---
title: "Chub AI API key: reverse proxy URL, key and model id"
canonical: https://unrestrictable.ai/guides/chub-ai
generated: 2026-11-16T02:57:40.034Z
source: https://unrestrictable.ai/guides/chub-ai
base_url: https://api.unrestrictable.ai/v1
---

# Chub AI API key: connect an uncensored model by reverse proxy

Run Chub AI on your own prepaid key: where the reverse proxy URL and key go in Secrets, which model id to type, the sampler values we open with, and the price of 100 messages.

Steps checked November 6, 2026 · prices as of November 16, 2026

## What you need

- A Chub AI account with a character open in a chat. Both panels you will touch, Secrets and Configuration, sit in the options menu; Chub's docs put that menu behind the hamburger button, top right on the chat page.
- A key from the [console](https://app.unrestrictable.ai/console). It begins with `sk-unr-` and is displayed a single time, so copy it before the dialog closes. Top-ups start at $10, no plan is required, and bigger top-ups carry bonus credit (+5% at $200, +10% at $500, +15% at $1,000, the console's own bonusTiers).
- A model id from the [Models page](https://unrestrictable.ai/models). The string printed there is the string Chub's model field wants, character for character. Four ids are priced further down.
- Nothing else. If you want to hear the model before any setup, [/chat](https://unrestrictable.ai/chat) runs the same catalog in the browser with three free turns a day and no account, and [/telegram](https://unrestrictable.ai/telegram) does the same through a bot.

## The three values

When you finish, each message in a Chub chat will go to an uncensored model billed to your own prepaid key rather than to one of Chub's bundled APIs. Three values make that connection. The URL and the key go in Secrets. The model id goes in a field which, in the walkthroughs ranking for this query, sits under Configuration next to the generation settings.

- Secrets, API: OpenAI, reverse proxy
- Proxy URL: https://api.unrestrictable.ai/v1 Ends in /v1; switch to https://api.unrestrictable.ai/v1/chat/completions when the check returns 404 route_not_found
- Proxy key: sk-unr-live-… Made in the console, displayed one time
- Model field: uncensored-ai-roleplay Any Models page id, character for character

Nothing leaves Chub: characters, personas, lorebooks and chat history stay put. Two things change, which model writes each reply and who sends the bill. The bill becomes a balance that settles [per token](https://unrestrictable.ai/pricing) rather than a Chub tier. The [docs page](https://unrestrictable.ai/docs) gives the same three values for every client; the guides for [SillyTavern](https://relayswitch.io/guides/sillytavern), [Janitor AI](https://relayswitch.io/guides/janitorai), [RisuAI](https://unrestrictable.ai/guides/risuai-custom-api) and [Agnai](https://unrestrictable.ai/guides/agnai-proxy) reuse them, and the [best uncensored LLM leaderboard](https://unrestrictable.ai/blog/uncensored-llm-leaderboard) helps choose the id.

## Step one, Secrets: URL and key

Chub's [API Connections page](https://docs.chub.ai/docs/the-basics/api-connections) says OpenAI can run on an API key or through a reverse proxy. In Chub's vocabulary a reverse proxy is any address that speaks the chat completions format, which this API does. Chub prints no label for its URL field or its key field, so every step below names the field by the thing it asks for; if your build words a label differently, go by that.

1. With a character open, open the options menu (hamburger button, upper right of the chat page, as Chub's [chat page docs](https://docs.chub.ai/docs/the-basics/just-chatting) describe). Select **Secrets**, which the docs call the place for setting up the API connection a chat uses.
2. Choose the **OpenAI** API, then move from the own-key mode to the **reverse proxy** mode, the one with an address field.
3. Paste `https://api.unrestrictable.ai/v1` into the proxy URL field, ending at `/v1`. That is correct when the app adds the chat completions path by itself. Should the next step's check come back 404 `route_not_found`, your build posts to the address verbatim: change the value to `https://api.unrestrictable.ai/v1/chat/completions`. The API answers both; the one matching what the app appends is the one that works.
4. Paste the key into the proxy key field, the full string from `sk-unr-` onward, without `Bearer` before it.
5. Save before leaving the panel. If your build offers a check button for the proxy (the ranking walkthroughs mention one, Chub's docs do not), use it: it fires a short chat completion at the URL using the key; passing means URL and key are both right. With no button, the first chat message is the check. A failure either way comes with a status code, and the Troubleshooting table below maps that code to a field.

