Introduction

chat-rs lets you build LLM clients with ease and attach them to your tools. It provides one unified, type-safe API across many providers — Google Gemini, Anthropic Claude, OpenAI, OpenRouter, and the whole OpenAI-compatible ecosystem — with compile-time-safe builders, automatic retry loops, streaming, and native tool execution.

Why chat-rs

• One API, many providers. Swap providers without rewriting your call sites. Every provider implements the same CompletionProvider / StreamProvider traits.
• Type-state builders. Misconfiguration is a compile error, not a runtime surprise — required fields like the model are enforced by the type system.
• Streaming first. Token-by-token output (SSE and, where supported, WebSocket) behind a single StreamEvent enum.
• Tools & HITL. Native tool calling and human-in-the-loop pause/resume flows.
• Pluggable transport. Bring your own HTTP/WS transport, or use the built-in reqwest / tungstenite implementations.

Architecture at a glance

chat-core defines the shared types and traits. Wire crates implement the on-the-wire protocols:

• chat-completions — the OpenAI Chat Completions wire.
• chat-responses — the OpenAI Responses API wire.

Provider crates are thin wrappers that preset URLs, auth, and provider-specific niceties on top of a wire crate. For example, chat-openai and chat-openrouter wrap chat-responses; chat-deepseek and chat-cerebras wrap chat-completions.