What "proxy" means here
The term Janitor AI API proxy is confusing because the word proxy has other uses in networking. On Janitor AI, it usually refers to the option that lets you use an LLM API of your choice instead of the site's default model. You don't need to set up any intermediate server.
In practice, the site sends the conversation to the address you entered, in the OpenAI chat format, and displays the response. Any service that supports this format can take its place, and this API is one such service. The labels and field positions change over time, so use the text below as a general map and consult the Janitor AI help if there is a difference.
The same logic applies outside Janitor AI. SillyTavern, for example, also allows adding an OpenAI-compatible endpoint, and the same URL, key, and model work there. Learning the process once is useful for multiple tools, which is why it is worth understanding what each field does, rather than just copying values.
The four pieces of data you need
| Data | Value | Note |
|---|---|---|
| API URL | https://api.llmsemcensura.com/v1 or the full form with /chat/completions at the end | The field on the site indicates which form it expects |
| Key | your API key | Obtained on the key page |
| Model | uncensored | Available unique ID, in lowercase |
| Context | 8,000 to 16,000 tokens | The context window reaches 100,000, but large context costs more |
The new key can be tested at no cost: after registration, one click on the key page redeems US$ 0.50 in free trial credit, once per person, valid for 7 days.
Three-step diagnosis
When the connection does not work, the error may be in the key, URL, model, or site. Separate the causes by testing outside Janitor AI.
Step 1: the key and the network
List the models. If the response comes in JSON, the key and base URL are correct.
# 1) A chave e a rede estão certas? (modelo esperado: uncensored)
curl https://api.llmsemcensura.com/v1/models \
-H "Authorization: Bearer SUA_CHAVE"Step 2: the chat method
Send a minimal call. The response should return the text in choices[0].message.content.
# 2) O método de chat responde com o modelo uncensored?
curl https://api.llmsemcensura.com/v1/chat/completions \
-H "Authorization: Bearer SUA_CHAVE" \
-H "Content-Type: application/json" \
-d '{"model": "uncensored", "messages": [{"role": "user", "content": "Diga apenas: ok"}], "max_tokens": 10}'Step 3: what the server returned
If something fails, check the HTTP code and the error body. The same test in Python also shows the X-Request-Id, useful if you need to contact support or look up the documentation, and the remaining balance in X-Balance-USD.
import os, requests
BASE = "https://api.llmsemcensura.com/v1"
H = {"Authorization": "Bearer " + os.environ["API_KEY"]}
r = requests.post(
BASE + "/chat/completions",
headers=H,
json={"model": "uncensored", "messages": [{"role": "user", "content": "Diga apenas: ok"}], "max_tokens": 10},
timeout=60,
)
print(r.status_code, r.headers.get("X-Request-Id"), r.headers.get("X-Balance-USD"))
print(r.text[:300])A detail that catches many people: when pasting values, the browser or editor may include a space or a line break at the end. For the URL and the key, this is enough for the request to fail silently. Paste into a notepad, check, and only then copy to the site.
If curl and Python work but Janitor AI does not, the problem is in the site fields: the form of the URL, the model name, or some extra space pasted along with the key.
Common errors and likely cause
- Network failure immediately upon sending. Typo in the URL or missing
https://. Compare with the one that worked in curl. - Authentication error. Incomplete key or with a space. Paste again and check the beginning and end.
- Model not found. Names like those from other companies do not exist on this endpoint. Use only
uncensored. - HTTP 429 with
concurrency. More parallel requests than the key allows: 2 on the test key, 8 on the full key. Avoid generating in multiple tabs. - HTTP 429 with
rate_limited. Exceeded the per-minute limit: 60 on the test key, 300 on the full key. Wait a bit. - 429 with
Retry-After: 5. Temporary overload. Wait 5 seconds and try again. - Response cut off. The site's new token limit is low. Increase it in the generation settings.
Quick list before opening the chat
- Did the curl from step 2 return text?
- Is the URL on the site exactly the one that worked, in the form the field requires?
- Is the model written as
uncensored? - Is the context at a reasonable value, from 8,000 to 16,000 tokens?
- Is there a balance? The
X-Balance-USDanswers that question.
Beware of third-party free proxies
Free "proxies" circulate in roleplay communities, where someone provides an address for everyone to use. Think carefully before: you don't know who operates the intermediary, what records they keep, or if the service disappears tomorrow. There is also the risk of sending your conversations to a stranger. Using an API with your own key, where you know who provides the service and how it is billed, avoids this problem. On this API, prompts are not used for training.
For security, use a key only for Janitor AI and keep a small balance on it. Never paste the key into screenshots, videos, or public chats.
Context, cost, and quality
The price is $0.15 per 1M input tokens and $0.50 per 1M output tokens. In roleplay, input dominates because the site resends the character sheet and history with every response. Example with 12.000 context tokens and a 400-token response:
- Input: 12,000 x 0.15 / 1,000,000 = US$ 0.0018
- Output: 400 x 0.50 / 1,000,000 = US$ 0.0002
- Total: US$ 0.002 per message
The $0.50 free trial credit gives 250 messages on this profile. A $20 top-up gives 10,000. Reducing the context to 6,000 tokens cuts the input cost in half to $0.0009, and the total drops to $0.0011. Less context also makes responses faster, at the cost of the character remembering less of the history. Top-up ranges are on prices.
Service limits
The model is open, hosted by an independent provider, and responds without content refusals for legal adult use. It is not the official API of any large company and does not route between models. It generates text only, without image, audio, or embeddings. Sexual content involving minors is prohibited, without exception, and failed or refused requests are not charged.
To top up, amounts range from US$ 20 to US$ 500, in USDT on the TRC20 network or USDC on the Base network, with a 5% bonus from US$ 50, 10% from US$ 100 and 15% from US$ 200. Want to better understand what changes when a model has no filters? Read uncensored AI.