Chub's docs steer people toward official keys and away from any third-party reverse proxy, since whoever runs the proxy can see your data and IP address. That holds for any address pasted here, this one included. The retention section lower down lists what this operator stores, so you can weigh it against a free proxy that discloses nothing.

## Step two, Configuration: model id and prompts

Chub's chat page docs place prompt structure and generation parameters inside **Configuration**, found in that same options menu, and **Presets** store them. The docs leave the model field unplaced; the walkthroughs ranking for this query show it inside the prompt structure part of Configuration, beside the API choice. Three settings go there.

1. First, where the panel asks what API the prompt is for, pick **OpenAI**, so the chat routes through the reverse proxy saved in Secrets.
2. Type a model id into the model field spelled the way the Models page spells it, `uncensored-ai-roleplay` say. The API matches the id and nothing else: a display name like "GLM-5.3 Flash Uncensored", or an id from another vendor's list like `gpt-4`, returns 404 `model_not_found`. If your build presents models as a fixed list with no free-text field, the check returns that code; send support the request id and we will say how the current build behaves.
3. Of the API Settings Chub lists for every API, keep the **System Prompt** (which the docs also name the Jailbreak Prompt) clear of jailbreak text, and keep **Post History Instructions** blank unless some character relies on a formatting rule. The models here stay in character without a prefill, and a pasted jailbreak is resent with every message at token prices.

Store the preset under a name you will know later, one preset per model if you switch between them. The URL and key live in Secrets and do not change; a single key reaches every id on the Models page.

## Sampler values for an uncensored model

Chub's docs list five generation settings (Temperature, Top P, Top K, Frequency Penalty and Presence Penalty) and give one rule: move Temperature or Top P, never both. An uncensored or abliterated model has no refusal to steer past, so these values shape the writing and nothing more. Our opening values for a Chub preset, with Temperature as the one dial we touch:

| Setting | Opening value | Reason |
| --- | --- | --- |
| Temperature | 0.9 | At 1.0 you get the model's raw distribution. 0.9 keeps the range and cuts the occasional nonsense; go down to 0.7 for a model that wanders. |
| Top P | 1.0 | Left off, following Chub's one-or-the-other rule. Adjust Temperature instead. |
| Top K | 0 | 0 disables it. Cutting the token set on top of Temperature makes the prose flat. |
| Frequency Penalty | 0 | 0.1 or 0.2 once a model starts recycling pet phrases late in a chat, and no more: past that it starts avoiding its own characters' names. |
| Presence Penalty | 0 | The same reasoning. 0.2 is the ceiling, for a model that keeps re-describing the scene. |

Those five are where Chub's docs end. Two further values shape each request Chub sends, and the ranking walkthroughs expose both as settings; set each under whichever name your build gives it. Response length, in tokens: 300 is roughly a paragraph, while a model that reasons before replying needs 1,000 or more, or the cap lands inside the reasoning and the reply comes back blank.

