incredible value.
One thing to keep in mind: DeepSeek models sometimes produce slightly less polished prose compared to OpenAI or Anthropic models. You might notice more grammatical quirks or awkward phrasings, especially in longer conversations. However, for casual chatting or high-volume usage, the cost savings more than make up for these minor issues.
DeepSeek-R1 (Reasoning Model):
- DeepSeek-R1: $0.28 input (cache miss) / $0.028 input (cache hit) / $0.42 output per 1M tokens
The R1 model includes reasoning capabilities, meaning it can show its “thought process” before generating responses. This is particularly useful for complex problem-solving or when you want the AI to explain its character interpretation decisions.
Other Notable Models
KoboldAI (Self-Hosted):
KoboldAI isn’t a single model but rather a platform for running open-source models locally. Popular options include Pygmalion, LLaMA variants, and other community-fine-tuned models.
The main advantage is zero ongoing costs once you’ve set up your local infrastructure. You need a decent GPU (at least 6-8 GB VRAM), but after the initial hardware investment, you can chat unlimited without worrying about API bills.
I tried running KoboldAI with a Pygmalion 13B model on my RTX 3080. It worked, but the setup was tedious and the response quality didn’t match GPT-4o, let alone GPT-5. That said, if you’re technically inclined and want complete control over your data and costs, it’s a viable option.
JanitorLLM (JLLM Beta):
This is Janitor AI’s proprietary free model. It requires no API setup and provides unlimited conversations at zero cost.
JLLM holds about 8,000-9,000 tokens in memory (roughly 6,000-7,000 words). It’s decent for casual use but has limitations:
- Occasional character inconsistency
- Performance degrades during peak hours
- Less nuanced emotional understanding compared to premium models
- Shorter context memory
That said, JLLM is constantly being improved. The team announced plans for JLLM v2 with better performance and consistency. They’re exploring advertising as a funding mechanism to keep the model free while upgrading its capabilities.
Key Features of Janitor AI
Now that you understand which models are available, let’s talk about the platform features that make Janitor AI different from other chatbot platforms. These are the tools you’ll actually use every day.
Character Creation and Customization
The platform lets you build detailed AI personas from scratch. You can define physical appearance, personality traits, speech patterns, and background stories. This level of customization enables highly specific roleplay scenarios that feel authentic.
You can create characters ranging from historical figures to original fictional personas. The character definition uses natural language rather than code. You simply describe how your character should act, what topics they care about, and how they should respond to different situations.
From my experience, well-defined characters produce more engaging and consistent conversations. The platform translates these instructions into behavioral patterns that the AI follows throughout interactions. I’ve created characters for everything from Victorian-era detectives to futuristic space explorers, and the system handles them all pretty well.
Here’s a practical example. When I created a character based on a grumpy librarian, I included details like “speaks in short, clipped sentences,” “deeply knowledgeable about rare books,” and “secretly enjoys helping people but hides it behind sarcasm.” The AI picked up on these nuances and maintained them throughout hours of conversation.
The character creation form includes several fields:
- Character image (required): Upload a picture representing your character
- Character name (required): Give your character a memorable name
- Character bio (highly recommended): Describe personality, appearance, background, and motivations
- Character tags: Add searchable descriptors like “fantasy,” “horror,” “romantic”
- Visibility settings: Choose public, unlisted, or private access
- NSFW toggle: Mark whether your character involves adult content
- First message: Write the opening message users receive
- Example dialogues: Provide sample conversations demonstrating the character’s voice
The more detail you provide, especially in the bio and example dialogues, the better your character performs. I usually spend 15-20 minutes crafting a good character definition. It makes a huge difference in conversation quality.
Multi-Model Support and Flexibility
One feature I didn’t expect to appreciate as much as I do is the ability to switch between different AI models mid-conversation. You’re not locked into a single model choice for your entire Janitor AI experience.
