Docs · 03
Powering your twin
The twin thinks with an AI model. You bring your own provider key from Anthropic, OpenAI, Azure, Gemini, Grok, or a custom OpenAI-compatible endpoint. This page covers the recommended models, setup, and a rough sense of cost.
01Bring your own key
You power the twin with your own AI provider key. Create (or use an existing) account with an AI provider, then paste a key into Twinly's Settings. You pay the provider directly for what the twin uses, and the key lives in your macOS Keychain, never leaving your Mac. Most control, most transparency.
Everything below covers how to set that up, the models we recommend, and a rough sense of cost.
02Bring your own key
Open Settings in the app and choose a provider: Anthropic, OpenAI, Azure OpenAI, or a custom endpoint (anything that speaks the OpenAI chat-completions API, including local servers). Paste your key. Twinly stores it in the macOS Keychain and never writes it anywhere else; it is only ever sent to that provider.
- Get a key from your provider's dashboard (Anthropic Console, OpenAI Platform, or your Azure resource).
- In Twinly, open Settings and select the provider.
- Paste the key and save. The provider chip at the bottom of the app flips to connected.
03Azure OpenAI
Azure needs three values from your Azure resource page:
- Endpoint — your resource URL, like
https://your-resource.openai.azure.com/. - Deployment — the deployment name you created in Azure AI Foundry. This is the model field for Azure.
- Key — from the same resource page.
04The brain / vision / call model fields
The twin does three different jobs, and each wants a different kind of model. Twinly lets you set them independently in Settings, so you spend on horsepower only where it helps:
| Field | What it does | Recommended |
|---|---|---|
| Brain | Plans tasks, writes drafts, runs the agent loop. Wants steady reasoning. | gpt-5.4-mini and up (OpenAI / Azure), or Claude Opus 4.6 and up |
| Vision | Reads screenshots while the twin works your Mac. Wants sharp eyes. | gpt-5.4-mini and up, or Claude Opus 4.6 and up |
| Call | Holds live phone conversations. Wants low latency. | gpt-5.4-mini (OpenAI), or Claude Opus 4.6 and up |
Leave the vision or call field blank and the brain model covers that job too. On the plain OpenAI provider the single Model field defaults to gpt-5.4. The custom-endpoint provider takes a full chat-completions URL plus the same three model fields.
05Roughly what it costs (bring-your-own-key)
On your own key you pay your provider for usage. The bill is dominated by two things: the screenshots the twin reads while working your Mac, and long multi-step reasoning runs. The defaults are deliberately tuned to the cheapest models that hold up in practice, and Twinly's execution engine (see How Twinly executes) avoids calling the model at all for tasks it can do deterministically.
| How you use it | Rough monthly model spend |
|---|---|
| Light: a few texts, calendar, email each day on the recommended models | a few dollars |
| Regular: daily mix of computer-use tasks, drafts, and the odd call | low tens of dollars |
| Heavy: lots of vision-driven computer use or flagship models | tens of dollars and up, scales with the model you pick |
06Switching and fallback
You can switch providers any time; tasks pick up the new brain on their next turn. For live phone calls there is also a fallback chain: if your selected provider has a hiccup mid-call, the twin quietly falls through any other configured providers rather than dropping the conversation.