The context limit is the value that moves the bill. Every message carries the definition of the character, your persona, any lorebook entries and all the history the limit lets through. Derive it from the catalog's `context_window` for the model: 16K below the window, capped at 64K, and never above the window, because an oversized prompt is refused rather than trimmed. Chub's **Chat Memory** (a summary passed as context, per the docs) keeps a long story alive inside a short window.

| Model | Window in the catalog | Limit we type |
| --- | --- | --- |
| [Uncensored AI Roleplay](https://unrestrictable.ai/models/uncensored-ai-roleplay) | 128,000 tokens | 65536 |
| [GLM-5.3 Flash Uncensored](https://unrestrictable.ai/models/glm-5.3-flash-uncensored) | not published | 32768 |
| [GLM-5.3 Uncensored](https://unrestrictable.ai/models/glm-5.3-uncensored) | not published | 32768 |
| [DeepSeek V4 Pro Uncensored](https://unrestrictable.ai/models/deepseek-v4-pro-uncensored) | not published | 32768 |

Windows are read live from the catalog on 2026-11-16. A "not published" entry means the developer gives no figure; start at 32K in that case.

## Test it from a terminal

Getting a reply inside the chat proves the UI side. The request below proves the API side with the shape Chub uses: a system message holding the character, a single user turn, and the settings from above. Swap in your key and run it.

```bash
curl -i https://api.unrestrictable.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-unr-live-…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "uncensored-ai-roleplay",
    "max_tokens": 300,
    "temperature": 0.9,
    "messages": [
      {"role": "system", "content": "You are Tomas, the keeper of a mountain inn snowed in for the week. Stay in character. Reply in two short sentences."},
      {"role": "user", "content": "The road is closed again. Is there a room left?"}
    ]
  }'
```

A 200 arrives in the [chat completions format](https://platform.openai.com/docs/api-reference/chat). The text sits in the first choice. Billing follows the `usage` block, and Chub's definition, persona and history count toward `prompt_tokens` in just the same way.

```
HTTP/2 200
content-type: application/json; charset=utf-8
x-uplink-request-id: ul_…
x-uplink-model: uncensored-ai-roleplay
x-uplink-charge: 0.00…

{
  "id": "chatcmpl-…",
  "object": "chat.completion",
  "model": "uncensored-ai-roleplay",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "…" },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": …, "completion_tokens": …, "total_tokens": … }
}
```

Two headers come with each answer: a request id to quote to support, and the charge that call settled for. When `finish_reason` reads `length`, the response length cap truncated the reply; on a reasoning model that is the blank-reply case from the troubleshooting table. Nothing is charged when a request fails before the model replies.

## Which id to type

The pages ranking for "chub ai api key" combine a stock vendor model and a jailbreak that is resent in the System Prompt every turn. An uncensored model needs no prefill; the model field alone does the work. Four ids to start with, priced live from the catalog:

| Model | Build | Input $/M | Output $/M |
| --- | --- | --- | --- |
| [Uncensored AI Roleplay](https://unrestrictable.ai/models/uncensored-ai-roleplay) uncensored-ai-roleplay | Roleplay | $0.60 | $2.34 |
| [GLM-5.3 Flash Uncensored](https://unrestrictable.ai/models/glm-5.3-flash-uncensored) glm-5.3-flash-uncensored | Uncensored | $0.1053 | $0.3685 |
| [GLM-5.3 Uncensored](https://unrestrictable.ai/models/glm-5.3-uncensored) glm-5.3-uncensored | Uncensored | $0.282 | $0.846 |
| [DeepSeek V4 Pro Uncensored](https://unrestrictable.ai/models/deepseek-v4-pro-uncensored) deepseek-v4-pro-uncensored | Uncensored | $0.97 | $1.94 |

Prices per million tokens, read from the live catalog on 2026-11-16.

- [Uncensored AI Roleplay](https://unrestrictable.ai/models/uncensored-ai-roleplay), `uncensored-ai-roleplay`: trained for roleplay and our first pick on a new character. No reasoning step, so it replies at once, with 128,000 tokens of context.
- [GLM-5.3 Flash Uncensored](https://unrestrictable.ai/models/glm-5.3-flash-uncensored), id `glm-5.3-flash-uncensored`: the budget pick for long sessions, when the per-message figure counts for more than the prose does.
- [GLM-5.3 Uncensored](https://unrestrictable.ai/models/glm-5.3-uncensored), id `glm-5.3-uncensored`: the full GLM for long scenes, once the flash model's replies start to feel thin.
- [DeepSeek V4 Pro Uncensored](https://unrestrictable.ai/models/deepseek-v4-pro-uncensored), id `deepseek-v4-pro-uncensored`: the largest of the four, for plots that call for it. Set its response length to 1,000 or more.

