Introduction
Search for Janitor AI and you’ll find dozens of guides that say the same things: it’s free, it’s for roleplay, connect an API key. Many also repeat numbers that don’t match what the company says itself. One popular claim is a 128,000-token memory for the free model. Another is a daily message cap that never appears in any official post.
This guide takes a different route. I went through Janitor’s official newsroom, its safety page and its help centre, then worked out what the numbers mean in practice. The most useful thing to understand about Janitor AI in 2026 isn’t a feature list. It’s the token budget that decides whether your character remembers anything, plus the age-verification rules and the new Router that most guides skip.
What Is Janitor AI?
Janitor AI is a web and mobile platform where people create and chat with community-built AI characters, mostly for roleplay and interactive storytelling. It is a front end rather than a single chatbot. You can talk to its in-house model, called JLLM (JanitorLLM), or connect a third-party model through a proxy or, since July 2026, through janitor+ Router.
Janitor reported 1.4 million daily active users in its March 2025 update. Its safety page now refers to “millions” of daily users. The company operates as JanitorAI Inc., and the platform is for adults only.
Janitor AI in 2026: A Verified Timeline
Official dates cut through a lot of confusion. Here is what Janitor published this year.
- February 7: the official app entered public beta on the App Store and Google Play, days after Janitor warned about fake apps.
- March 31: public chats launched, letting users publish chats for others to read, react to, fork and continue.
- June 15: an age-verification notice for users in Brazil, Australia and the UK.
- June 24: janitor+ launched at $12.99 a month.
- July 15: the short-response model went to all users.
- July 30: janitor+ Router launched for subscribers on web.
The Token Budget: What Actually Decides Memory
Most guides call JLLM “limited” and stop there. The numbers matter more.
Janitor’s help article on tokens as your AI’s memory budget says the free model’s token “wallet” often holds 8,000 to 9,000 tokens. Every word you type and every word the bot returns uses that wallet. The article also separates two kinds of tokens:
- Permanent tokens are the character’s personality, scenario, advanced prompts and chat memory. They are sent with every message.
- Temporary tokens are the chat itself. They get pushed out as new messages arrive.
Janitor’s guidance is to keep permanent information under about 1,500 tokens. Here is what that means in numbers. These are my calculations from the help-centre figure, not Janitor’s claims.
| Setup | Permanent tokens | Space left for chat (of ~9,000) | Messages that fit at ~150 tokens each |
|---|---|---|---|
| Lean character card | 1,500 | ~7,500 | ~50 |
| Heavy character card | 3,000 | ~6,000 | ~40 |
| Same lean card on janitor+ (5x context, ~45,000) | 1,500 | ~43,500 | ~290 |
Illustrative: real message length varies.
Reply length matters too. On July 15, Janitor released shorter responses to everyone. Choosing Response Length → Short uses the same base model as standard JLLM, and Janitor says replies average 70% shorter. My own reasoning, not a Janitor claim, is that shorter replies also leave more of a fixed window for older messages. If your replies average 300 tokens, a 70% cut brings them to about 90.
Free vs janitor+ vs Router
Free. In its launch post, Janitor says the version people know stays completely free. You get JLLM and access to the community library.
janitor+ ($12.99/month). According to the launch post, you get five times more context, unlimited priority messages, monthly “swipes” on frontier models and a badge. Janitor says the extra capacity needs expensive GPUs and that it priced the plan to hopefully break even. It names voice and image only as future plans, so don’t assume they exist today.
Router. Router lets you use third-party models inside Janitor without copying API keys or configuring a proxy. Per-token prices are shown before you chat, and billing runs through a built-in wallet. At launch it’s web-only and limited to subscribers. Subscribers also get 10% off some popular models and $5 in launch credits. Janitor says JLLM is still its recommended model for everyday roleplay and that Router isn’t meant to replace it.
If you already use proxies, the older route still exists. Router is the easier, more transparent option.
Age Verification: What Changed for Brazil, Australia and the UK
This is the least-covered part of the Janitor AI story. Janitor first blocked UK access in July 2025 after the Online Safety Act. In June 2026 it said users connecting from Brazil, Australia and the UK must verify their age instead.
The company cites Australia’s Age-Restricted Material Codes (in effect 9 March 2026) and Brazil’s Digital ECA (in effect 17 March 2026). It says the alternative was blocking those regions or risking fines of up to $35M in Australia or $9.5M in Brazil.
Verification runs through k-ID and offers three routes: on-device facial age estimation, an ID scan deleted after confirmation, or a reusable AgeKey from another k-ID platform. Janitor says it receives only a yes/no result. The system relies on your IP address, not GPS or your physical location.
For UK readers, our AI Policy & Ethics coverage tracks how these rules are spreading.
Mobile vs Web: Why Content Looks Different
If a character appears on the website but not in the app, that’s by design. Janitor’s mobile moderation post explains that the app defaults to Restricted Mode to meet Apple and Google rules. In that mode, higher-scored characters are blocked and definitions are hidden. You change this in your web profile settings, not in the app.
Safety, Moderation and Privacy
Janitor’s safety page says every piece of content is screened by automated systems before publication, and human moderators review reports. It reports more than 100,000 items reviewed by AI daily and more than 20,000 human moderation actions per month. These are company figures I can’t independently verify.
