Introduction

chat-rs is a runtime for model interaction. It gives you one consistent, type-safe API across every major provider, with streaming, multimodal content, tools, structured output, embeddings, and multi-provider routing.

Why chat-rs

• One API, every provider. 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 configuration like the model is enforced by the type system.
• Streaming first. Token-by-token output behind a single StreamEvent enum, plus full-duplex input streaming.
• Tools and human-in-the-loop. Native tool calling with pause and resume for approvals.
• Bring your own provider. Implement one trait and any backend plugs into the same reason-and-act loop.

How it fits together

chat-core defines the shared types and traits. Two wire crates implement the on-the-wire protocols, and provider crates are thin wrappers on top of them:

• chat-completions is the OpenAI Chat Completions wire.
• chat-responses is the OpenAI Responses API wire.
• chat-router composes several providers behind one, with fallback.

For example, chat-openai and chat-openrouter build on chat-responses; chat-ollama, chat-deepseek, chat-cerebras, and chat-huggingface build on chat-completions.