The opposite direction is a 24B roleplay fine-tune: `cydonia-24b-v4.3` was trained on long character chat, publishes its context window, and today lists at $0.60 per million output tokens. Every model is on the [Models page](https://unrestrictable.ai/models) with roleplay and abliterated builds marked; the [uncensored hub](https://unrestrictable.ai/uncensored) covers the difference. The same models answer in the browser at [/uncensored-ai-roleplay](https://unrestrictable.ai/uncensored-ai-roleplay) if you want to compare before pointing Chub at one.

Every listed id gets the same fixed prompt once a day, and the reply is matched against what that model is known to say. On 2026-11-16 the check ran on 169 of the models on this site; 147 replied as the named model. That included all four above. How the check works, and today's results, are at [/proof](https://unrestrictable.ai/proof).

## The price of 100 messages

What you type is a small part of a Chub message. Each turn also sends the character definition, your persona, any matching lorebook entries, and all the history that fits the context limit. We price one message at 4,000 tokens in and 300 out, then the same message deep into a long chat at 16,000 tokens in. Per-million prices are read from the live catalog on 2026-11-16; the math is tokens multiplied by price and nothing more.

| Model | Per message | Per 100 | Per 100 at 16k context |
| --- | --- | --- | --- |
| `uncensored-ai-roleplay` | $0.0031 | $0.3102 | $1.0302 |
| `glm-5.3-flash-uncensored` | $0.0005 | $0.0532 | $0.1795 |
| `glm-5.3-uncensored` | $0.0014 | $0.1382 | $0.4766 |
| `deepseek-v4-pro-uncensored` | $0.0045 | $0.4462 | $1.6102 |

In other words, the smallest top-up, $10, covers roughly 3,223 messages that size on `uncensored-ai-roleplay` at the current price. Credit stays good for 12 months, so a quiet month costs nothing.

Two levers move the figure. Context limit outweighs reply length: the 16,000-token column shows the same reply carrying more history. And a System Prompt jailbreak is billed on every single message, the second reason to leave it blank on an uncensored model. The [pricing page](https://unrestrictable.ai/pricing) lists each model's per-token price and the bonus bonusTiers.

## What this operator stores

Chub's warning is right: every message travels to the address in Secrets, and whoever operates that address can read it. The real question is who that is and what they retain. Per the [privacy page](https://unrestrictable.ai/privacy): a usage record for each request (request id, model, token counts, charge); SHA-256 fingerprints of the request and the response, never the text itself, so a signed receipt can be verified against the bill; and request bodies only while replay is switched on for a key, kept 7 days. Replay off means prompts and replies pass through unwritten. The key is held as a salted hash and displayed a single time.

Two free habits. Put a monthly spend limit on the key from the console; a key that escapes in a shared screenshot then stops at 402 `monthly_limit_reached` instead of emptying the balance. And use one key per app, one for Chub and another for SillyTavern, so revoking one leaves the other alone.

## Troubleshooting

A failed generation carries one of the API's status codes, and each code points at a field.