On privacy, the company states it never sells personal data or shares private conversations for advertising. It also lists itself as a sponsor of the Coalition for Secure AI. The caveat: if you route chats through a third-party model, that provider’s terms apply, not Janitor’s.
Behind the Scenes: Why It Mostly Stays Online
A Janitor engineer wrote an unusually candid infrastructure post in November 2025. At roughly 500,000 concurrent websocket connections, the Node.js cluster used about 120GB of RAM across 30 servers. After a four-month rewrite in Elixir, the same workload ran on three nodes, with edge workers handling stateless requests. The post also mentions migrating a chats table of over 20 billion rows.
That explains why outages still happen but the site survives enormous traffic. The newsroom changelog shows short incidents on September 11 and October 1, and a September 14 write-up covers a week of search problems.
Practical Tips: Getting More From Janitor AI
- Audit tokens first. Check the permanent token count in the bot editor and trim before you blame the model.
- Put essentials in the Scenario field. Load-bearing context belongs there.
- Use Short replies for conversation. Switch to default for scene-setting.
- Summarize regularly. Use Chat Memory to hold plot points, and keep it short.
- Use the test scenario. One character, a twist at message 10, a question about it at message 60.
Common Mistakes
- Using bloated character cards of 3,000+ tokens and wondering why the bot forgets in minutes.
- Pasting unknown “free proxy” links, which can expose your API key.
- Assuming the app and web show the same content.
- Believing blog claims about a 128K free window.
- Installing look-alike apps. Janitor itself warned about fakes in February.
Personal Experience: What Verifying Janitor AI Taught Us
We didn’t run a month of chats on Janitor AI for this guide, and we won’t pretend otherwise. What we did was verify every claim against primary sources. That process taught us a few things.
Official pages beat review sites. The 128K-token figure appears in several guides. Janitor’s own help centre says 8,000 to 9,000 for JLLM. A third-party source separately describes 128K as a site-wide cap for proxy models, which is a different thing.
Beware “independent” sites with affiliate links. One top-ranking page was framed as an independent review based on official announcements, but its links pointed to a competing service. It still helped us locate Janitor’s real announcements, so we followed those and not its commentary.
Official posts show priorities. The newsroom lists app updates, safety, creator tools and infrastructure, while voice and image generation remain “plans.” That tells you more about the product than any feature table.
The lesson: run the 20-minute test yourself. One character card under 1,500 tokens, a twist at message 10, a question about it at message 60. If the bot remembers, your setup works. If not, trim tokens before upgrading. See our Character AI vs FictionLab AI vs Janitor AI comparison for how it compares to other platforms, and our Editorial Process for how we verify claims.
FAQs
What is Janitor AI?
Janitor AI is an adults-only platform for chatting with and creating AI characters, available on web and as an official mobile app. It runs an in-house model (JLLM) and supports third-party models.
Is Janitor AI free?
Yes. Janitor says the version people know stays free. The paid janitor+ plan costs $12.99 a month and adds more context, priority messages and enhanced swipes.
How big is Janitor AI’s memory?
Janitor’s help centre says the free JLLM often holds 8,000 to 9,000 tokens. janitor+ advertises five times more context, which works out to roughly 45,000 by my arithmetic.
What is janitor+ Router?
It’s a subscriber-only, web-only feature launched July 30, 2026, that lets you use third-party models inside Janitor with per-token billing and no API key or proxy.
Why is Janitor AI asking me to verify my age?
If you connect from Brazil, Australia or the UK, local laws require it. Verification goes through k-ID, and Janitor says it only receives a yes/no result.
Why can’t I see some characters in the app?
The app starts in Restricted Mode to meet app-store rules. Change your content settings on the website to sync mobile with web.
Does Janitor AI have a mobile app?
Yes, the official app entered public beta on February 7, 2026. Fake apps exist, so only download it through links on Janitor’s own site.
How do I get shorter replies?
Open the model settings and choose Response Length → Short. Janitor says replies average 70% shorter.
Is Janitor AI safe?
It has moderation, an 18+ rule and a privacy commitment, but safety depends on your setup. Third-party models and proxies handle your chats under their own terms.
Conclusion and Actionable Takeaways
Janitor AI in 2026 is less about a single chatbot and more about a platform with a memory budget, a growing set of paid options and real regional rules. Once you understand the token budget, most “my bot forgets” problems become fixable.
- Trim permanent tokens to about 1,500 before paying for anything.
- Try Short replies to stretch memory.
- Treat janitor+ and Router as convenience upgrades, not requirements.
- Check age-verification rules if you’re in Brazil, Australia or the UK.
- Change content settings on the web if the app looks restricted.
- Verify any number you read against Janitor’s official newsroom.
For more verified coverage, see AI LLM News and our AI Tools & Reviews section.
Disclaimer: This article is for general information only and isn’t legal, financial or mental health advice. Janitor AI is intended for adults, and some content is mature. Prices, limits and policies change often and vary by region, so confirm current details on Janitor’s official site before paying. We’re not affiliated with Janitor AI. Company statistics are self-reported and unverified. If an AI chat habit is affecting your wellbeing, consider speaking to a professional.