Janitor AI help: fixes, proxies and getting it to talk
Janitor AI slow, stuck in a queue or not working? A step-by-step fixer, a DeepSeek and OpenRouter proxy setup, and how to have Janitor AI speak to you.
Janitor AI is brilliant right up until it isn't. The bot you've been building a scene with for an hour goes quiet, the reply spinner turns for a full minute, or the page just refreshes itself in a loop. It happens to everyone, and nearly always for one of a handful of reasons.
So start here. Pick what's going wrong, try the most likely fix first, and only move on if it doesn't work.
Fix 1 of 4
Check the clock
The free model is shared by everyone on the site, and US evenings are its rush hour. A reply that lands in seconds in the morning can take minutes at 9pm Eastern. If you can, play earlier in the day.
Ask for shorter replies
Turn down max new tokens in the generation settings. A shorter reply finishes sooner, and you can always ask the character to keep going.
Stop hammering regenerate
Every retry is another request to the same busy pool. Give a slow reply a minute before you swipe again.
Skip the queue
Janitor+ gets priority access when the servers are busy. Or bring your own model through a proxy, which never waits in the free queue at all.
Set up a proxy
Fix 1 of 5
Check it isn't down for everyone
Outages and maintenance usually get announced on Janitor's Discord and social accounts, and status sites like Downdetector show whether other people are stuck too. If it's down, nothing on your side will fix it.
Try a private window
A private or incognito window loads the site without your extensions or saved cookies. If it works there, the problem is on your browser's side.
Turn off extensions
Ad blockers, voice add-ons and script managers are the usual culprits for a page that refreshes in a loop. Switch them off, reload, then turn them back on one at a time.
Switch off your VPN
Janitor checks for bots before it lets you in, and some VPN servers get stuck in that check over and over. Try without it, or pick a different server.
Check your broadband filter
Some UK and other providers block adult sites by default, and Janitor can land on that list. If it never loads on home wifi but works on mobile data, that's your provider's filter, which you can switch off in your broadband account.
Fix 1 of 5
Check the model name
It has to match the provider's model ID exactly, capitals and all. Providers rename and retire models often, so copy the current ID from their model page rather than an old guide.
Check the URL
Janitor wants the full chat-completions address, not the provider's home page. For OpenRouter that's https://openrouter.ai/api/v1/chat/completions, and for DeepSeek's own API it's https://api.deepseek.com/chat/completions.
Check your balance and limits
A paid key with no credit fails, and free models come with daily request caps. An error that mentions 429 or a rate limit means wait, or top up.
Save, then refresh
Save the configuration, reload the chat and make sure that configuration is the one selected. A stale page can keep using the old settings.
Make a fresh key
If the key was ever pasted somewhere public, revoke it on the provider's site and make a new one. A leaked key can be drained by strangers or switched off by the provider.
Fix 1 of 5
Raise max new tokens
It caps how long a reply can be. Raise it a step at a time; longer replies take longer to arrive.
Adjust the temperature
Temperature controls how adventurous the writing gets. Most guides settle between about 0.8 and 0.9; much lower turns repetitive, much higher starts to ramble.
Steer it out of character
Put a note to the model in double brackets, like ((OOC: describe the room before she answers)). It reads the instruction without your character saying it.
Put key facts where they're always read
The free model remembers roughly 9,000 tokens of chat, so long scenes lose their start. Put names, relationships and rules in your persona or the bot's description, which stay in view.
Try a stronger model
The house model is free but fairly basic. A bigger model through a proxy usually writes longer and follows the scene better.
Set up a proxy
Fix 1 of 3
Use your device's read-aloud
Janitor is text first, with no voice of its own. Every phone and browser can read a page aloud: Speak Screen on iPhone, Select to Speak on Android, and Read aloud in Microsoft Edge.
Add a voice extension
Browser add-ons can read each new reply in a chosen voice, and some connect to voice services such as ElevenLabs. Your messages then go to that add-on's company too, so read its privacy policy first.
Make it speak to you by name
If you meant the character should talk to you, set up a persona with your name and a line about who you are. Bots use {{user}} for your persona's name, so it shows up in their replies.
Why is Janitor AI so slow?
Because the free model is a shared kitchen. Janitor's own model, JanitorLLM, costs nothing to use, and every free message from every user lines up for the same pool of servers. When a big chunk of the user base is online at once, usually US evenings, the line gets long. That's the queue or waiting room people talk about, and it's why the same message can come back in five seconds one morning and a minute that night.
You've got three ways round it:
- Free
- The house model, unlimited chats on the standard models, about 9,000 tokens of memory. Slowest at busy times.
- Janitor+
- Reported at $9.99 a month. About five times the memory and priority access when the servers are busy.
- Your own model
- Plug in DeepSeek, Claude, GPT and others through a proxy. You pay the provider per message, and you skip the free queue.
