Janitor AI Custom API Setup: Proxy URL, Key, Model
Janitor AI needs your own API for better models. This guide covers the four API Settings fields, a free test model, 401/404 fixes, and Taiwan invoices.
Query date: 2026-09-09 (Taipei time). The field names and settings locations for Janitor AI follow its official help center as of that day. Platform interfaces change often, so check the actual fields on your screen before you configure. Model unit prices are taken from the BazaarLink model catalog on the same query date.
Once you play Janitor AI for a while, most people hit the same issue: to switch to a better model, you have to connect your own API. The settings page asks for a URL, a key, and a model name. If any one field is wrong, you get a red error, and the error message usually does not tell you which field caused it.
This article explains those three fields clearly and covers the two things Taiwan users care about most: whether you can run a free test first, and how to get an invoice when your company needs to expense it.
Bottom line first: The settings are in API Settings → + Add Configuration on the chat screen, and you fill in four fields: Configuration Name, Model, Proxy URL, and API Key. As long as your provider offers an OpenAI-compatible endpoint, these four fields will work. With BazaarLink, you can first run the flow with a free model, confirm all four fields are correct, and then decide whether to top up for a stronger model.
What Each of the Four Fields Does
| Field | Meaning | Symptom when wrong |
|---|---|---|
| Configuration Name | Just a label for you; any name works | Does not affect the connection |
| Model | The model ID, which must exactly match the string in the provider's catalog | Usually returns 404 or "model not found" |
| Proxy URL | Where the request is sent | Usually returns 404, or nothing responds at all |
| API Key | Your key, which determines whose account the usage is billed to | Returns 401 |
You can create multiple configurations and switch between them at any time, so you do not have to keep overwriting the same one when comparing models.
How to Fill It In with BazaarLink
- Sign up at BazaarLink and create an API key on the keys page (the format begins with
sk-bl-). - Go back to Janitor AI and open API Settings → + Add Configuration.
- Fill in the four fields like this:
| Field | What to enter |
|---|---|
| Configuration Name | Anything, for example BazaarLink Free |
| Model | qwen/qwen3.7-flash:free (start with the free model to get it working) |
| Proxy URL | https://api.bazaarlink.ai/v1/chat/completions |
| API Key | The sk-bl-... key you just created |
- Save, then send a test message. If you get a reply, the connection works.
How far down should the Proxy URL go? Janitor AI's official example uses the full
/chat/completionsendpoint, so the table above uses the full path. If your screen version only accepts up to/v1, enterhttps://api.bazaarlink.ai/v1instead. Try both once; whichever works is correct.
Run a Free Test First, Then Decide Whether to Pay
You do not have to top up at the start. These two models are zero-cost:
| Model ID | Description |
|---|---|
qwen/qwen3.7-flash:free | Free model with a 1-million-token context |
auto:free | Automatically routes to whichever free model is available at the time |
Use a free model to confirm the four fields are correct before switching to a paid model. That way, if something goes wrong, you know the problem is not in the settings but somewhere else.
Choosing a Better Model: Prices and How to Pick
Roleplay has one defining trait: every turn resends the earlier conversation, so input tokens build up much faster than in normal chat. When choosing a model, give the input price more weight.
Prices below are the unit prices as of the 2026-09-09 query (USD per million tokens):
| Model ID | Input | Output | Best suited for |
|---|---|---|---|
deepseek-v4-flash | $0.20 | $0.40 | Long conversations, keeping costs down; the starting point for most people |
gpt-5.6-luna | $0.20 | $1.20 | Cheap input, stable storytelling |
gemini-3.7-flash | $0.75 | $3.75 | Fast responses, large context |
grok-4.6 | $2.00 | $6.00 | A livelier tone |
claude-sonnet-4.6 | $3.00 | $15.00 | Best long-form plot consistency, but the most expensive |
For the latest prices, see BazaarLink Models. To estimate how much a long conversation will cost, see the cost section in our SillyTavern guide; the method is the same.
Three Most Common Errors
- 401: The key was mistyped, or there are extra spaces before or after it. Paste it again, and make sure you do not copy a leading space in front of
sk-bl-. - 404 or "model not found": The string in the Model field does not match the catalog. The model ID must match character for character, including suffixes such as
:free. - Nothing responds at all: The Proxy URL is wrong. Following the explanation above, try both the full endpoint and the
/v1-only version.
Taiwan Expense Reporting: Company Tax ID and Electronic Invoices
This is the most practical difference between us and typical overseas providers. If you pay an overseas service by card, you receive an English receipt, and your accountant will often send it back when you try to claim the expense.
When you top up through BazaarLink, you can use your company tax ID, receive a Taiwan electronic unified invoice, and pay in NT$. If the money is going through company accounts, this saves a lot of back-and-forth. For details, see How to Expense AI API Costs at a Taiwan Company.
If you are an individual player and do not need an invoice, you can ignore this section; the setup is exactly the same.
Other Frontends That Also Require Your Own API
The same key can be shared across different frontends without applying again:
- How to Connect an API to SillyTavern in Taiwan: the most features, but you host it yourself
- RisuAI Beginner's Guide: no install, works on your phone
- What Is Tavern AI? A 10-Minute Beginner Guide: start here if you are completely new
Sources and Query Notes
- Janitor AI official help center, proxy setup tutorial (query date 2026-09-09): field names Configuration Name / Model / Proxy URL / API Key; settings location API Settings → + Add Configuration; the official example uses an endpoint in the
/v1/chat/completionsform - BazaarLink model catalog
/api/v1/models(query date 2026-09-09): all model IDs, context lengths, and unit prices in the tables above
FAQ
What should I enter as the Proxy URL in Janitor AI?
Enter your provider's OpenAI-compatible endpoint. With BazaarLink, enter https://api.bazaarlink.ai/v1/chat/completions; if your screen version only accepts up to /v1, enter https://api.bazaarlink.ai/v1 instead. The official example uses the full /chat/completions form, so try both once and use whichever one works.
Where do I set up my own API in Janitor AI?
In the chat screen's API Settings, click + Add Configuration, and four fields will appear: Configuration Name, Model, Proxy URL, and API Key. You can create multiple configurations and switch between them at any time, which makes it easy to compare different models (query date 2026-09-09; the interface follows your current screen).
Can I test the Janitor AI API setup for free first?
Yes. Enter qwen/qwen3.7-flash:free or auto:free in the Model field. Both are zero-cost models. Confirm all four fields are correct first, then decide whether to top up for a stronger model.
Which model is the best value for roleplay?
Roleplay resends the earlier conversation with every turn, so input tokens accumulate quickly, and the input price should carry more weight. Most people start with deepseek-v4-flash ($0.20 input, $0.40 output per million tokens) to control costs. For the best long-form story consistency, but the most expensive option, use claude-sonnet-4.6 ($3.00/$15.00).
What should I do if I get a 401 or 404 after setup?
A 401 is almost always a mistyped key or extra spaces before or after it; paste it again. A 404 or 'model not found' means the Model field does not match the catalog string exactly; the model ID must match character for character, including suffixes like :free. If nothing responds at all, the Proxy URL is most likely wrong.
Can I get a unified invoice for Janitor AI spending for expense reporting?
Yes. Topping up through BazaarLink lets you use your company tax ID, receive a Taiwan electronic unified invoice, and pay in NT$, so you do not have to handle English receipts from overseas for company expense claims. Individual players who do not need an invoice can follow the exact same setup.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API