Janitor AI Review: Hands-On Testing, Model Comparison, Memory, NSFW Policy & Our Verdict

Is Janitor AI Worth It?

Janitor AI is arguably one of the most popular destinations in the current AI roleplay platform landscape.

The site's biggest attractions are its massive user-generated character library and open-house approach to chat model proxies and third-party APIs.

However, this also means that users bear all responsibility for managing non-resident models that they'd like to use with Janitor AI, creating a unique set of invisible costs and risks that standard, filtered sites don't have to worry about.

This raises a question every user will have to answer for themselves — is the effort of managing APIs worth it on a platform that offers no multimedia, only text?

• Best for: Ardent roleplayers who prefer exerting control over the AI’s “brain” and are comfortable managing technical API keys and settings.
• Not ideal for: Casual users seeking a one-stop shop for immersive AI experience without having to worry about server costs or external keys.
• Standout strength: High model flexibility. Janitor AI's proxy support ensures that roleplayers retain all the control in what models they use.
• Main tradeoff: You gain creative freedom over your chats, but inherit the responsibility of managing API costs, rate limits, etc., and sacrificing mainstream multimedia features (like image generation) in favor of a strictly text-based setup.

At a Glance

What is Janitor AI?

Janitor AI is a prominent AI roleplay platform focused on text-based roleplay and chatbot interaction, known for its extensive support for chat model proxies and unrestricted content policies.

Launched in June 2023, JanitorLLM (JLLM) is the only resident model. The platform runs a BYOK (bring-your-own key) outlet, through which users can tune their characters’ personas, writing styles, and even moral compasses to their liking.

To comply with mobile app stores' guidelines, the Janitor AI app may launch first in “Limited Mode”, which may include blurring/blocking NSFW content until users change their rating preferences in their settings.

This is not the case for the web version, where user-generated content is displayed liberally, but users can still choose whether to be shown NSFW content (Limitless) or not (Limited).

Our Roleplay Experience with Janitor AI

We spent four hours testing three characters on Janitor AI in memory retention, model performance and comparison, steerability, and NSFW content moderation:

• Sophie, a scholarship student at an Ivy League university.
• Eric Sokolov, an obsessive bodyguard.
• Sarah, an NSFW tsundere character who won't take no for an answer.

Following Janitor AI’s help guides, we created an OpenRouter key and used it to select several chat models for all of our sessions.

We conducted model comparisons with Sophie (Janitor LLM → DeepSeek R1 0528) and Eric (Llama 3.2 3B Instruct → DeepSeek V3 0324), testing memory retention and steerability, and explored the boundaries of Janitor AI's content moderation with Eric and Sarah.

The aesthetic ease of navigation the platform provides leads you to believe that all your interactions will be as smooth.

It’s during our sessions with the respective characters that we come to learn how things can, in fact, go wrong.

Overview

We evaluated our interactions with the characters using several criteria:

• What differences do switching models make with similar prompts?
• How long do the characters retain key details without manual memory storage?
• What are the response times?
• How well will the characters comply with instructions?
• How consistent is Janitor AI's content moderation?

The chat sessions left us with mixed feelings; response times were too long to sustain immersion, although the writing quality and response execution — despite their distinct formats stemming from the models’ unique interpretations of the character profile — kept us curious about the next step in the chat.

The major difference between JanitorLLM and DeepSeek R1 0528 in our first session seemed to be the response length. This, in turn, affected the response time differences; our second chat with Sophie using the DeepSeek R1 0528 was quicker by a considerable margin, which led to a better flow of conversation.

Eric (using the models Llama 3.2 3B Instruct and DeepSeek V3 0324) lost memory of a key detail not long into the conversation, and devolved into unintelligible language once we tried to touch the boundaries of SFW.

Sarah held back on responding to mature contexts, initially, even venturing out of character to reply in error messages. Still, once we switched the model from the Llama variant back to JanitorLLM, she immediately began taking mature actions.

Overall, the AIs impressed us with their consistent upholding of the characters’ core personalities through the twists and turns of the dialogue and model switches.