If speed is the only problem, try quieter hours first. If memory bugs you as well, a proxy fixes both.
How to set up a proxy (DeepSeek and OpenRouter)
A proxy here just means Janitor sends your chat to another company's model instead of its own. You bring an API key, which is a password for that company's service, and you pay them for what you use. DeepSeek is the popular pick because it's cheap and writes well; OpenRouter is handy because one key reaches dozens of models, DeepSeek included.
- Get a keySign up with the provider (OpenRouter, or DeepSeek directly), add a little credit if the model needs it, and create an API key. Copy it somewhere private.
- Open API settingsIn a Janitor chat, open the settings menu and find the API settings. Choose the proxy option and add a new configuration.
- Fill in three fieldsThe full chat-completions URL, your key, and the model ID exactly as the provider writes it.
- Save and testSave, refresh the chat, select your configuration and send a message. If it errors, the fixer above has the usual culprits.
Menus move between updates, so if a label doesn't match, Janitor's help centre has a current walkthrough. Two things don't change, though:
Which model is best for Janitor AI? There's no single answer, and the line-up changes monthly. DeepSeek's chat models are the usual value pick, while the big names like Claude and GPT cost more per message. Start cheap, compare a few replies in the same scene, and only pay more if you can feel the difference.
How to have Janitor AI speak to you
People mean two different things by this, so here are both.
If you want to hear it. Janitor doesn't do voice. It's text first, with no voice notes or calls. But your phone or browser can read any reply aloud, and that's free and private. On an iPhone, turn on Speak Screen in the accessibility settings and swipe down with two fingers. On Android, Select to Speak does the same. In Microsoft Edge, Read aloud is in the menu. Voice extensions go further and read each reply automatically in a voice you choose, but they pass your messages to another company, so treat them like any other app you'd share a chat with.
If you want it to talk to you, not at you. Make a persona: a name and a few lines about who you are in the scene. Bot creators write {{user}} where your name goes, so a good bot will use yours, react to your description and stop treating you as a stranger.
If voice is the whole point, Janitor is the wrong tool. Companion apps such as Fantasy.ai send voice notes and even call, and our AI sexting apps guide sorts the options by words, pictures or voice.
Commands and formatting
Most of steering a Janitor chat comes down to a few conventions every bot creator uses:
- *asterisks*
- Actions and description: *she leans on the bar*.
- "quotes"
- Spoken dialogue.
- ((OOC: …))
- A note to the model, outside the story. Use it to steer the scene or fix a mistake.
- {{char}}
- Stands in for the character's name, in bot descriptions.
- {{user}}
- Stands in for your persona's name.
Can Janitor AI make images?
No. Janitor is a text roleplay site. It doesn't generate pictures, voice notes or video of its characters. If you want a character who sends selfies, look at companion apps like Candy AI. If you want to make the pictures yourself, our NSFW AI prompt guide covers the generators.
Keep it private, keep it adult
Two last things. Your chats live on Janitor's servers, and if you use a proxy they go to that provider too; our guide to who can read your AI chats breaks down who sees what. And Janitor has zero tolerance for minor-coded characters and personas, with permanent bans for breaking it. Keep every character a clear adult and you'll never meet that filter.
Questions people ask
Why is Janitor AI not working?
Usually one of four things: an outage or maintenance, a browser extension or VPN interfering, a broadband adult filter blocking the site, or a broken proxy setting. Check Janitor's Discord or a status site first, then try a private window with extensions and VPN off.
Why is Janitor AI so slow?
The free model is shared by everyone, so replies queue up at busy times, especially US evenings. Shorter replies, quieter hours, Janitor+ priority access or your own model through a proxy all help.
How do I make Janitor AI faster?
Lower max new tokens so replies are shorter, chat outside US evening peak hours, and avoid repeatedly regenerating. For a lasting fix, connect your own model through a proxy, which skips the free queue.
How do I have Janitor AI speak to me?
Janitor has no built-in voice, but your device can read replies aloud: Speak Screen on iPhone, Select to Speak on Android or Read aloud in Microsoft Edge. Voice browser extensions can read each reply automatically. If you want the character to address you, set up a persona with your name.
How do I use DeepSeek with Janitor AI?
Create an API key at DeepSeek or OpenRouter, then in a Janitor chat open the API settings, add a proxy configuration and enter the full chat-completions URL, your key and the exact model ID. Set a spending limit on the provider's dashboard.
Is there a free API key for Janitor AI?
Not a safe one. Free keys shared online are usually stolen and get shut off. Janitor's own model is free, and some providers offer free models with daily limits on your own account.
Can Janitor AI generate images?
No. Janitor AI is text roleplay only, with no pictures, voice notes or video. Companion apps such as Candy AI send pictures, and image generators let you make your own.
