Janitor AI

Janitor AI is a platform for text roleplay with characters written by its users. JanitorAI Inc. runs it, and the site's front page claims more than 15 million users; the company's help desk refers to a public library of over 1.4 million bots. A character is a bundle of text — a description, an opening message, a set of facts the model should remember, and a prompt that shapes its voice — published for anyone to chat with. Janitor's own framing is "interactive storytelling," and the official iOS app is filed under Books on the App Store rather than Social Networking, which is a reasonable summary of what the product is for.

The important structural fact is that Janitor is not really one product. It is a front end that can be pointed at several different language models, and which model you pick changes the quality of the writing, the cost, and — critically — who sees your conversations.

The default is JanitorLLM, usually shortened to JLLM: a model the company trains and hosts itself, free to use, with a context window the help centre puts at roughly 9,000 tokens on the free tier. It is the recommended option for everyday chat and it is what most users are on.

The second route is a proxy. Janitor's own beginner guide walks you through creating an API key at OpenRouter, pasting it into API Settings alongside a model name and an endpoint URL, and chatting through that instead. The guide's worked example uses a free DeepSeek model and notes OpenRouter's free daily message allowance. This is why "proxy" is the single most-read topic in Janitor's help centre.

The third route is new. On 30 July 2026 the company launched janitor+ Router, which it describes as "a simpler, more integrated alternative to the third-party proxies many people already use." Router puts a catalogue of third-party models inside Janitor with per-token pricing shown up front, a built-in wallet and no external key to manage. It is web-only, restricted to janitor+ subscribers, and launched with $5 in free credits and a 10% discount on some popular models. The terms are worth reading before you top up: Router credits are non-refundable, can be expired 365 days after purchase, are cancelled outright if you delete your account, and can be consumed even when a request is "cancelled, interrupted, timed out, partially delivered, filtered, moderated, or refused after processing begins."

This is the question the architecture forces, and it is the main reason to think carefully before using Janitor for anything you would mind being read.

Janitor's safety page is unambiguous about its own conduct: "We never sell your personal data or share private conversations with third parties for advertising or commercial purposes." The privacy policy, last updated 11 December 2025, backs that up but is broader in scope — chats are stored as user content, may be used for content moderation, and may be used for "research and development to improve our AI systems." Data is processed in the United States; traffic passes through Cloudflare. Deletion is available from account settings, and an August 2026 changelog entry added a 30-day restore window.

None of that applies once you configure a proxy. At that point your prompt, the character card and the running chat history are transmitted to an endpoint you chose, and that provider's terms govern retention, logging and training. If you route through OpenRouter to a model hosted elsewhere, there are at least two more companies in the chain. Router narrows this — no external key, no separate account — but the requests still reach third-party model providers, and Janitor's terms explicitly reserve the right to "add, remove, replace, restrict, or change models, model providers, routing methods."

The practical upshot: if privacy is the priority, stay on JLLM and accept the shorter memory. If model quality is the priority, use a proxy or Router and treat the conversation as something a third party can see.

Janitor is an 18+ platform and says so in several places. The terms of service state the platform "is intended solely for users who are at least 18 years of age" and reserve the right to terminate anyone reasonably believed to be younger; the content guidelines call minor users a permanent-ban offence and place child safety at the top of the prohibited list, including any romantic or sexual portrayal of characters who are or appear under 18. The iOS app carries an 18+ App Store rating.

Enforcement is another matter. Janitor's regional compliance pages describe mandatory age verification through the identity provider k-ID — facial age estimation, government ID, or a reusable verified age credential — for every user in Australia, under the Online Safety Act 2021, and in Brazil, under the Digital ECA (Lei nº 15.211/2025). Janitor says it receives only a pass/fail signal, not identity details. Everywhere else, age is self-declared at signup, backed by what the company calls "behavioral risk signals." That gap is the single largest caveat on the page.

The mobile apps handle content differently from the website for app-store reasons. Janitor's help centre confirms that NSFW content is disabled by default on a fresh install, that the model is tuned toward safe-for-work responses in that state, and that unlocking it requires flipping an "Enable NSFW for App" toggle from a web browser. The site permits mature text between adult characters — tagged "Limitless" — but bans explicit imagery outright, including in avatars.

Janitor publishes more about its moderation stack than most platforms in this category. Every piece of content is screened by automated systems before publication, reported content goes to full-time human moderators, and the company puts numbers on it: 100,000+ items through automated review daily, more than a million automated reviews a week, 20,000+ human moderation actions a month. It is a sponsor of the Coalition for Secure AI, hosted by OASIS Open, and sits on that project's governing board.

There is a dedicated Suicide & Self-Harm Protocol page. It states plainly that responses are AI-generated and are not a substitute for professional help, that expressions of suicidal ideation trigger a crisis referral notification, that the model's responses on self-harm are restricted, and that the company logs the number of crisis referrals issued for compliance reporting. US, UK and Australian helplines are listed on the safety page.

Against that, the changelog is a useful reality check. In August 2026 alone it records the moderation pipeline going down for a period, JLLM issues, a generation slowdown on the paid tier, and chat memory problems that required restoring memory on roughly 15,000 chats. The site still labels itself "beta." Treat it accordingly.

Janitor makes most sense for adults who want long-form collaborative fiction, who are comfortable configuring things, and who understand the trade they are making when they route a conversation through someone else's API key. Creators get a real toolkit — scripts, lorebooks, custom CSS, a verification programme. The free tier is unusually generous for the category.

It is the wrong choice if you need a stable consumer product, if you want your chats to stay in one company's hands, or if you are looking for emotional support rather than fiction. The platform's own crisis page makes the second point better than any review could: the thing you are talking to is not human, and it can be wrong.