Model providers (your own AI key)
Kothadesk answers with your own AI provider key. Your provider bills you directly at their price, and Kothadesk never adds a margin to model usage.
Updated
On this page
Supported providers
#Every bot needs two kinds of model: a chat model to write answers and an embedding model to search your knowledge.
| Provider | Chat | Embeddings | Notes |
|---|---|---|---|
| OpenAI | Yes | Yes | One key covers both |
| Google Gemini | Yes | Yes | A Google AI Studio key; one key covers both |
| Anthropic | Yes | No | Add a second key (OpenAI, Gemini or compatible) for embeddings |
| xAI Grok | Yes | No | Add a second key for embeddings |
| OpenAI-compatible | Yes | Yes | OpenRouter, vLLM, Ollama or a gateway, on a public https address. Costs are not estimated |
Add a key
#- 01
Open Model providers
In the sidebar, choose Model providers, then Add provider key. You need to be an owner or admin with a verified email.
- 02
Fill in the key
Choose the provider, give the key a label (for example OpenAI production), paste the API key and tick what to use it for: Chat, Embeddings or both. OpenAI-compatible keys also need the base URL and an embedding model to test.
- 03
Verify and continue
Kothadesk makes one small test call. For embeddings it also checks that the model returns 1024 dimensions. A rejected key is not saved.
- 04
Choose default models
Pick the default chat and embedding models for new bots, or skip and choose per bot later.
Keys are stored encrypted and never shown again; only the last four characters appear in the list. If the provider is briefly unavailable, the key is saved as unverified: use Re-verify later.
| Message | What to do |
|---|---|
| The provider rejected the key (401/403) | Check the key is correct, active and belongs to the right project |
| This key's quota is used up | Wait for your provider's limit to reset or raise it with them |
| The provider is rate limiting this key | Try again in a few minutes |
| The embedding model must return exactly 1024 dimensions | Pick an embedding model that supports 1024 dimensions |
| The base URL is not allowed | Use a public https address |
Choose models for a bot
#In a bot's Settings, the Model card picks which key and model the bot uses.
| Preset | When to use it |
|---|---|
| Fast | Lowest cost; fine for simple FAQs |
| Balanced (recommended) | Good answers at a sensible cost for most support bots |
| Premium | Best quality for complex questions, at the highest cost |
| Specific or custom model | Any model your key can use, exactly as your provider names it |
Presets follow the recommended model when it changes, so you get improvements without editing anything. The estimated price per million tokens is shown under the model.
Watch out: Changing the embedding key or model re-indexes all of the bot's knowledge, with one embedding call per chunk billed by your provider. The old index keeps answering until the new one is ready.
Describe images with AI (Bot settings, Images) runs on the bot's own chat key and needs a vision-capable model: the Fast preset's model if it can read images, otherwise Balanced. Most current OpenAI, Anthropic and Gemini chat models can; models found only through an OpenAI-compatible endpoint or Ollama are treated as unable to. If neither preset can, the settings page warns and those images are marked Model can't read images.
Rotate or delete a key
#- Rotate key: paste the new key; it is verified first, then every bot switches to it at once.
- Choose models or Refresh models: update the defaults or the list of models your key can use.
- Delete: bots that use the key stop answering until you pick another key. Point them at another key first; forcing the delete asks you to type the key's label.
- If your provider rejects a key during normal use, it is marked invalid and owners get an email.
Costs
#Settings, Usage shows estimated cost, tokens and model calls per day, per bot, per type and per provider and model, with a CSV export. The figures are estimates: your provider's bill is the final word.
- Set a monthly spending limit with your provider as a safety net.
- Use a separate key for Kothadesk so its usage is easy to see in your provider's dashboard.
- Start with Balanced and use an evaluation set before switching to a cheaper model.
- Image descriptions (when Describe images with AI is on) appear as their own type. Each is one small call per image without alt text or a caption, capped at 2000 per bot a month.
- AI assist for agents (suggested replies, Ask AI and translations in the Inbox) appears as its own type, on the bot's chat key.
Worked examples
#Online shop
Keeping costs low
Answer a high volume of simple delivery and returns questions cheaply.
One Gemini key for chat and embeddings, chat preset Fast. Run an evaluation after a week of real questions to confirm the answers hold up. Gemini models read images, so the same key can describe product photos that have no alt text.
Clinic or bookings
Careful, polite answers
Clear, well-worded replies to patients.
An Anthropic key for chat (preset Balanced) and an OpenAI key for embeddings, both with a monthly spending limit.
Software company
Using your own gateway
Route all AI traffic through the company's existing gateway.
Add an OpenAI-compatible key with the gateway's public https base URL and pick the exact model names. Costs are not estimated, so track them in the gateway.
Coaching institute
A free-tier key while you try it
Try Kothadesk before paying for AI usage.
A Google AI Studio key works for chat and embeddings. If you see "quota is used up", the free daily limit was reached; it resets the next day.