Janitor AI is a free character-roleplay front end that hosts thousands of community bots and lets you connect your own model backend through an API. To use it, create an account, pick a character, open the API settings, choose a backend, and select a model tuned for roleplay. This guide covers setup, models, cards, and fixes.
Janitor AI confuses a lot of newcomers because it is not a single chatbot. It is a front end, meaning it provides the interface and the character library, but the actual text generation can come from different model backends that you configure. Get that one concept and everything else clicks. This guide walks through account setup, the API config screen, choosing a model, building character cards, dialing in uncensored settings, and fixing the errors people hit most.
If you want matching visuals for your characters, you can generate them free with the free generator on our homepage and keep them consistent using our character consistency techniques guide.
What Janitor AI actually is
Janitor AI is a hosted platform with a massive catalog of user-created character cards. You browse characters, click one, and start chatting. The catch is that high-quality, uncensored generation usually requires you to point Janitor AI at a model backend rather than relying solely on any built-in option. That backend can be a hosted proxy service or your own API key from a model provider.
Because of this design, Janitor AI is the most flexible app in the category and also the one with the steepest learning curve. The payoff is control: you choose the model, the latitude, and often the price.
Step 1: Create an account
Go to the official site and sign up with an email. Confirm you meet the minimum age requirement, since the platform hosts mature content. Once you are in, you can browse characters immediately, but you will want to set up a backend before serious roleplay.
Step 2: Understand the API config screen
The settings panel is where people get stuck. Here is what the key fields mean.
API or backend type. This tells Janitor AI where to send your messages for generation. Options typically include a hosted proxy, an OpenAI-compatible endpoint, or another provider’s API.
API key. If you bring your own model, you paste your provider key here. This is what gets billed, so guard it like a password.
Model name. The specific model the backend should use. Roleplay-tuned models behave very differently from general assistants.
Proxy URL or endpoint. For OpenAI-compatible setups, this is the base URL the requests go to. Get this exactly right or you will see connection errors.
Context size and temperature. Context size controls how much conversation the model can see, which affects memory. Temperature controls randomness, where higher is more creative and lower is more consistent.
After filling these, send a test message. A reply means the backend is wired correctly.
Step 3: Free versus paid backends
| Backend type | Cost | Setup effort | NSFW latitude | Best for |
|---|---|---|---|---|
| Built-in free option | Free | Lowest | Variable | Casual trying |
| Hosted proxy | Free to low | Low | High when available | Beginners wanting open chat |
| Your own API key | Pay per use | Medium | Depends on model | Power users wanting control |
| Self-hosted / local | Hardware cost | High | Very high | Privacy-focused tinkerers |
Hosted proxies are popular because they are easy and often allow mature content, but availability fluctuates. Bringing your own key gives stable access and control but costs per message. Self-hosting via a local model offers maximum privacy and latitude if you have the hardware.
Step 4: Choosing a model for roleplay
Not all models roleplay well. General assistant models tend to be cautious and break character. Roleplay-tuned or instruction-following models with higher creativity hold persona better.
When choosing, prioritize a larger context window for memory, a model known for following character instructions, and one whose policy or tuning permits mature themes if that is your goal. In our testing, roleplay-focused models held character noticeably longer than generic assistants before drifting. Read the backend’s documentation for the exact model names it supports.
Step 5: Importing and building character cards
Character cards are the heart of Janitor AI.
Using existing cards. Browse the library, open a character, and start. You can favorite cards and tweak some settings per chat.
Building your own. Use the create flow and fill these fields:
- Name and avatar.
- Personality, written concisely with the most important traits first.
- Scenario, which sets the setting and situation.
- Greeting, the character’s opening line that anchors tone.
- Example dialogue, a few sample exchanges that teach the model the voice.
Front-load the critical details, since models weight early text more heavily. Keep the personality tight rather than dumping a novel, which can dilute the most important traits.
Step 6: Settings for uncensored chat
To keep roleplay open and in-character, combine a permissive backend with good card writing. A clear scenario and greeting that state the tone reduce the model steering away. A slightly higher temperature adds creativity, while a generous context size preserves memory. If a model resists, the issue is usually the backend’s policy rather than your settings, so switching backends often solves it.
For visuals to match your character, generate a reference image with the free generator on our homepage and reuse it. If you want a fully custom look, our LoRA training guide shows how to lock a character’s appearance.
Common errors and fixes
No reply or blank message. Usually a backend connection problem. Recheck the endpoint URL, API key, and model name. Send a test message after each change.