Let’s say you start a conversation using JLLM (free) just to test out a new character concept. Once you’ve confirmed the character works well, you can switch to GPT-5.1 or Claude 4.5 Sonnet for higher-quality responses. The platform maintains your conversation history and character definitions across model switches.
This flexibility extends to cost management too. For casual, low-stakes conversations, you might use JLLM or DeepSeek V3. When working on serious creative writing where quality really matters, you switch to GPT-5.2 or Claude Opus 4.5. This approach lets you balance quality against budget effectively.
The platform currently supports API connections to:
- OpenAI (GPT-3.5 through GPT-5.2)
- Anthropic (Claude 3 through Claude 4.5)
- Google (Gemini 2.5 and Gemini 3)
- DeepSeek (V3, V3.1, V3.2, R1)
- OpenRouter (access to multiple providers through a single API)
- KoboldAI (local self-hosted models)
- JanitorLLM (free proprietary model)
Immersive Mode
Immersive Mode transforms the chat interface into a more story-focused experience. When enabled, it removes editing controls and displays responses in a streamlined format. This keeps your focus on the conversation flow rather than interface elements.
The mode also enables text streaming in supported configurations. Instead of responses appearing all at once, you see them generate word by word, like someone typing in real time. This makes interactions feel more dynamic and natural.
To enable Immersive Mode, open any chat and look for the settings menu (typically a three-bar icon in the top right). Toggle the option on, and your conversation interface will shift to the immersive layout.
Honestly, once I started using Immersive Mode, I never went back. It makes such a big difference in how engaged you feel with the conversation. The streaming text effect might seem like a small thing, but it really does change the experience. You’re not just reading completed thoughts—you’re watching them form in real time, which creates a sense of presence that static text blocks can’t match.
Memory Management and Context Window
Janitor AI includes conversation memory features that help maintain context over long chats. The system remembers previous interactions, character details, and ongoing plot points. This creates continuity in your roleplay sessions.
Actually, memory capacity depends entirely on which model you’re using. Here’s what you can expect:
- JLLM: 8,000-9,000 tokens (about 6,000-7,000 words of history)
- GPT-3.5 Turbo: 16,000 tokens
- GPT-4o: 128,000 tokens (roughly 96,000 words)
- GPT-5 series: Up to 400,000 tokens (around 300,000 words)
- Claude models: Up to 200,000 tokens (about 150,000 words)
- Gemini models: Up to 200,000 tokens depending on the variant
In practical terms, this means the AI can remember details from earlier in your conversation and reference them naturally. If you mentioned your character has a fear of heights in the first five minutes of chatting, models like GPT-5 or Claude will remember that three hours later when you describe climbing a tower.
You can also manage memory manually through the chat settings. There’s a “Chat Memory” option where you can:
- View what the AI currently remembers
- Edit specific details or correct misunderstandings
- Add important information you want the AI to prioritize
- Clear memory entirely if you want to start fresh
One technique I’ve found useful is periodically reviewing the chat memory to make sure the AI hasn’t picked up any wrong details. Sometimes the AI misinterprets something you said early in the conversation, and that misunderstanding gets reinforced over time. Catching and correcting these early saves a lot of confusion later.
Community Character Library
One feature I didn’t expect to use as much as I do is the community character library. There are literally millions of user-created characters available. You can browse by tags, search for specific types, or explore trending characters.
The library includes everything from anime characters and fictional personalities to historical figures and original creations. Each character page shows you the creator’s name, character description, example dialogues, and user ratings. You can also see how many chats the character has had, which gives you a sense of popularity.
What’s cool is you can use these characters as-is or fork them to create your own modified versions. I’ve taken community characters and tweaked their personalities to better fit my specific roleplay scenarios. This saves a ton of time compared to building everything from scratch.
Popular character categories include:
- Fantasy roleplay (dragons, elves, wizards, medieval settings)
- Romantic companions (various personalities and relationship dynamics)
- Horror scenarios (creepy characters, psychological thrillers)
- Anime/manga characters (both canon and original)
- Historical figures (for educational or creative purposes)
- Productivity coaches (surprisingly popular for accountability)
- Language tutors (for practice conversations)
The search and filtering system works pretty well. You can filter by:
- Content rating (SFW vs NSFW)
- Character type (human, fantasy creature, robot, etc.)
