Models and providers
Relay uses models in three roles, each set separately in Settings → AI:
| Role | Used for | Runs |
|---|---|---|
| Understand | Classifying every incoming event. | On every message that isn’t trivial. Keep it small and fast. |
| Answer | Replies, clarifying questions and plans. | Once per request that needs a reply. |
| Code | Making approved changes. | Once per approved plan. |
Providers that support it also take a reasoning effort per role.
Agent tools on your own plan
Section titled “Agent tools on your own plan”Relay can run an agent tool you’ve installed and signed in to. It never sees their credentials: it runs the tool, so usage counts against your own plan.
| Tool | Notes |
|---|---|
| Claude Code | Recommended. Also provides Slack, Linear and ClickUp connections. |
| Codex | Uses your ChatGPT plan. |
| Antigravity CLI | |
| OpenCode | |
| Cursor | Preview. |
| Grok Build | Preview. |
These tools read your repo with their own read-only tools when answering, and edit it when making a change.
API providers
Section titled “API providers”| Provider | Key |
|---|---|
| Anthropic | ANTHROPIC_API_KEY |
| OpenAI | OPENAI_API_KEY |
GOOGLE_GENERATIVE_AI_API_KEY |
|
| Mistral | MISTRAL_API_KEY |
| Groq | GROQ_API_KEY |
| xAI | XAI_API_KEY |
| DeepSeek | DEEPSEEK_API_KEY |
| OpenRouter | OPENROUTER_API_KEY |
| OpenAI-compatible (Ollama, LM Studio, vLLM…) | OPENAI_COMPATIBLE_API_KEY and a base URL |
The model picker lists models live from each provider’s own “list models” endpoint, so it’s never stale and only shows models your key can call.
API models get three small read-only tools, confined to the repo folder, for answering questions about code.
Token use
Section titled “Token use”Every model call’s tokens (and cost, where the provider reports it) are recorded. Settings shows today’s use and the last 14 days. See Guardrails for the daily budget.