Connect
Get an AI key
Relay uses an AI model to read emails, fill in skills and draft replies. You bring your own key from an AI provider, so you pay them directly at their prices and the account is yours. You can also run a model on your own hardware: see Use a local AI model.
How to add a key in Relay
- Create a key with your provider (see below). Copy it straight away: most providers show it only once.
- In Relay, open Settings > AI.
- Pick the provider under AI provider. Relay fills in the standard address.
- Type the Model name exactly as the provider lists it.
- Paste the key into AI provider API key and click Test connection.
Tip: Your key is stored only on your PC. Treat it like a password: anyone with it can run up charges on your account. Most providers let you set a monthly spending limit; it's worth doing.
Providers
These links were checked on 26 October 2026. Providers change their consoles often, so if a link has moved, start from the provider's home page and look for API keys.
| Provider | Where to create a key | Pricing |
|---|---|---|
| Claude (Anthropic) | console.anthropic.com | Anthropic pricing |
| OpenAI | platform.openai.com/api-keys | OpenAI API pricing |
| Google Gemini | Google AI Studio | Gemini API pricing |
| Mistral | console.mistral.ai | Mistral pricing |
| DeepSeek | platform.deepseek.com | DeepSeek pricing |
| xAI (Grok) | console.x.ai | xAI docs |
| Groq | console.groq.com/keys | Groq pricing |
| OpenRouter | openrouter.ai/keys | OpenRouter models |
| Xiaomi MiMo | Xiaomi's MiMo developer platform (Relay uses https://api.xiaomimimo.com/v1) |
See the provider's site |
| Other | Any OpenAI- or Anthropic-compatible endpoint: choose Other: OpenAI-compatible API (or Anthropic-compatible) and enter its address | Your provider's terms |
Which model should I pick?
Pick a current, mid-sized model from your provider: fast enough for email, good enough to follow your skills carefully. The biggest models write a little better and cost more; the smallest can miss details. You can switch at any time and compare drafts.
What does it cost?
Each email usually takes a few AI calls: one to read and sort it, one more to pick a skill when you have skills switched on, and one to draft the reply. Reports and skills with look-ups take a few more. Providers charge per token (roughly per word), so cost depends on the model and how long your emails and documents are.
To see your own numbers, enter the provider's Input price and Output price (per million tokens) in Settings > AI. Relay then tracks what it spends.
Does the provider see my data?
The provider sees the text Relay sends for each task, with private details replaced by placeholders when masking is on (the default). Each case shows exactly what was sent under What the AI saw. Most business API plans don't train on your data, but check your provider's terms. For the strongest guarantee, run a local model.
The decision engine
Relay also needs something to sort emails before drafting. By default it uses Jev when you have a Jev key, and otherwise your main AI, so one AI key is enough to start. See The decision engine for Jev, the free Kev model and how to keep everything on your PC.