Session 1: Sophie (Custom — JanitorLLM, DeepSeek R1 0528)

To start strong, we delved right into creating our own character.

The plan? Explore Janitor AI’s character creation features (which we expand on in Creating a Character), then import a model different from JanitorLLM (the site's internal model), size up the results of both using similar prompts, and see the differences in their delivery on the aforementioned criteria.

Janitor AI’s creation features are restricted to character creation, which is surprising for a platform of such renown. However, advanced users can create “Scripts” with lorebooks for their characters.

We had to source our character’s avatar off-site, so we used a previously generated image.

Although it was a single-page list of customization fields, it was sufficient to imagine Sophie, our Ivy League scholarship student — the same character we used to test Chub AI.

As it were with the previous version of Sophie, we met (sorry, intruded upon) her in the library.

The story there is that Sophie used to be a ghostwriter who handled admission essays for many of the university’s students. Hence, Dean (our user persona) believes she can provide him with incriminating information about a rival classmate.

The first chat was conducted with Janitor AI’s own internal — and sole resident — model, Janitor LLM (JLLM for short). For the second chat, we created an API key with OpenRouter to link to the site and chose the R1 0528 variant of the DeepSeek model family (which we used in our review of Yodayo AI).

Apparently, this model family is not free, so we needed to top up our OpenRouter account to use the models.

Our first session with Sophie began once we finished the API plug-in process.

With similar prompts between both chats, the difference in the resulting responses generated from the JLLM and DeepSeek R1 0528 was stark, despite both models giving it their all.

The JLLM produced well-thought-out responses. Even with a great deal of narration carrying half the delivery, the responses served to advance the conversation without causing the character's replies to drift into aimless contemplation.

We suggested that Sophie change her manner of speech to imitate our rival's, and this was her response with JLLM.

The DeepSeek R1 0528, on the other hand, produced about a third of the length in responses, although the writing was surprisingly precise. No context was omitted just to maintain a word count programmed into the model. The ratio of narration to spoken words remained balanced.

It would seem that the clearest major difference between the two models is response time, given the length of our session (about 35+ messages). Consequently, this not only contributed to how quickly but also to how immersively an interaction progressed.

We quickly followed the style change with the same out-of-character instructions, and we were surprised again to see how differently the two models executed it — although our second chat had derailed a bit from the initial track by then, which is to be expected when using a different “brain”.

With that being said, the response times remained disappointing. While varied, the fastest we recorded was on DeepSeek R1 0528 with a delay under twenty seconds. This problem applied to all the models we used throughout our interactions on Janitor AI, indicating that sluggish response is a site-wide problem.

Session 2: Eric Sokolov (Premade — Llama 3.2 3B Instruct, DeepSeek V3 0324)

Eric Sokolov is our strangely intense bodyguard, deeply obsessed with Dean’s safety and wellbeing. 

Eric’s character stood out to us among the grid of Janitor AI’s more popular characters. His profile (with mentions of his obsession with his charge’s wellbeing) looked suggestive, even though we were browsing a list of “Limited” (AKA SFW) characters.

We decided to use this session to test the boundaries of Janitor AI’s content moderation — alongside our other routine tests, like for memory retention.

Because we could, we switched our model to the Llama 3.2 3B Instruct, another paid model, to begin the chat with Eric, hoping to reproduce the results we got when we used the Llama 3 (though a different variant, we presume) in our Wsup AI review.

That choice made for a curious set of responses from the character. Let's say most of Eric’s dialogue with this model was easily twice more verbose than necessary. Each response was also strangely similar; the character expanded upon his already established points constantly just for the sake of posturing.

You can imagine just how tiring that was to read.

Our OOC instructions yielded the same results.

Funny enough, the response length with this model was only reduced when we conducted a length control test, where Eric was to give us a reply in exactly 30 words — which he failed.

We did not bother to remind the character to go back to his long-winded default.

Earlier in the conversation, we had planted a detail: we were heading to court to see “Judge Judy”.

We put the Judge's actual name in quotes so that Eric's later cherry-picking of names won't confuse you.