- Scenario type (adventure, romance, horror, slice-of-life)
- Popularity metrics (most chatted, highest rated, trending)
- Creator (follow your favorite character creators)
From my observation, the most successful community characters share a few traits: detailed backstories, clear personality markers, well-written example dialogues, and engaging opening messages. If you’re creating characters to share publicly, these elements really matter.
Text Formatting and Advanced Commands
Janitor AI supports various text formatting options and special commands that enhance your roleplay experience. You can use markdown-style formatting for emphasis, create action descriptions, and use specific commands to control AI behavior.
Most users develop their own formatting style, but here are some common conventions:
- Dialogue: Use quotation marks – “I think we should head north,” she suggested.
- Actions: Use asterisks – He reached for his sword, eyes narrowing suspiciously.
- Thoughts: Use italics or parentheses – (Why does this place feel so familiar?)
- Narration: Plain text without special formatting
The AI learns from your formatting style and typically mirrors it back. If you consistently use asterisks for actions, the AI will do the same. This creates a consistent reading experience throughout the conversation.
There’s also a command system for power users, though it’s not officially documented everywhere. The community has put together guides on commands like:
- Adjusting response length
- Changing narrative perspective (first-person vs third-person)
- Controlling tone (serious vs humorous)
- Triggering specific plot events
These commands vary by which AI model you’re using, so you’ll need to experiment to see what works with your chosen backend.
Profile Customization with CSS
If you’re into customization, Janitor AI lets you modify your profile appearance using CSS. This means you can change colors, fonts, layouts, and visual elements to match your personal style.
The CSS customization feature appeals mainly to users with some web development knowledge, but you don’t need to be an expert. The community has shared tons of CSS templates that you can copy and paste directly into your profile settings.
I’ve seen profiles with custom backgrounds, animated elements, and completely redesigned layouts. It’s a nice touch that makes the platform feel more personal. If you’re interested in web development and CSS styling, this feature offers a fun playground to experiment.
To customize your profile:
- Go to your profile settings
- Find the “Custom CSS” section
- Paste your CSS code
- Preview the changes
- Save when you’re happy with the result
The community subreddit and Discord server have dedicated channels where users share CSS snippets and templates. This is honestly one of the best resources if you want to customize your profile but don’t know where to start.
Mobile App (Beta)
In 2026, Janitor AI finally released official mobile apps for both Android and iOS. The apps are still in beta, but they’re functional and offer most of the features available on the web version.
The mobile experience is optimized for on-the-go chatting. The interface is touch-friendly, with easy access to your favorite characters, quick model switching, and streamlined chat controls. Immersive Mode works particularly well on mobile devices since the smaller screen naturally focuses your attention on the conversation.
One limitation I’ve noticed: complex character creation is still easier on desktop. The mobile app is great for chatting with existing characters, but building detailed characters with extensive bios and example dialogues feels cramped on a phone screen.
You can download the Janitor AI Beta app from:
- Google Play Store (for Android devices)
- Apple App Store (for iOS devices)
The app syncs with your web account, so your characters, chats, and settings carry over seamlessly between devices. I often start conversations on my desktop and continue them on my phone during commute time.
Understanding JanitorLLM (JLLM)
Let’s talk about JanitorLLM in detail, because this is probably the most important thing to understand if you want to use Janitor AI without spending money.
JLLM is Janitor AI’s proprietary language model that launched in beta. It’s completely free to use, requires zero API setup, and provides unlimited conversations. You just select “Janitor LLM” in your API settings, and you’re good to go.
What Makes JLLM Different
JLLM is specifically engineered to provide natural language processing services without requiring a paid subscription. It uses advanced machine-learning techniques to understand and generate human-like text based on your input.
