Janitor AI review

What Janitor AI is

Janitor AI (janitorai.com) is a character-chat platform run by JanitorAI Inc., a US company incorporated in 2023 and headquartered in Boulder, Colorado. It launched in mid-2023 as a community site where users write and share character “cards” and chat with them, initially by plugging in their own OpenAI or other API keys. It has since added its own hosted model, JLLM, a native iOS app called “Janitor: Interactive Stories”, an Android app with the same name, and, on 24 June 2026, its first paid tier, Janitor+. It is a different animal from Candy or OurDream: the emphasis is on user-written fiction, romance and fandom rather than a single “girlfriend” product, the library runs to millions of cards, and there is no visual side at all.

Janitor’s iOS app, “Janitor: Interactive Stories”, in the App Store, September 2026 (rated 18+).

Chat and roleplay

The free experience uses JLLM, Janitor’s in-house model. Its strengths are that it is uncensored within the terms, cheap to run and tuned for the format the community writes in. Its weaknesses were on display in 2026: an architecture change on 20 April and the removal of a faulty GPU endpoint on 8 May produced widely reported repetition loops, flattened character voices and more refusals, and the “proper fix”, a JLLM V2 trained on a new B200 cluster, had no confirmed release date at the time of writing. Peak-hour queues (US evenings, weekends) slow free replies from a few seconds to a minute or more, though the old 50-message daily cap has gone. The alternative that has always defined Janitor is bring-your-own-model: paste an API key or reverse-proxy URL for DeepSeek, Claude, GPT, Gemini or a self-hosted model, and Janitor becomes a front end for whatever you can afford, with quality to match. Since June 2026, Janitor+ subscribers can do the same without leaving the site through the Router, which lists around 70 models with their context sizes and per-token prices. Our conversation score of 7 is for the house model; with a good external model it would be higher.

Memory

This is the honest weak point. JLLM’s working context is about 8,000–9,000 tokens, shared between the character card, your persona, the “Chat Memory” field, any advanced prompt and the recent messages. The help centre says so plainly and recommends keeping the memory field to concise notes and prompting periodic summaries. In practice a long scene starts losing early details after 25–30 exchanges. Janitor+ multiplies the context five times, to roughly 40,000–45,000 tokens, which is the single most useful thing the subscription buys. An experimental feature that replaced old messages with memory summaries was accidentally switched on for everyone between 18 and 20 August 2026; characters forgot agreements and a related bug cleared memories when messages were deleted, affecting around 15,000 chats. It was reverted and most data restored within two days, but it illustrates that Janitor’s memory system is still under construction. Users on external models get whatever context that model provides.

Models and technology

JanitorAI is unusual in this group for saying quite a lot about its models. The free default is JLLM, the ‘Janitor Large Language Model’, an in-house model served on the company’s own GPUs with a context window of roughly 8–9k tokens. The changelog records an FP8 build for H20 GPUs with 16,384-token support in January 2026, a partial reasoning layer in February, and a full base-architecture swap on 20 April 2026 that the team describes as ‘heavily tuned with high quality gemini and opus data’. What the base architecture actually is has not been disclosed. A short-reply variant, jllm-short, reached general access in July.

janitor+ ($12.99 a month, launched 24 June 2026) does not change the model; it gives JLLM five times the context (about 40–45k tokens by third-party estimate), priority queueing and a monthly allowance of ‘swipes’ on unnamed frontier models. The janitor+ Router, added on 30 July, goes further: around 70 third-party models, including Claude, GPT, Gemini, DeepSeek, Grok and GLM, can be selected inside JanitorAI and billed per token from a built-in wallet, with 10% off popular models. Janitor does not publish the list; it appears in settings.

Before the Router, and still for free users, the same thing is done by hand: the API Settings page accepts any OpenAI-compatible endpoint, and the help centre walks beginners through OpenRouter, DeepSeek, Gemini and Anthropic keys. There is no image, video or voice model at all; the janitor+ announcement names voice and image as future modalities the subscription is meant to fund.