Invalid API key. The key is wrong, expired, or out of credit. Regenerate it in your provider dashboard and confirm billing is active.
Network or CORS errors. Often a proxy issue. Confirm the proxy is online and the URL is exact, including the path.
Character keeps breaking persona. The model is too cautious or the card is weak. Switch to a roleplay-tuned model and strengthen the personality and example dialogue.
Memory loss mid-chat. Context size is too small. Increase it if the backend allows, and write a short pinned summary of key facts.
Replies are too short or get cut off. The max output token setting is too low. Raise it modestly so replies finish their thought without rambling for paragraphs.
Rate limited or throttled. Free proxies and budget tiers often cap how fast you can send messages. Slow down, or move to a paid key for stable throughput. Persistent throttling usually means the backend is overloaded rather than a problem on your end.
Replies feel generic and bland. Either the model is a cautious general assistant or the character card is thin. Switch to a roleplay-tuned model and enrich the personality and example dialogue. The card does most of the heavy lifting, so invest there before blaming the model.
Choosing the right backend for your situation
The backend decision is the one that shapes everything else, so match it to your priorities. If you want zero cost and minimal setup, start with a free hosted proxy and accept that availability and policies are outside your control. If you want stable, predictable access and are willing to pay per message, bring your own API key from a provider whose terms permit your use case. If privacy is paramount and you have capable hardware, run a local model so nothing leaves your machine.
A practical path for newcomers is to begin on a free proxy to learn the interface and confirm Janitor AI suits you, then graduate to your own key once you know which model behaves best for your roleplay. That way you avoid paying before you understand what you are buying, and you do not waste time self-hosting until you are sure you want the control it offers.
Realistic expectations on quality
Set your expectations by the model, not the platform. Janitor AI is a window onto whatever model you connect, so a strong roleplay model produces strong roleplay and a weak one produces weak roleplay regardless of how nice the front end looks. The most common disappointment comes from pairing a great character card with a cautious general-purpose model and concluding the app is bad. It is not; the model simply is not built for open roleplay. Swap it and the same card comes alive.
Understanding tokens, context, and cost
If you bring your own API key, the money question comes down to tokens. A token is roughly three quarters of a word, and you pay for both the input (your message plus the character card plus the conversation history the model sees) and the output (the reply). This matters because a long conversation with a big character card sends a lot of input tokens on every single turn, since the model re-reads the history to stay coherent.
Two practical consequences follow. First, a bloated character card costs you on every message, so keep personas tight. Second, a larger context window improves memory but increases per-message cost, so there is a tradeoff between recall and budget. Many power users set a moderate context size and lean on a pinned summary of key facts to get most of the memory benefit at a fraction of the token spend.
If you use a free hosted proxy instead of your own key, you skip the metering but accept variable availability and policies you do not control. Neither approach is wrong; they suit different priorities.
Optimizing replies: temperature, tokens, and prompts
Beyond picking a model, a few settings shape the feel of every reply.
Temperature. Higher values, around 0.9 to 1.1, make replies more creative and varied, which suits open roleplay. Lower values, around 0.6 to 0.8, make the character more consistent and predictable. If your character feels random and forgets its own traits, lower the temperature. If it feels stiff and repetitive, raise it.
Max output tokens. This caps reply length. Set it too low and replies get cut off mid-sentence; too high and the model rambles. A moderate cap keeps pacing natural.
The system prompt. Many backends let you add a system instruction that applies to every message. A short instruction reinforcing the roleplay format, the perspective, and the tone can dramatically reduce out-of-character breaks. Keep it concise and specific.
Restating context. When a long chat starts drifting, drop a one-line reminder of the key facts. Because the model weights recent text heavily, this nudges it back on track without restarting.
Janitor AI versus dedicated apps
The contrast with plug-and-play apps like SpicyChat and CrushOn AI is stark. Setup effort on Janitor AI is high, since you configure a backend, while dedicated apps let you chat in seconds. Model choice is the big differentiator: Janitor AI gives you full control, whereas dedicated apps fix the model for you. On cost, Janitor AI is pay-per-token or a free proxy, while dedicated apps run a flat subscription or free tier. NSFW latitude on Janitor AI depends entirely on the backend you choose, whereas dedicated apps set it themselves, often generously. In short, Janitor AI is for tinkerers and control seekers, and dedicated apps are for users who just want to chat.