From an implementation perspective, JLLM works differently than connecting external APIs. The model runs on Janitor AI’s own servers, which means you don’t need to sign up for OpenAI or Anthropic accounts. You don’t need to manage API keys or worry about token costs.
In my testing, JLLM handles basic conversations pretty well. It understands context, maintains character personalities reasonably well, and generates coherent responses most of the time. For casual roleplay or creative writing practice, it gets the job done.
The model was designed with the community in mind. Jan Zoltkowski and his team wanted to ensure that anyone could access Janitor AI regardless of their budget. JLLM makes that possible.
JLLM Limitations You Should Know
Here’s the truth—JLLM isn’t as sophisticated as premium models like GPT-5 or Claude Opus 4.5. You’ll notice some limitations:
- Response quality varies during peak hours: When lots of users are online, JLLM can produce repetitive or lower-quality responses
- Shorter context memory: The model holds about 8,000-9,000 tokens, which is less than premium alternatives
- Occasional character inconsistency: Users on Reddit have reported issues with misgendering, writing for the user instead of staying in character, or making characters behave out of character
- Less nuanced emotional understanding: JLLM sometimes misses subtle emotional cues or produces responses that feel slightly off-tone
- Performance degradation: Some users have noticed JLLM’s quality has declined in recent months, though the team is actively working on improvements
- Repetitive phrasing: JLLM sometimes falls into repetitive patterns, especially in longer conversations
- Limited reasoning capabilities: Complex problem-solving or multi-step planning isn’t JLLM’s strength
That said, JLLM is constantly being improved. The development team has announced plans for JLLM v2, which should address many of these limitations. They’re exploring advertising as a potential way to fund the free model’s development while keeping it accessible to everyone.
According to announcements on Janitor AI’s newsroom, the team is working on:
- Improved consistency in character behavior
- Better handling of pronouns and gender
- Reduced repetitiveness in responses
- Enhanced emotional intelligence
- Longer context windows
When to Use JLLM vs Premium Models
Based on my experience, here’s when each option makes sense:
Use JLLM when:
- You’re just starting out and want to explore the platform
- You’re having casual, low-stakes conversations
- Budget is a concern and you don’t want ongoing costs
- You’re testing character definitions before committing to paid API usage
- You’re okay with occasional quality fluctuations
- You’re in scenarios where perfect consistency doesn’t matter
Use premium models when:
- You need consistent, high-quality responses
- You’re working on serious creative writing projects
- Long-term memory and context retention matter for your use case
- You want the best possible character consistency
- You’re willing to pay for better performance
- Emotional nuance and subtlety are important to your roleplay
- You’re running extended campaigns with complex plot lines
Honestly, I recommend starting with JLLM to get familiar with the platform. Create a few characters, experiment with different roleplay scenarios, and see how the system works. Once you understand the basics and know you’ll use Janitor AI regularly, then decide whether upgrading to a premium model is worth it for your specific needs.
From a practical standpoint, many users end up using a hybrid approach: JLLM for testing and casual chats, premium models for serious projects. This lets you enjoy Janitor AI without constantly worrying about API costs.
How to Set Up and Use Janitor AI
Let me walk you through the actual setup process step by step. I’ll break this down so anyone can follow, even if you’re not particularly tech-savvy.
Creating Your Janitor AI Account
First, you need to create an account. Head over to janitorai.com and click the “Register” button in the upper right corner. You’ve got a few options here:
- Sign up with your email address and create a password
- Use your Google account for quick registration
- Connect through Discord if you’re already active in the community
- Sign up via X (formerly Twitter)
I went with the Google option because it was faster, but any method works fine. The whole process takes about 30 seconds.
Once you’re logged in, take a minute to set up your profile. Add a username, upload a profile picture if you want, and fill out your user description. This isn’t strictly necessary, but it helps AI characters understand how to interact with you.