Media

There is none. Janitor generates no images, no video and no voice, and never has. The product is text, and the community has historically preferred it that way: card creators supply their own artwork, and the absence of image generation keeps the platform out of the moderation and payment-processor problems that dog visual services. The Janitor+ announcement mentions “other modalities (think voice and image)” as future plans, with no dates. Anyone who wants pictures of their companion needs a different service or an external image tool; that is why the media score is left blank rather than set to a low number.

Beyond chat: what else it does

JanitorAI is a text platform first, and its extras are mostly about writing, sharing and tooling rather than media.

• Character creation and catalogue: free, public or private bots with definitions, scenarios and example dialogue; creators get analytics dashboards, scheduled releases with countdowns, event tags and, since July 2026, a reopened Creator Verification programme. There is no tipping or revenue share.

• Scripts: the platform’s lorebook system, with keyword triggers, weighted random events and state engines, plus version control with snapshots and diffs. Free.

• Public chats: since March 2026 any conversation can be published, reacted to, forked from any message and continued, with attribution chains and persona switching. Free.

• Memory tools: a manual memory field with a recommended template; an experimental auto-summary was made opt-in in August 2026 after it overwrote chat history.

• Proxy and Router: bring-your-own API keys are free; the janitor+ Router bundles third-party models with per-token billing for subscribers.

• Mobile apps: ‘Janitor: Interactive Stories’ on iOS and Android since the February 2026 beta, with NSFW enabled from the website; janitor+ is sold in-app.

• Localisation and age checks: a multilingual interface since June 2026, and k-ID age verification for Australia, Brazil and the UK.

• Cosmetics: janitor+ adds a golden username, badge, profile styling and GIF comments.

There is no native group chat, no image or video generation, no voice messages or calls, no games or scoring, and no SMS, Discord or Telegram bots. Group scenes are done by prompting, and voice only through third-party browser plug-ins.

The free tier, honestly

It is the most generous in the category. Chat with JLLM is unlimited, there is no card on file, and the full library, character creation, proxies and API keys are all available without paying. The practical limits are the 9,000-token context, queueing at peak times and the current quality of the house model. Janitor states outright that the free version “stays exactly the same, and stays completely free”, and as of September 2026 that holds; the paid tier adds context, priority routing, a monthly allowance of “swipes” (rerolls) on frontier models, Router access and a badge, without removing anything from free users. The one cost the free tier does carry is that your chats go through Janitor’s servers rather than a provider you chose.

Platforms and apps

Web, iOS and Android. The iOS app is listed as “Janitor: Interactive Stories”, rated 18+, iOS 15.5 or later; the Android version has more than a million installs and is rated “Mature 17+” on Google Play, an odd mismatch with the 18+ terms. An APK is also distributed. Store versions are subject to Apple’s and Google’s content rules, so explicit content is more constrained there than on the website; subscriptions bought through the stores must be cancelled through the stores, and the Router is web-only for now. There is a Discord community that doubles as the fastest support channel.

Support, cancellation and refunds

Web subscriptions are cancelled in Settings, then Billing, then Manage Subscription, which opens the Stripe billing portal; the plan stays active until the end of the period and the account reverts to free. Store subscriptions are cancelled in the App Store or Google Play. The FAQ states that subscriptions are “generally non-refundable, except where required by our Terms of Service or applicable law”; the terms add a pro-rata refund only if Janitor itself terminates a paid plan for reasons other than the user’s breach. Router credits are prepaid, billed per token, non-refundable except where the law requires, and cannot be spent while a subscription is inactive; new subscribers could claim $5 of launch credit, expiring after a year. Support is by email (support@ and billing@janitorai.com) and a public help centre; there is no trial of Janitor+.

Who it is for

Janitor is for people who read and write interactive fiction, want a huge library of community characters, and are happy with text. It is the obvious free starting point for anyone new to the category, and the obvious home for anyone who already has an API key. Janitor+ makes sense for heavy JLLM users who want memory and speed without configuring a proxy. It is not for anyone who wants images, video or voice, or who expects a polished single-companion experience out of the box.

Research basis: desk research from the official site, help centre and public user reports, September 2026.