What “proxy” means on Janitor AI (it’s not a VPN)
On Janitor AI, a proxy is a bridge to a better brain. The community’s own pinned guide defines it exactly that way: you connect an outside language model to Janitor with three strings, an API URL, a model name, and an API key, and your roleplay runs on that model instead of the free built-in JanitorLLM. That is the entire concept.
It is not an IP proxy and not a VPN, which is worth saying because two of the top pages ranking for this exact search are network-proxy vendors selling IP masking that has nothing to do with Janitor. If a page starts talking about “bypassing geo-restrictions” or “rotating residential proxies,” it is answering a different question. So does the top-voted reply in the beginner thread this SERP ranks first, which recommends a residential-proxy company for what is actually an API question.
What you need before you start
Three values, all pasted, never typed:
- API URL: the provider’s chat completions endpoint. It is an address for Janitor to call, not a link for you to click.
- Model name: the exact model string from the provider’s list. One typo, like “conpletions” for “completions,” and you get an error; a user in the beginner thread lost an evening to exactly that.
- API key: shown once at creation. Copy it immediately and store it somewhere safe. Treat it like a password.
One setting matters more than people expect: context size. The community’s long-standing answer is 16k, and their pinned guide’s phrasing has become a meme: “16k. Yes, really!” Bigger numbers mostly make the model slower and more forgetful at the edges, not smarter.
Set up your first proxy, step by step
The flow below matches JanitorAI’s official help-center guide, updated December 2025, using OpenRouter as the example provider. One honest caveat before you start: Janitor has been changing its API settings screen, and the community megathread’s own comments note the instructions do not always match the current interface. The names below are the stable parts.
- Create an account at openrouter.ai and generate a key under Keys. It starts with
sk-or-v1-and is shown once. - In a Janitor chat, open the menu and go to API Settings, then choose Proxy and the Custom configuration option.
- Paste the API URL:
https://openrouter.ai/api/v1/chat/completions - Paste the exact model name from OpenRouter’s model list, copied from the page, never typed.
- Paste your key, with no trailing spaces.
- Skip the check button. The community guide is blunt about the “Check API Key/Model” button being unreliable; the real test is sending a message.
- Save, then hard-refresh the page. The official guide calls the refresh “a crucial final step” for a reason.
One thing my verification found that no guide mentions: the official help doc still recommends a specific free DeepSeek model string, and that model is no longer on OpenRouter’s free list. I pulled OpenRouter’s live model index on July 26, 2026: 15 free models remain, and none of them are the uncensored roleplay favorites the guides were written around. Pick your model from the live list, not from a guide, including this one.
Other providers that work
DeepSeek official: the cheapest paid route
When the free DeepSeek variants left OpenRouter, the community’s answer was to go direct: https://api.deepseek.com/v1/chat/completions with a small top-up, community-reported at a $2 minimum. The recurring verdict in the threads is that a few dollars lasts a long time, and “if you’re gonna use DeepSeek, just pay.”
Google AI Studio (Gemini): the biggest free quotas
Google’s AI Studio issues free API keys for Gemini models with per-model daily caps. Community write-ups from mid-2026 reported daily allowances in the hundreds for the flash models; Google has since moved the exact numbers behind its live rate-limit pages, so treat any specific figure you read, including old ones like “100 a day for Pro,” as dated and check the live page. Gemini’s filter behavior on explicit roleplay varies by model and prompt, which the community works around with custom prompts; that is their report, not my test.
Chutes: what changed when it went paid
Chutes was the community’s free workhorse until it “swapped to a payment plan” in 2026, in the megathread’s words: a mandatory top-up, community-reported at $5 for a daily message allowance. The lesson the megathread drew is the one to remember for every provider on this page: “Free things rarely stay free.”
Janitor AI free proxy options, compared
| Route | Cost as of Jul 26, 2026 | The catch |
|---|---|---|
| Built-in JLLM | Free, no setup | The model everyone is trying to upgrade from |
| OpenRouter free models | Free, daily caps | The uncensored favorites are gone from the free list; what remains is corporate-aligned models |
| Google AI Studio (Gemini) | Free tier, per-model daily caps | Caps move; filter behavior varies by model |
| DeepSeek official | Paid, small top-up (community-reported ~$2 minimum) | Not free, but the community’s price-performance pick |
| Chutes | Paid top-up (community-reported $5) since its 2026 change | The precedent for why free routes die |
Quota numbers older than a few weeks are unreliable in this ecosystem; the community’s own megathread went stale within months. Where I could not verify a number against a live official page on July 26, 2026, I have labeled it community-reported.
Janitor AI proxy errors: 429, 401, 404, and how to fix each
This is the part no article on page one has: the error strings Janitor actually prints, and what each one means. Sources: the community’s pinned proxy megathread, the official troubleshooting doc, and the error reports inside both threads.
| Error | What it means | The fix |
|---|---|---|
429 rate limited | You hit the provider’s daily or per-minute cap | Wait, or raise the cap by adding credit with the provider. A blocked generation still burns your daily count on some providers, which is why the community treats 429 as the worst one |
401 invalid token / no auth credentials | Wrong key for the provider, or a key pasted with spaces | Check the key prefix matches the provider (OpenRouter keys start sk-or-v1-), re-paste with no trailing spaces |
400 not a valid model ID | The model string is wrong | Copy the exact string from the provider’s live model list |
402 quota exceeded | The provider wants payment | Top up or switch providers |
404 no matching endpoint | Wrong API URL, a dead model name, or a provider data-policy setting blocking free models | Re-paste the URL; re-check the model exists; on OpenRouter, review your privacy settings, which gate some free models behind a data-sharing opt-in |
500 / 503 no instances available | The provider is overloaded, not you | Wait it out; “the problem you face is pretty much a very small service being overloaded,” as one megathread reply put it |
| ”A network error occurred / Failed to fetch” | Usually a typo in the URL, sometimes the provider being down | Re-copy the URL from the source. The user who coined this thread’s best confession had written “conpletions" |
| "No response from bot” | A reasoning model returned its thinking instead of a reply | Set max new tokens to 0 in Janitor’s settings, the megathread’s standard fix for reasoning models |
If an error string is not in this table, the pattern still holds: 4xx codes mean something on your side or your account, 5xx codes mean the provider is struggling, and network errors mean a typo until proven otherwise.
Are proxies safe? Where your chats actually go
Every proxied message leaves Janitor and lands on your provider’s servers under that provider’s privacy terms. Janitor’s own privacy policy, which I read in full for my Janitor AI safety read, does not address proxied chats at all. The practical rules: use a provider with a data policy you have actually opened, never use a shared or scraped key from a stranger’s paste, and treat “free proxy” links in comment sections as scams by default. The community megathread’s own disclaimer says the same: “be vigilant about free proxy scams.”
If you’re done fixing error codes
Fair enough. The apps below are what I recommend when someone wants the uncensored roleplay without the API keys, the rate limits, or any of this page: I paid for each one myself and documented the tests.