In the profile settings, you can add details like:
- Your preferred name
- Physical description (if you want characters to “see” you a certain way)
- Personality traits
- Preferences for roleplay scenarios
The more information you provide, the better characters can tailor their responses to you. That said, many users skip this step entirely and just jump into chatting. Both approaches work.
Choosing Your First Character
After registration, you’ll see the main interface with thousands of characters to choose from. Don’t get overwhelmed by all the options. Here’s how to find something good:
Browse by categories: The platform organizes characters into segments like Fantasy, Anime, Horror, Romance, Game, Male, Female, and more. Pick a category that interests you.
Check trending characters: These are currently popular with other users and tend to be well-made. The trending section updates regularly based on chat volume and user ratings.
Read character descriptions: Click on a character to see their full profile, including personality traits, backstory, example dialogues, and the creator’s notes. This gives you a sense of whether the character matches what you’re looking for.
Look at user ratings and chat counts: Characters with higher ratings and more chats usually deliver better experiences. A character with 50,000 chats and 4.8 stars is probably a safer bet than one with 50 chats and no ratings.
For your first character, I’d recommend picking something simple and well-reviewed. Don’t go for the most complex multi-character scenarios right away. Start with a straightforward one-on-one conversation to get comfortable with how the platform works.
Some good starter character types:
- Friendly NPCs (tavern keepers, shopkeepers, guides)
- Tutorial characters specifically designed for new users
- Simple romantic interests with clear personality traits
- Adventure companions for fantasy scenarios
Avoid as your first character:
- Multi-character scenarios with complex interactions
- Characters with extensive lore you need to know
- Highly specific fetish or niche content
- Characters requiring deep knowledge of a particular fandom
Configuring API Settings (The Free Way)
Once you’ve picked a character, you need to configure your API settings before you can start chatting. Here’s the easiest way to get started for free:
- Click on the character you want to chat with
- Look for the three-bar menu button (hamburger icon) in the top right corner
- Select “API Settings”
- In the dropdown menu, choose “Janitor LLM”
- That’s it—you’re done!
With JanitorLLM selected, you can start chatting immediately without any additional setup. No API keys, no payment information, no complicated configuration. Just pick your character and start typing.
The interface will remember your API selection, so you don’t need to do this every time. You can switch between different API options anytime through the same menu.
Starting Your First Conversation
Now for the fun part. With your character selected and API configured, click the “Start a new chat” button. Most characters will send you an opening message automatically. This sets the scene and gives you context for the conversation.
Your first message matters more than you might think. Instead of just saying “hello,” try to engage with the scenario the character presented. If the character opened with a scene description, respond to what’s happening in that scene.
Here’s a practical example. If the character’s opening message is:
“You walk into a dimly lit tavern. A hooded figure at the corner table motions for you to sit down. Rain patters against the windows, and the smell of ale fills the air.”
Don’t just respond with “hi.” Instead, try something like:
“I cautiously approach the table, keeping one hand near my weapon. Water drips from my cloak onto the wooden floor. ‘You wanted to see me?’ I ask, sliding into the seat across from the hooded figure.”
From my testing, more detailed responses generally produce better AI replies. The AI picks up on your writing style and matches it. If you write short, simple sentences, you’ll get short, simple responses. If you write detailed, descriptive paragraphs, the AI will do the same.
Another tip: include sensory details and character actions, not just dialogue. This gives the AI more to work with and creates a richer roleplay experience.
Using the Chat Interface
The chat interface has several features worth knowing about:
Message editing: Hover over your previous messages and click the edit icon (usually a pencil). This is useful if you made a typo or want to rephrase something to get a better response.
Message regeneration: If you don’t like how the AI responded, look for the regenerate button (usually a circular arrow icon). The AI will create a different response based on the same prompt. You can regenerate multiple times until you get something you like.
Message deletion: You can delete messages from the conversation history. This is helpful if the chat went in a direction you didn’t like. Just select the message and choose delete.
Chat memory options: Access this through the menu to see what details the AI currently remembers about your conversation. You can edit these details or add new ones manually.