| Code | Body | What happened | What to do |
| --- | --- | --- | --- |
| 401 | `invalid_api_key` | Chub's proxy key field has a typo or a trailing space. | Paste the key again from the console into Secrets, then save. A revoked key gives the same answer. |
| 404 | `route_not_found` | The address Chub sends to is not the chat completions route. | Swap the proxy URL between `https://api.unrestrictable.ai/v1` and `https://api.unrestrictable.ai/v1/chat/completions`; one matches what your build appends. |
| 404 | `model_not_found` | The model field contains a display name, another vendor's id or a misspelling. | Copy the id off the Models page as printed. The body echoes the id it got. |
| 402 | `insufficient_balance` | No prepaid balance left. | Add funds in the console. `monthly_limit_reached` under the same code means the key's own monthly cap was reached; raise it or wait for the new month. |
| 403 | `scope_denied` | The key is restricted to certain models, and the model field names a different one. | Allow the model on the key in the console, or pick an allowed id. |
| 429 | `rate_limited` or `upstream_rate_limited` | Too many requests on this key, or a saturated route for the model. | Regenerate once `retry-after` has passed, or use another id for a minute. |
| 502 | `upstream_error` | The route to the model failed, or the reply was cut before it reached you. | Wait out `retry-after` and regenerate. No charge was made. |
| 200 | a blank reply | The response length cap ran out inside a reasoning step, so `content` is empty and `finish_reason` says `length`. | Set the response length to 1,000 or above, or use a model that replies without reasoning, `uncensored-ai-roleplay` for one. |

Below are the API's own bodies for the common codes. The final token on each line is the string to look for in Chub's error text.

```
HTTP 401, the key is wrong
{"error":{"message":"The API key is not valid.","type":"authentication_error","param":null,"code":"invalid_api_key"}}

HTTP 404, the URL does not reach the chat completions route
{"error":{"message":"No such route.","type":"not_found_error","param":null,"code":"route_not_found"}}

HTTP 404, the model field names an id that is not on the list
{"error":{"message":"No such model: gpt-4.","type":"not_found_error","param":null,"code":"model_not_found"}}

HTTP 402, the balance is empty
{"error":{"message":"The prepaid balance does not cover this request. Top up at … to continue.","type":"billing_error","param":null,"code":"insufficient_balance"}}
```

A 200 whose reply ignores the character is a settings issue, not a connection issue: a System Prompt that contradicts the definition, or a Temperature far off 1, looks like that. Loading the same id in [SillyTavern](https://relayswitch.io/guides/sillytavern) with that guide's samplers is the fast comparison.

## Questions

### Where does the API key go in Chub AI?

Into Secrets, the panel Chub’s docs name as the place to set up the API connection for chatting. From an open chat, open the options menu at the top right, pick Secrets, choose OpenAI, switch to reverse proxy, then paste the URL and the key. The model id has a field of its own, which the ranking walkthroughs put under Configuration.

### How can I get a key for Chub AI?

From whichever provider runs the model you want. For this one: make a key in the console, top up from $10, and paste the sk-unr- string into Chub’s proxy key field. It is displayed once and works in Chub, SillyTavern and any SDK at the same time. See the console.

### Does a free Chub AI API exist?

Chub’s docs point out that a reverse proxy operator sees your data and IP address, which is the price of a free proxy. A prepaid key bills per token with no subscription, and the cost table above prices 100 messages on four uncensored models in cents. For a free taste outside Chub, /chat gives three turns a day with no account.

### What is the best API to use with Chub AI?

One that names its model, prints its price and says what it retains. Here the model field accepts any id on the Models page; uncensored-ai-roleplay is the roleplay default and glm-5.3-flash-uncensored the budget pick. See the Models page.

### Why is the API reply empty in Chub AI?

Most often the response length cap was used up inside a model’s reasoning step, so content arrives empty with finish_reason length. Set the cap to 1,000 or more, or pick a model that does not reason first. The other cause, a wrong model id, shows as 404 model_not_found rather than as an empty reply.

### What is behind "Error during generation" in Chub AI?

The request to the proxy URL returned no reply. The API’s status code says why: 401 is the key, 404 the URL or the model id, 402 the balance, 429 a rate limit. The troubleshooting table above ties each code to the field to fix.

Every model above is on the [uncensored model list](https://unrestrictable.ai/models) with its live price per million tokens. [Per-token pricing](https://unrestrictable.ai/pricing) sets the whole catalog side by side, and the [family directory](https://unrestrictable.ai/uncensored) groups the builds by base model.

[Get a key](https://app.unrestrictable.ai/console) Prepaid from $10, no subscription.

## New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.
