Chub AI makes sense if you want to edit a character and choose the model behind the conversation. The catch is that the character, its model settings and its chat history are separate parts of the setup. If a familiar character changes after a move or model switch, I would check those parts before rewriting it or paying for more context.
This review focuses on preserving a useful character setup. Public interface observations are dated below; explanations of authenticated chat controls come from Chub’s official documentation.
Choose the model connection before the subscription
Chub is a character library and chat frontend with several inference routes. Its API guide lists Chub’s own Mars service alongside external providers and self-hosted Kobold/Ooba connections. A free frontend does not make every route free.
In an anonymous desktop Chrome check on November 2, 2026, the public subscription page displayed one $20/month offer and explained its crypto-only payment approach. We did not enter checkout; the displayed price is not a verified tax, renewal or payment quote.
Its free-option FAQ described one-time trial credits of a few dozen messages being distributed to new users periodically. That does not establish a daily refill or an immediately available balance for every signup.
The same FAQ describes single-user access and says concurrent requests from an account return HTTP 429. That is Chub's stated condition, not an error we reproduced. If you connect an API client, queue requests before treating a parallel-generation failure as a broken character.
For an external API, compare that provider’s usage charges with your actual conversation length. A Chub subscription does not mean an unrelated provider’s API bill is included. Self-hosting also brings hardware requirements; renting a GPU still costs money. These are billing distinctions, not a quality comparison between models.
Keep the character, transcript and setup notes separate
For a move to another frontend, I would check these three pieces separately:
- Character definition
- Check the description, greeting and any embedded character book in the destination. A familiar avatar is not evidence that all instructions survived the move.
- Conversation
- Use a chat export for the session you want to continue. Chub's guide lists JSONL for SillyTavern, PNG and Text under Chat Settings. Choose the format the destination accepts.
- Model setup
- Record the API, model, preset and important settings separately. Keep secret keys out of shared notes. Reusing a character does not establish that its old model configuration is active.
The chat guide describes presets as model-specific. It also distinguishes edits to a character inside a particular chat from updating the underlying character. That matters when a saved session works but a new session behaves differently.
Do not delete the original conversation after seeing an import succeed. I have not tested an export round trip or preservation of every branch, Stage or attachment. Before committing to a move, reopen a harmless exported session in the destination and check its ending and one earlier turning point. That is a suggested check, not a migration result.
Check these controls before rewriting the character
The following troubleshooting order is based on official documentation. I did not run generation tests or observe these settings in an authenticated chat.
The character ignores its own instructions
Chub's character guide says character-specific System Prompt and Post History Instructions require Use V2 Spec. Check that setting before assuming the card is broken. This condition does not mean every ordinary description or greeting requires V2.
Also distinguish Creator's Note from Character's Note. The first is metadata; the latter is inserted into the prompt. Putting a behavior rule only in Creator's Note will not give it the documented prompting role.
The character changes after an API or model switch
Check the active model and use a suitable model-specific preset before editing the character. Keep the same short, neutral scene for the comparison. If you change the card, preset and model together, you cannot tell which change affected the reply.
A lorebook fact stops appearing
Check the keyword, scan depth and token budget. Chub documents whole-word, case-insensitive matching by default, so a keyword inside a longer word need not activate the entry. A matching entry can also be excluded when the lorebook budget fills.
Character books require Use V2 Spec. Chat-level scan depth and budget settings can override the book's defaults. Check that configuration before buying a larger-context model.
Sources: Character Creation, Just Chatting and Lorebooks. These mechanisms explain what to inspect; they do not establish that a particular model will follow every instruction.
A visible transcript can exceed the model’s context
Chub’s prompting guide says the model receives the history that fits its context, rather than necessarily the entire conversation. Being able to scroll to an old message does not establish that the next request includes it.
My suggested diagnostic is to note one harmless fact, continue the conversation, then ask about it without supplying the answer. Record the model, preset and context settings. A correct answer could still come from recent repetition or another prompt field; it does not by itself prove durable memory.
If the workflow you want is a locally managed frontend, compare SillyTavern or Agnai with the backend you intend to use. Changing the interface alone does not establish where inference runs.
Next step: verify one setup before moving your main chat
Start with one non-sensitive character and a short conversation. Confirm the active model and preset, inspect any V2-dependent fields, then try the destination’s supported chat format. Keep the original until that check passes. If you want a companion service with almost no configuration, Chub’s extra control may be more work than you need.