Temperature settings: Adjust how creative or predictable the AI’s responses are. The temperature slider usually ranges from 0 to 1:
- Lower temperature (0.3-0.5) = more focused and predictable responses
- Medium temperature (0.6-0.8) = balanced creativity and consistency
- Higher temperature (0.9-1.0) = more creative and random, but potentially inconsistent
Generation settings: Some models let you adjust other parameters like:
- Response length (short, medium, long)
- Repetition penalty (reduces repetitive phrases)
- Top P sampling (controls response diversity)
- Frequency penalty (discourages repeating the same topics)
One tip I learned the hard way: save your chat periodically if it’s important to you. While Janitor AI does auto-save, I’ve had a few instances where chats got lost due to browser issues or connection problems. You can export chats as text files through the chat menu.
Creating Your Own Character
Once you’re comfortable with the platform, you might want to create your own character. Click “Create a Character” in the upper right corner of the main page.
You’ll see a comprehensive form with several fields:
Character image (required): Upload a picture that represents your character. This can be:
- AI-generated image from tools like Midjourney or Stable Diffusion
- Commissioned artwork
- Stock photos (for realistic characters)
- Anime/manga artwork (ensure you have rights to use it)
The image should be appropriate and clear. Most successful characters use high-quality, visually appealing images.
Character name (required): Give your character a name that’s memorable and fitting for their personality. Consider:
- Cultural/historical context if relevant
- Nickname or title (e.g., “Marcus ‘The Iron Fist’ Blackwood”)
- Whether the name suggests personality traits
Character bio (optional but highly recommended): This is where you really define your character. Include:
Physical description: Age, height, build, distinctive features, clothing style
Personality traits: Core characteristics, quirks, strengths, flaws
Background story: Origin, significant life events, motivations
Speech patterns: How they talk (formal, casual, uses slang, has an accent)
Likes and dislikes: What they enjoy or avoid
Goals and fears: What drives them and what they’re afraid of
Relationships: Important people in their life
The more specific you are, the better. Instead of “friendly,” try “greets everyone warmly and remembers small details about people she’s met.” Instead of “tough,” try “survived years as a mercenary and rarely shows vulnerability, but has a soft spot for stray animals.”
Character tags (required): Add relevant tags so users can find your character:
- Genre tags (fantasy, sci-fi, horror, romance, slice-of-life)
- Character type tags (warrior, mage, detective, teacher)
- Scenario tags (adventure, mystery, comedy, drama)
- Content tags (SFW, NSFW if applicable)
Visibility settings: Choose whether your character is:
- Public: Anyone can find and use it
- Unlisted: Only people with the direct link can access it
- Private: Only you can use it
NSFW toggle: Mark whether your character involves adult content. This affects who can see it in search results and is required by the platform’s content policy.
First message: Write the opening message your character will send to users. This is crucial because it sets the tone for the entire conversation. Good first messages:
- Establish the setting and context
- Show the character’s personality
- Give the user something to respond to
- Create intrigue or emotional engagement
Example of a good first message:
“The tavern door slams open, letting in a gust of cold wind. I look up from polishing a glass and size you up with experienced eyes. ‘We don’t get many strangers here, especially not ones who look like they’ve been through hell.’ I set the glass down and lean against the bar. ‘You need a drink, information, or trouble? Because I can provide all three.'”
Example dialogues (optional but recommended): Provide 3-5 sample conversations that demonstrate how your character talks and behaves. Format them like this:
{{user}}: What brings you to this town?
{{char}}: *leans back in chair, arms crossed* Unfinished business. The kind that doesn't ask politely. You asking because you're curious or because you're involved?
Example dialogues teach the AI your character’s voice more effectively than any amount of description. They show rather than tell.
The more detail you provide, especially in the bio and example dialogues, the better your character will perform. I usually spend 15-20 minutes crafting a good character definition for characters I plan to use seriously. For quick experimental characters, 5-10 minutes is fine.
