Janitor AI not working: how to actually fix it

Quick answer: check these four things first

"Janitor AI is not working" almost never means the entire platform is down for everyone at once. In order of how often each one turns out to be the real cause: the built-in backend is overloaded during a peak evening window and queuing or timing out silently, your browser is holding a stale session or a cached bundle from before the last update, an ad blocker or privacy extension is quietly blocking the fetch calls the chat page depends on, or, if you have a custom API connected, the key behind it is dead, rate-limited, or misconfigured. Work through those four in order before you clear your whole browser or wait out an outage that may not even be happening. A hard refresh, a private window, and the 10-second connectivity test further down this page isolate the actual cause faster than reinstalling anything.

Why Janitor AI breaks down more than it looks like it should

None of this means your account or your setup is broken. It means a chat app built on a shared backend, a growing conversation, and a live network connection has more failure points than "is the website up" suggests.

• One shared backend, every free user: the built-in model runs on infrastructure sized for average load, not peak load. A US evening surge pushes it into queuing or dropped requests that look identical to "the site is broken."
• Chats grow past what a request can carry: every message resends the character card, persona, and recent history. Past a certain size, a single request can cross the backend's timeout window and simply never return, with no error shown.
• A filtered reply can come back empty instead of visible: some moderation layers send back nothing rather than a warning message, so a blocked response reads as a hang instead of a block.
• Extensions block requests without telling you: ad blockers and privacy tools that stop third-party network calls can silently drop the exact requests JanitorAI's chat page depends on.
• Sessions expire while the page still looks logged in: multiple open tabs, a long idle period, or a backend deploy can invalidate your session token in the background.
• Mobile browsers throttle backgrounded tabs: switching apps mid-generation on a phone can suspend the tab and kill the request before a reply lands.

The math: escaping a shared key costs cents, not dollars

If your diagnosis lands on "the same proxy dies every evening," the actual fix is not a bigger workaround, it is a personal key so your quota is not split across everyone else hitting the same free pool at the same hour. Roleplay chat is input-heavy: JanitorAI resends the character card, persona, and recent history with every message, so 6,000 input tokens and 300 output tokens per message is a realistic baseline for a detailed card and a mid-length chat. Run that baseline at a heavy 80 messages a day for a month and the difference between models tells the whole story. On a budget model, a personal key that never queues behind a crowd is a rounding error next to what a flaky free proxy already costs you in frustration.

Why "it's not working" catches people off guard

Each of these has a specific, boring fix. None of them require a support ticket or a reinstall, and most resolve inside the time it takes to read the next section.

• "Down for me" is not always down for everyone: peak-hour congestion on a shared backend is uneven, so your chat can queue for a minute while someone else's loads without a hitch.
• A 429 looks exactly like an outage: a rate limit on a shared or free key produces the same blank result as a real server failure, but the fix is different, wait or switch keys, not different infrastructure.
• Extensions rarely show an error: a blocked fetch call usually never resolves rather than failing loudly, so the page looks stuck instead of broken.
• Your own chat can be the actual cause: a card and history that quietly crossed a token or timeout limit behaves like a platform bug, but the fix is trimming the conversation, not waiting for a status update.
• Logging out fixes more than it gets credit for: a stale session token is invisible until you sign out, because the page keeps rendering as if you were still authenticated.

Default model vs a connected API: which one fails, and how

The two modes fail differently, which is useful once you know which one you are diagnosing. The built-in model shares one backend with every free user on the platform, so its failure mode is congestion: queuing, timeouts, and a generic error message with no code attached. A connected API, whether a provider's own endpoint or a gateway, fails the way any HTTP API fails: with a status code you can look up and a cause you can isolate in one curl request.

The fix list, in the order that actually works

Work down the list until one step changes what you see. Most people never get past the third item.

• Blank screen, nothing happens when you hit send: disable ad blockers and privacy extensions first, or open the chat in a private window. A blocked fetch call is the most common silent cause and never shows an error.
• "Error occurred, please try again" on every message: hard refresh and clear site data before anything else. A stale cached bundle from before the last update behaves exactly like a broken backend.
• A reply that comes back completely empty: shorten or edit your last message. A filtered response often comes back as nothing rather than a visible warning, so an empty bubble usually means the message got blocked, not that the app hung.
• Worked yesterday, fails today, only in one long chat: start a fresh chat or trim the persona and card. A conversation that quietly crossed a token or timeout limit fails the same way on every message until you shorten it.
• Stuck on a spinner right after logging in fine: log out completely and back in. A session token can expire in the background while the page keeps rendering as if you were still signed in.
• On proxy mode and it stopped overnight with no changes on your end: test the endpoint outside JanitorAI first with the curl command below, then match the error code against the table in the next section.
• The same proxy dies every evening at the same hour: that is a rate limit on a shared or free key, not a real outage. The durable fix is a personal pay-as-you-go key so your quota is not split across a crowd; pointing the proxy field at an OpenAI-compatible gateway such as APIsRouter is a settings change, not a rebuild, and budget models cost a fraction of a cent per message.

Confirm it is not your proxy: a 10-second test

If you have a custom API connected, isolate it before touching JanitorAI's settings again. The curl call below hits the chat completions endpoint directly, using APIsRouter as the example endpoint: swap in your own key, URL, and model, and a clean JSON reply back means your proxy is fine and the "not working" report is happening somewhere else in the chain. No reply, or an error code, points straight at the row to fix in the table below.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer sk-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Say ready."}],
    "max_tokens": 20
  }'

FAQ