Janitor AI doesn't run its own model on every message. It's a chat front-end that sends your conversation to whichever engine you point it at. Swapping in DeepSeek is a matter of pasting three values into the right panel, and most of the people who get stuck do so because the URL, the key, and the model name came from three different places.
What you're actually connecting
Janitor AI hosts the characters, the cards, and the chat history. The text of each reply comes from a model provider you select. Out of the box that's the site's own JanitorLLM, which queues up during busy hours and can leave you waiting on a one-line response.
Point it at DeepSeek and the replies come from DeepSeek's servers instead. The reason this works at all is that DeepSeek exposes an OpenAI-compatible API. Janitor AI was built around that same request format, so DeepSeek drops in without a plugin. You just change where the request goes.
That's the whole trick, and it's also where setups break. A URL from one provider plus a key from another plus a model name from a year-old video will fail every time. No amount of prompt tweaking rescues a mismatched trio like that.
Key names change, so check yours
Before you copy anything, one warning about model names. DeepSeek's current direct model IDs, as of September 2026, are the deepseek-v4-flash and deepseek-v4-pro entries in the model list. Older tutorials still quote deepseek-chat and deepseek-reasoner, which were the previous generation and no longer match the current catalog.
The people with the most patience for posting tutorials are often the least likely to update them. If your model name doesn't match one on the provider's live list, you get an error that reads like a network problem but isn't one. Copy the identifier from the provider dashboard, not from a blog post.
Pick your route before you open Janitor
Two paths work, and you should decide which one you want before you create any accounts.
The direct route. You make a DeepSeek platform account, add a payment method, and pull a key from DeepSeek itself. The base URL is https://api.deepseek.com. This is the steadier choice if you'll use the setup regularly, and for roleplay-length conversations the cost tends to land in cents rather than dollars, which makes the per-token billing easy to forget until you check the dashboard.
The proxy route. Services like OpenRouter or one of the free-proxy providers give you a key that can reach DeepSeek models through their own endpoint. You get a dashboard and the option to reach several models with one key, but you're now trusting a middle layer, and free tiers can throttle you during peak hours.
One rule covers both: keep every field inside a single route. A DeepSeek key against an OpenRouter endpoint should fail, and it will. When it fails, you know the configuration is inconsistent instead of guessing.
Step 1: Get the three values
- Sign in to whichever provider you picked and open its API keys section.
- Generate a new key. Copy it immediately, because most providers show it once and never again.
- Find the exact model ID and the endpoint URL on that same provider's page. On the direct route that means a
deepseek-v4-flashstyle ID and theapi.deepseek.combase. - Store all three in a note. You'll paste them into Janitor AI in the next step.
Step 2: Enter them in Janitor AI
This is the part where people stall, mostly because the panel is easy to miss.
- Sign in to Janitor AI and open any chat.
- Open the API or proxy settings for that chat. The control sits inside the chat window, not in your global profile.
- Switch the active integration from JanitorLLM to the custom or proxy option.
- Fill the three fields: the model name exactly as copied, the proxy or base URL from the same provider, and the API key that starts with
sk-. - Save the settings.
If Janitor AI shows separate fields for base URL and model, put the base URL only where it asks for a base URL. If it shows one proxy field expecting a full request path, paste the provider's chat-completions form. What matters isn't the slash count. It's that all three values describe the same route.
Step 3: Test small, then tune
- Reload the chat page after saving. Janitor AI reads the configuration at load, so a refreshless page keeps firing at the old model.
- Send a deliberately boring message, something like asking it to confirm the connection in one sentence.
- If a coherent reply comes back, the wiring is done.
Only after the plain test passes should you touch response length and temperature. DeepSeek handles roleplay well at moderate settings. Crank both to the ceiling and replies get unhinged or cut off mid-sentence.
The errors you'll actually see
A network error that won't quit. This is usually not your setup. DeepSeek's API faces regional throttling from some IP addresses, and routing the request through a residential or mobile proxy in a supported region fixes most cases. Plenty of guides skip this entirely. Why? Because it's easier to blame the user's typing than the network they're on.
Model not found. The pasted name doesn't match anything in the provider's catalog. Go back, copy the live identifier, and paste it without trimming characters.
A message that proxies are turned off. Some character cards disable proxies by default. Open the card, or your own character list, and turn Allow Proxies back on for that character.
A 401 or 404. Read the code before you change anything. A 401 points at the key or the account, a 404 at the endpoint or model, and a 429 at quota. Changing prompts won't fix any of them.
Two habits worth keeping
Save the model name and endpoint in a note outside the browser. Switching characters or accounts means re-entering them, and one mistyped character sends you back to the model-not-found error. Tedious, but faster than re-copying from a video.
Check the provider dashboard now and then. Credit balance, remaining free quota, and latency tell you whether the service is still free or has quietly started billing you.
One note on content: Janitor AI hosts unfiltered roleplay characters, and what the model produces depends on the card and your settings, not on which API you connected. If you share your key or endpoint with someone, they can spend your quota.
Where to go from here
Start on the free proxy route with a single short session. Confirm the key, model, and endpoint are pasted correctly, and only then decide whether the direct route's stability is worth funding. Once the connection holds, the rest is tuning length and temperature until the replies sound like the character in front of you.
Verified September 26, 2026. Model names, prices, and free-proxy policies change often, so confirm current values on your provider's site before setting them.