After filling out the form, click “Create Character.” The platform will process your character and make it available based on your visibility settings. You can always go back and edit your character later if you want to refine the definition.
Setting Up Premium API Options
While JanitorLLM is free and works fine for casual use, you might eventually want to upgrade to premium models for better conversation quality. Let me show you how to set up the most popular paid options.
OpenAI API Setup
OpenAI’s models (GPT-3.5, GPT-4, GPT-5 series) are among the most popular choices for Janitor AI users. Here’s how to get started:
Step 1: Create an OpenAI Account
Go to platform.openai.com and sign up for an account. You’ll need to provide:
- Email address
- Phone number for verification
- Payment method (credit card)
OpenAI used to offer $5 in free trial credits for new accounts, but this policy has changed over time. Check their current promotional offers when you sign up.
Step 2: Add Billing Information
Before you can generate an API key, you need to add a payment method:
- Go to Settings > Billing
- Add a credit card
- Set up a spending limit if you want (recommended for cost control)
- Choose between prepaid credits or post-paid usage
I recommend setting a monthly spending limit when you’re starting out. This prevents surprise bills if you accidentally leave a chat running or use an expensive model excessively.
Step 3: Generate an API Key
After setting up billing:
- Click on your profile in the top right corner
- Select “API Keys” from the dropdown menu
- Click “Create new secret key”
- Give your key a descriptive name like “Janitor AI Access”
- Leave permissions set to “All” (or customize if you know what you’re doing)
- Click “Create secret key”
- Copy the generated key immediately (you won’t be able to see it again)
Store your API key somewhere secure. I use a password manager like 1Password or Bitwarden to keep track of API keys. Never share your key with anyone or post it publicly—it’s linked directly to your billing account.
Step 4: Connect to Janitor AI
Back in Janitor AI:
- Open any character chat
- Click the three-bar menu button
- Select “API Settings”
- Choose “OpenAI” from the dropdown
- Select your preferred model (GPT-5.2, GPT-5.1, GPT-4o, etc.)
- Paste your API key in the designated field
- Click “Check API Key/Model” to verify it works
- Optionally add a system prompt to influence AI behavior
- Adjust temperature and other generation settings if desired
- Save your settings
That’s it. Your conversations will now use OpenAI’s models instead of JanitorLLM.
Managing OpenAI Costs:
OpenAI provides a usage dashboard where you can monitor your spending:
- Go to platform.openai.com
- Navigate to Usage section
- View daily/monthly token consumption
- Check cost breakdown by model
Set up email alerts for:
- When you’ve spent 50% of your budget
- When you’ve spent 80% of your budget
- When you’ve reached your budget limit
This helps you avoid unexpected bills.
Anthropic Claude Setup
Claude models are popular for their longer context windows and nuanced understanding. Here’s how to connect them:
Step 1: Create an Anthropic Account
Visit console.anthropic.com and sign up:
- Enter your email and create a password
- Verify your email address
- Complete the onboarding questions
Step 2: Add Payment Method
Anthropic requires payment information before API access:
- Go to Settings > Billing
- Add your credit card information
- Choose a spending limit (optional but recommended)
Anthropic offers $5 in free API credits for new accounts to test their models.
Step 3: Generate API Key
From the console:
- Navigate to API Keys section
- Click “Create Key”
- Name your key (e.g., “Janitor AI”)
- Copy the generated key immediately
- Store it securely
Step 4: Connect to Janitor AI
In Janitor AI:
- Open API Settings from any chat
- Select “Anthropic” or “OpenRouter” (if using OpenRouter for Claude)
- Paste your Anthropic API key
- Choose your preferred Claude model (4.5 Sonnet, 4.5 Haiku, Opus 4.5, etc.)
- Test the connection
- Save settings
Claude models work particularly well for longer roleplay sessions where context matters. I’ve found Claude 4.5 Sonnet to be the sweet spot for most users—it’s not the cheapest option, but it delivers consistently high-quality responses without the premium price of Opus.