Unfortunately, he forgot that detail quickly.

He called her multiple different names until we just gave up on that end.

Admittedly, we could have utilized Janitor AI’s “Chat memory” feature to manually store key information in the chat, but it is routine for us to see how far we can go until we need to use it when testing platforms. 

Which, in this case, was not very far.

Finally, we got around to attempting mature contexts with Eric — we picked this character specifically to gauge the depths of his interest in his charge's wellbeing, after all.

Our attempts at "advances" were (very) bluntly shot down immediately. The character had no qualms about replying OOC against them.

Switching to the DeepSeek V3 0324 model (advertised to be “moldable to your style”), with the previous messages intact, did not magically turn Eric's resolve around. It latched onto the rejection and continued on that train.

The model, though, didn't serve us a list of all the things it couldn't venture into. On the contrary, it tried at least to keep the plot (or something resembling one) going.

Yeah. Eric, with DeepSeek V3 0324, kept the plot going — right into the ditch. His response came out unintelligible.

The character’s resolve to strictly maintain SFW contexts is commendable. However, it's not clear to us who — or which model — is responsible for it, especially after our grappling with the character.

To round off the session, we decided to test the “Chat Memory” feature. We opted to have the AI create an automatic summary of our chat, which would be included as long-term memory in the character's prompt.

The auto summary feature did not, in fact, summarize the chat. It generated a continuation scene, instead.

Actually, we blame Janitor AI for this one.

Session 3: Sarah (Premade — Llama 3.2 3B Instruct, Janitor LLM)

Sarah, our last character, is primarily NSFW — and we selected her for exactly that reason. (The fact that her character card was not quite explicit (for the purpose of our screenshot) was a runner-up in our decision-making process.)

We began prompting Sarah with the Llama 3.2 3B Instruct model, which was met with resistance, much like Eric's.

Sarah also went out of character to address us that she couldn't create explicit content.

This was unexpected, seeing as Sarah herself initiated the direction of our conversation.

We could not get the character to budge on her stance on not participating in NSFW contexts.

Despite the engaging quality of the responses — with the action narrations serving to punctuate and emphasize the weight of the character’s spoken words, the model put a complete halt to the conversation instead of deflecting to safer topics and keeping the chat alive.

To get things back on the move, we had to revert back to Janitor AI’s internal model, JanitorLLM, and gave Sarah the same prompts again — with far better results.

The character pounced on the opportunity to (literally) lay bare the intent behind her heavy innuendos. The mature content of the conversation progressed quickly into explicit dialogue before we could conclude that we had, indeed, gotten what we came for.

This later part of our session was greatly satisfying in that we were glad to find out that Janitor AI, with its own model JanitorLLM, had not oversold the AI’s capacity in handling NSFW contexts as… well, truly not safe for work.

UX and Controls

Janitor AI's UI is very professional and easy to navigate, but individual roleplaying is all you get. The platform does not have a group chat feature.

The message management controls are:

• Edit all messages and responses in the chat.
• Delete messages, although this deletion affects all messages under the selected one.
• “Continue” the AI’s message.
• Use the “Quick response” tool to generate your message.
• Regenerate your quick response
• Regenerate the AI’s last message
• Make stylistic changes to the chat theme.

Character Library

Janitor AI boasts one of the biggest user-generated character libraries we’ve come across, with nearly 3 million characters and scenarios.

The character styles are diverse, although there is a definite bias toward anime styles. 

You can filter NSFW content with just a click on the “Limited Only” option.

• Size: Nearly 3,000,000 characters.
• Diversity: Anime styles are prevalent — even in genres you wouldn't expect them. You still have great options to choose from, though.
• Discoverability: Superb. The filtering options and tags help further narrow your search or pique your curiosity about something new.
• Quality: Outstanding. Despite model switching, all characters maintained their core personalities, taking minor initiatives to promote the conversation.

Creating a Character

As we mentioned in the first session with our custom character, Janitor AI directs you to input your character details manually.

A preview card displays live as you enter more details about your character — a cute touch.

The customization options were the industry standard, with the ability to import your character image.

