Janitor AI Review 2026: A Free Uncensored Library, Bring-Your-Own Model and Proxy Risks

Janitor AI is a platform for text roleplay with AI characters created by its community. You browse a huge library of detailed personas — each with a backstory, personality and scenario — and chat with them.

Janitor handles the interface, character management and conversation context, while the language model behind the chat is either its own hosted model or one you connect. The homepage reports more than 15 million users, and the library is one of the largest in AI roleplay.

It is text-only: there is no voice, image or video generation here.

The company is JanitorAI, Inc., and the platform was created by founder Jan Zoltkowski, an Australian developer who built it at a hackathon and launched it on 1 June 2023; according to Semafor, it reached a million users in 17 days.

Its early history is colourful and well documented: Semafor reported that Zoltkowski initially pitched investors "with the help of" the convicted "pharma bro" Martin Shkreli, who Zoltkowski said was no longer involved.

It grew by word of mouth as a less-filtered alternative to an increasingly restricted Character AI.

Janitor's defining choice is that you pick the brain behind your roleplay. There are three routes:

• JLLM (JanitorLLM) — Janitor's own hosted model, free and uncensored-capable, with a context window of around 9,000 tokens. Good enough for a lot of roleplay with no setup, and there is no hard daily cap.
• Bring your own API key — connect OpenAI, Anthropic's Claude, DeepSeek or OpenRouter. Janitor charges nothing for this; you pay the provider for the tokens you use, and stronger models give noticeably better writing and longer memory.
• janitor+ Router — a built-in, pay-per-token way to reach stronger models with no proxy or API key to manage. It is part of the janitor+ subscription (more on cost below) and is web-first.

Other strengths reward the time you invest.

Lorebooks let you store persistent facts — a world's rules, a character's history — that the model always has on hand, which is the closest Janitor gets to reliable memory; and personas let you keep several versions of yourself and reuse them across characters.

There is a character builder with custom CSS, group chats, and portability — character cards move to and from SillyTavern. In February 2026 Janitor added official iOS and Android apps, so it is no longer web-only.

Watch for look-alike "Janitor" apps in the stores: only the app published by JanitorAI, Inc. is the real one. For a polished companion with voice and images instead, pair Janitor with a tool like Candy AI.

Janitor AI is free to use, and its founder has said a free option will always remain. What you pay depends on how you power your chats.

In June 2026 Janitor launched janitor+, its first subscription, at $12.99/month (or $99/year on some web accounts) — so older guides citing a "$9.99 Basic / $19.99 Premium" layout are out of date; that structure was never Janitor's current model.

The janitor+ Router sits inside the subscription: it runs premium third-party models per token with no proxy or key to set up, which is the no-fuss paid path for better writing.

If you bring your own API key instead, you pay the external provider directly — a few dollars a month for light use, more if you lean on premium models heavily — so the real cost depends on how much you chat and which model you choose.

For light users, Janitor can genuinely cost nothing; for heavy users on strong external models it can cost more than a flat-rate app like Candy AI ($13.99), but you gain far more variety and model choice.

Subscriptions are generally non-refundable, and an app-store subscription must be cancelled through the store, not the website. Prices checked October 2026; confirm current rates in the app.

The single most important safety point about Janitor has nothing to do with its own servers. To avoid paying for premium models, many users route chats through community-run reverse proxies that share someone's API access.

These are a genuine hazard: whoever runs a proxy can log every message you send and can harvest any API key you paste into it, and some proxies are honeypots set up to do exactly that.

In 2023, community warnings circulated that the well-known "moxxie" proxy had leaked users' chat logs and IP addresses — a user-reported incident rather than a hack of Janitor, but a clear illustration of the risk.

There is a reason this matters for Janitor specifically: the whole appeal of cheap, uncensored access pushes newcomers toward proxies they do not understand.

The safe routes are the free JLLM, the janitor+ Router, or your own API key entered only on official provider and Janitor pages. Treat any "free premium" proxy as untrusted.