Google Gemini Setup
Google’s Gemini models offer excellent value, especially for budget-conscious users who want good performance.
Step 1: Access Google AI Studio
Go to aistudio.google.com:
- Sign in with your Google account
- Accept the terms of service
- Access the API key section
Step 2: Generate API Key
- Click “Get API Key”
- Choose “Create API Key”
- Select a Google Cloud project (or create a new one)
- Copy your API key
Google offers generous free tiers for Gemini models, particularly Gemini 2.5 Flash. Check their current quota limits—they typically allow thousands of requests per day for free.
Step 3: Enable Billing (For Paid Tiers)
If you want access to Gemini 3 Pro or higher usage limits:
- Go to Google Cloud Console
- Enable billing for your project
- Set up budget alerts
Step 4: Connect to Janitor AI
Using OpenRouter is often easier for Gemini access:
- In Janitor AI API Settings, select “Proxy” or “OpenRouter”
- Add a new configuration
- Enter the Gemini model name (e.g., “google/gemini-3-pro”)
- Paste your Google API key or OpenRouter key
- Set the correct API endpoint
- Test and save
Alternatively, you can connect directly if Janitor AI supports native Gemini integration (check the latest platform updates).
Using OpenRouter for Multiple Models
OpenRouter is a platform that gives you access to multiple AI models through a single API key. This is particularly useful if you want to try different models without setting up separate accounts for each provider.
Why OpenRouter is Useful:
- Single API key for dozens of models
- Easy model switching without reconfiguring
- Often offers free or discounted access to certain models
- Handles routing and fallbacks automatically
- Provides unified billing across providers
Setting Up OpenRouter:
Step 1: Create Account
Visit openrouter.ai and sign up:
- Enter email and password
- Verify your email
- Complete profile setup
Step 2: Add Credits
OpenRouter uses a prepaid credit system:
- Go to Account > Credits
- Add funds via credit card
- Start with $5-10 to test different models
Step 3: Generate API Key
- Navigate to API Keys section
- Create a new key
- Copy and store it securely
Step 4: Browse Available Models
OpenRouter provides access to:
- OpenAI models (GPT-3.5, GPT-4, GPT-5 series)
- Anthropic models (Claude 3, Claude 4, Claude 4.5)
- Google models (Gemini 2.5, Gemini 3)
- DeepSeek models
- Open-source alternatives
- Experimental models
Browse the model list and note the exact model names you want to use.
Step 5: Configure in Janitor AI
- In Janitor AI API Settings, select “Proxy”
- Click “Add Configuration”
- Enter a friendly name (e.g., “OpenRouter – Claude 4.5 Sonnet”)
- Paste the model name from OpenRouter (e.g., “anthropic/claude-4.5-sonnet”)
- Set API URL to: https://openrouter.ai/api/v1/chat/completions
- Add your OpenRouter API key
- Test the connection
- Save
You can create multiple configurations for different models, making it easy to switch between them.
OpenRouter Advantages:
- Free tier access to some models
- Competitive pricing (sometimes cheaper than direct API access)
- No need to manage multiple API keys
- Automatic failover if a provider is down
- Usage analytics and monitoring
OpenRouter Considerations:
- Adds a small markup on some models
- Response times might be slightly slower due to routing
- Not all models are always available (depends on provider status)
From my experience, OpenRouter is perfect for users who like to experiment with different models. You can try GPT-5.2 one day, Claude Opus 4.5 the next, and Gemini 3 Pro after that—all without switching API keys or accounts.
You May Also Like
AI TechnologyAI Agent Use Cases for Business: What’s Actually Working Right Now
SaaSWhat Is No-Code Automation? The 10 Best Platforms in 2026
AI TechnologyWill AI Replace Customer Service?
AI ToolsHow to Create Amazing Photos with Gemini AI: Complete Guide for Beginners (2026)
AI Tools16 Best AI Coding Assistant Tools in 2026
AI Tools