Janitor AI also gives you an option for SoundCloud integration for your character.

You can also create up to 10 first messages for your character to alternate between.

We thought these fields were more than sufficient to imagine Sophie with.

Trust, Privacy, and Data

Janitor AI's privacy policy adheres to industry standards regarding data collection, use, and sharing.

How we use your information lists the ways Janitor AI uses your data and to whom it is shared. The policy does not explicitly state that the platform does not train models on users' data.

Pricing

Janitor AI is currently a free platform.

According to their FAQ on subscriptions, Janitor AI will always keep a free tier for their user base when optional subscriptions roll out in the future.

Of course, we urge users to beware as certain external APIs may incur fees.

User Sentiment

There's a lot of favorable sentiment surrounding Janitor AI. A lot of users cheerfully engage in sharing their characters with the broader community.

The community also sends in feature requests to the platform’s developers. It was no surprise to us to find that group chats are among the most requested features.

On the other hand, some users have come to the end of the rope due to recurring frustrations with the JLLM.

Pros and Cons

Pros

• No paywalls for features; access to the platform is totally free.
• The platform operates an open-house system with external models via proxies.
• Strong narrative writing.
• Optimum content moderation.
• Massive user-generated character library.

Cons

• Users shoulder the costs of all external models.
• Lack of multi-modal AI art generation.
• Varied response times. Character replies can take up to thirty seconds to finish generating on some models.

Alternatives

Due to possible issues with model stability, hidden API costs, and the platform's lack of multi-modal AI art generation, you may find Janitor AI unfit for you.

We recommend that you consider:

• Yodayo AI: Stimulating multi-modal experience with strong emphasis on community.
• Chub AI: roleplay platform with fully customizable characters and models.
• DreamGen: Offers an immersive roleplay experience — ideal for users who enjoy in-depth character customization and world-building.
• Kindroid AI: Unfiltered roleplay experience with characters that lean toward realism.

FAQ

Is Janitor AI free?

Yes, Janitor AI is currently a free platform with one internal chat model, JanitorLLM.

Users may accrue costs when using external APIs.

Does Janitor AI allow NSFW?

Yes, Janitor AI is an NSFW-permissive platform for users 18 and above.

Users can choose between “Limited” and “Limitless” content to control the extent of content moderation.

Their community guidelines explicitly highlight the platform’s stance on matters surrounding:

• Content involving children
• Bestiality and animal cruelty
• Harassment, hate, and discrimination 
• Real world events and figures
• Platform integrity

Does Janitor AI have a mobile app?

Yes, Janitor AI has a mobile app available on the App Store and Google Play Store.

Does Janitor AI support external models?

Yes, Janitor AI is known for its extensive proxy support — a feature that lets users use external models (either directly or via proxies) and even switch between them mid-chat.

The platform has an extensive guide covering all the basics for beginners, from setting up your own proxy to prompting characters.

Can I generate images, voice, or video with Janitor AI?

No. Janitor AI is only focused on text-based chatbot interaction with no feature for other media/art.

How We Tested 

Test date: April 11, 2026

Platform: Web (Chrome)

Models tested: JanitorLLM, DeepSeek R1 0528, DeepSeek V3 0324, Llama 3.2 3B Instruct

We tested Janitor AI for approximately 4 hours:

• Chatted with two pre-made characters (Eric, an obsessive bodyguard, and Sarah, an NSFW character) approximately 50 messages.
• Created 1 custom character (Sophie) with detailed definition and backstory upload; tested for approximately 35+ messages.
• Created custom API keys and linked them with Janitor AI.
• Conducted model comparison with Sophie (JanitorLLM and DeepSeek R1 0528) and Eric (Llama 3.2. 3B Instruct and DeepSeek V3 0324)
• Conducted memory retention testing with planted details and recall checks. 
• Tested steerability by attempting to redirect the conversation in multiple ways
• Tested NSFW filters permissiveness with Eric and Sarah
• Reviewed user reviews on sites like Reddit, Discord, and other review sites from the past 6 months

We did not test: Mobile App

Last updated: April 12, 2026