This is also the history that shaped the platform: Janitor launched on OpenAI's models and, as Semafor reported, drew a cease-and-desist from OpenAI in July 2023 over its NSFW use, which spurred the pivot to its own JLLM.

Janitor is 18+ and hides NSFW by default; you confirm your age and toggle it on. Crucially, the toggle controls Janitor's own library visibility and JLLM behaviour — it does not change a third-party model's training.

Run an uncensored model (or JLLM in its uncensored mode) and adult roleplay is open; route through OpenAI or Anthropic and their moderation applies and may refuse. So "how filtered is Janitor?" honestly depends on the model you connect, not a single global switch.

The hard limits are strict and set out in the terms regardless of model: no content involving minors (including "aged-up" or minor-coded characters), no real-animal or incest content, and no sexual depictions of real, identifiable people. Accounts reasonably suspected of belonging to under-18s are banned.

Those are the right lines.

It is worth knowing, though, that the platform's openness can be abused by others: the AI Incident Database records a January 2025 Massachusetts case in which a man who pleaded guilty to cyberstalking was alleged to have used tools "like CrushOn.ai and JanitorAI" to create deepfakes — a misuse of the tool, not a failure of Janitor's own security.

Janitor's own privacy posture is weakly documented, and an independent tracker says so. The VerifyWise AI Trust Index grades Janitor AI D, 44/100 (assessed 13 August 2026).

Its privacy policy may use your data — including chats — to improve its AI systems, and VerifyWise lists a training opt-out as "Not stated": there is no disclosed way to keep your inputs out of model training.

The policy also grants the company a broad, "worldwide, non-exclusive, royalty-free, sublicensable, and transferable" licence to your content, and that licence is not clearly revoked when you delete your account.

Retention is described only as "as long as necessary," with no named timeframe, and no end-to-end encryption or two-factor authentication is documented. If you bring your own API key, your conversations also reach that provider under its policies.

On the brighter side, accounts are pseudonymous, no ID is required, the policy states it does not sell your personal data, and no breach of Janitor's own servers is on record (do not confuse it with unrelated AI-chat leaks reported elsewhere).

The practical advice is simple: use a burner email, never paste an API key into an unofficial proxy, and assume your chats are stored and may help train the model.

That combination of a broad, lingering content licence and no stated training opt-out is common in this corner of the market, but it is worth knowing before you pour months of character and lorebook work into the platform.

Use Janitor if you want the widest character variety for little or no money, you enjoy picking and tuning your own model, and uncensored text roleplay is the goal. For experienced roleplayers, nothing else quite matches the library plus model control at this price.

Skip it if you want plug-and-play polish, voice or image companions (Candy AI or Kindroid), strong privacy guarantees, or you are not comfortable with API keys and the proxy risks above.

The easiest free alternative with a similar feel and no model setup is SpicyChat.

In short: Janitor AI wins on library size and model freedom, and is unbeatable value for tinkerers — provided you respect the proxy risks and its weak documented privacy.

This is a research-based review. No paid subscription was purchased and the platform was not evaluated first-hand.

In October 2026 the sources checked were: Janitor AI's own site and official help-center articles (read through the r.jina.ai reader proxy, which returns the text of pages that block direct fetching — the text is the official page) for the free tier, JLLM context, janitor+ price and cancellation terms; the VerifyWise AI Trust Index for the privacy grade, content licence and training stance; Semafor's September 2023 reporting for the founding, the OpenAI cease-and-desist and the Shkreli detail; the AI Incident Database for documented misuse; and official app-store listings for the native apps.

That shapes the scores. Operator, pricing, the free tier, age and content rules and the data stance come from Janitor's own documents and reputable reporting, and carry higher confidence.

Judgements about writing quality, memory and model behaviour rest on documented capability rather than first-hand observation, so the Features and Content sub-scores are held deliberately modest, and the Safety score leans on the dated VerifyWise assessment and primary incident records.

Volatile facts are dated so you can re-check them in the app.