The takeaway: Janitor AI trades convenience for control. If configuring an endpoint and weighing token costs sounds tedious, a dedicated app will make you happier. If you want to pick your exact model and tune everything, nothing else comes close.
Privacy and self-hosting
The most private way to use Janitor AI is to point it at a local model running on your own hardware. Your conversations never leave your machine, latitude is entirely your choice, and there is no per-message cost beyond electricity. The downside is that capable local models need a decent GPU and some setup patience. For most people a hosted backend is the practical choice, but privacy-focused users should know the local path exists.
Who Janitor AI is for
Janitor AI rewards people who like to tinker and want control over their model, latitude, and cost. If you want plug-and-play simplicity instead, our roundup of the best NSFW AI roleplay generators lists easier options like SpicyChat and CrushOn AI.
Pairing Janitor AI with visuals
Because Janitor AI is text-first, your visuals come from elsewhere, and that is actually an advantage: you get full control instead of mediocre in-chat images. Generate a reference portrait of your character once, then reuse it so the look stays consistent across scenes. The NSFW AI character image generator is built for exactly this, and our best NSFW image generators roundup compares the strongest paid and free tools.
If budget is a concern, the best free uncensored generators guide lists no-cost options, and our prompt examples library gives you templates so your first prompt lands. The easiest place to start is the free generator on our homepage. If you prefer a guided companion app over a configurable front end, our Candy AI review covers the polished alternative.
When Janitor AI is the wrong choice
Be honest with yourself about how much configuration you enjoy. Janitor AI shines for people who like turning dials, comparing models, and weighing token costs. If reading about endpoints and API keys already feels tedious, you will be happier on a flat-rate app where everything just works. The flexibility that makes Janitor AI powerful is the same thing that makes it frustrating for users who only want to chat. There is no shame in picking the simpler tool; the best app is the one you actually use.
Verdict
Janitor AI is the most flexible roleplay platform in 2026, but only if you embrace its front-end nature. Set up a backend, pick a roleplay-tuned model, write strong character cards, and you get a deeply customizable experience that few apps match. Pair it with a dedicated image generator and consistency techniques for the full package.
Frequently asked questions
Is Janitor AI free to use?
Janitor AI itself is free to access, and you can browse and chat with characters at no cost. However, the best uncensored roleplay usually requires connecting a model backend, which may be a free hosted proxy or your own paid API key. So while the front end is free, high-quality generation can carry a cost.
What does the API setting in Janitor AI do?
The API setting tells Janitor AI where to send your messages for text generation. You choose a backend type, paste an API key if bringing your own model, set the model name and endpoint URL, and tune context size and temperature. Getting these fields right is what makes the chat work and stay uncensored.
What is the best model for Janitor AI roleplay?
Roleplay-tuned or strong instruction-following models with a large context window work best. General assistant models tend to be cautious and break character. Prioritize a model known for following character instructions and one whose tuning permits mature themes if needed. Check your chosen backend’s documentation for the exact supported model names.
Why is Janitor AI not responding?
A blank or missing reply almost always means a backend connection problem. Recheck the endpoint URL, API key, and model name, then send a test message after each change. Invalid keys, offline proxies, or exhausted credit are the usual culprits. Confirm your provider billing is active if you use your own key.
How do I make Janitor AI uncensored?
Combine a permissive backend with strong card writing. Choose a roleplay-tuned model whose policy allows mature themes, write a clear scenario and greeting that state the tone, raise temperature slightly for creativity, and use a generous context size. If a model resists, switching backends usually fixes it, since the limit is often the backend policy.
How do I create a character card in Janitor AI?
Use the create flow and fill in name, avatar, personality, scenario, greeting, and example dialogue. Write the personality concisely with the most important traits first, since models weight early text more heavily. Add a few example exchanges to teach the voice, and a strong opening greeting to anchor the tone and reduce drift.
Is Janitor AI safe and private?
Janitor AI publishes terms and a privacy policy, but adult chat is sensitive, so review them before sharing personal details. Guard any API keys you enter, use a dedicated email, and avoid real identifying information in roleplay. For maximum privacy, advanced users self-host a local model backend rather than routing through a third-party proxy.
Can Janitor AI generate images?
Janitor AI is primarily a text roleplay platform, and its image capabilities are limited compared to dedicated apps. Most users generate visuals separately and reference them in chat. For consistent character art, generate a reference image with a dedicated tool and reuse it, or train a custom model to lock the character’s appearance.
More AI companion guides:
