Your own AI with your own API key

With Premium, Dex can run on your own AI account. You pay the provider directly, and Dex uses no Draftdeck credits.

Updated Oct 10, 20264 min read

On this page

In short

Premium workspaces can run Dex on their own OpenAI, Anthropic, Google Gemini or Mistral account. Dex then uses no Draftdeck credits: you pay the provider directly, at its prices.
Create an API key at the provider (see below), and add credit or a payment method there.
In Draftdeck open Settings › Dex and AI › Your own AI. Pick the provider, paste the key, keep the suggested model and click Test and save.
What Dex sends then goes to your provider under your own contract with it. Draftdeck stores the key encrypted and shows only its last 4 characters.
When Premium ends, the key stays saved but unused, and Dex runs on Draftdeck's AI with credits again. You can remove it at any time.

Connect your own AI in one minute

The provider bills you by use (tokens). A chat answer with the suggested models usually costs a fraction of a cent; see the provider's price list. A web search is billed by the provider on top.
With your own key the provider becomes your workspace's own processor: what Dex sends (see the privacy notice) goes to it under your contract and its terms. Draftdeck is not responsible for the processing there. Check its data processing terms (DPA) and whether it trains on API data.
EU or US: OpenAI, Anthropic and Google process in the USA (with EU standard contractual clauses). Mistral is a French company; its EU endpoint keeps requests in the EU but has no web search, the global endpoint has it.
The key belongs to the workspace, not to a person: it is not part of anybody's data export, and it is deleted with the workspace.

Sign in to the OpenAI platform (platform.openai.com). This is not ChatGPT: a ChatGPT subscription does not include API use.
Billing: under Settings › Billing, add a payment method and buy some credit.
Key: Settings › API keys › Create new secret key. Copy it right away, OpenAI shows it only once. It starts with sk-.
Model: gpt-4o-mini (suggested, fast and cheap); gpt-4.1 or gpt-4o write better and cost more. Web search: yes.

Sign in to the Anthropic Console (console.anthropic.com). A Claude.ai subscription does not include API use.
Billing: under Settings › Billing, buy credits.
Key: Settings › API keys › Create key. Copy it right away. It starts with sk-ant-.
Model: the suggested Claude Haiku model is fast and cheap, a Sonnet model writes better. Web search: yes; on some accounts an admin has to allow it in the Console first.

Open Google AI Studio (aistudio.google.com) and sign in with a Google account.
Key: Get API key › Create API key, in a new or existing Google Cloud project. It starts with AIza.
Billing: the free tier has tight limits, and on it Google may use what you send to improve its products. For work, set up billing for the project (in AI Studio or the Google Cloud console).
Model: gemini-2.0-flash (suggested) or gemini-2.5-flash; gemini-2.5-pro writes best and costs more. Web search: yes.

Sign in to La Plateforme (console.mistral.ai).
Billing: choose a plan under Billing and add a payment method. Free plans have tight limits and other data terms.
Key: API keys › Create new key. Copy it right away.
Model: mistral-small-latest (suggested); mistral-medium-latest or mistral-large-latest write better. In Draftdeck choose the endpoint: EU (data stays in the EU, no web search) or Global (with web search).

Open Settings › Dex and AI › Your own AI, choose the provider, paste the key, keep or change the model and click Test and save. Draftdeck sends one small test request and saves the key only if it works.
Connected shows the provider, the model, the last 4 characters of the key and whether web search works. At the top, Own AI replaces the credits. Only owners and admins can connect a key.

Key refused (401): copy the whole key again, and check that it belongs to the chosen provider and was not deleted.
No credit or limit reached (429): add credit or a payment method at the provider, or wait until the limit resets.
Model not known (404): check the spelling or use the suggested model; some models need extra access at the provider.
No answer from the provider: check your connection and try again; if the provider has an outage, wait a few minutes.