Arc AI

A provider-agnostic AI SDK for Rust

中文  ·  English

Version	0.2.0
Stars
Forks
License	AGPL-3.0-or-later

---

Core Features

• Unified multi-provider API — one interface for DeepSeek / Qwen / OpenAI / Anthropic Claude / Google Gemini
• Streaming + tool calling together — text streams in real time while the tool loop runs automatically; no trade-off needed
• Native protocol engines — Anthropic and Gemini have dedicated engines, not OpenAI-compat wrappers

Provider	Engine	Streaming	Tool Calling
DeepSeek / Qwen / OpenAI	OpenAI-compat	✅	✅
Anthropic (Claude)	Native	✅	✅
Google Gemini	Native	✅	✅

Usage

[dependencies]
arc-ai = "0.2"

Chat

use arc_ai::{ArcClient, Message};

let client = ArcClient::deepseek("sk-...");  // switch provider here only
let msgs = vec![Message::user("Explain Rust ownership.")];
let resp = client.chat("deepseek-chat", &msgs).await?;
println!("{}", resp.text);

Stream

let mut stream = client.stream("gpt-4o", &msgs).await?;
while let Some(StreamPart::TextDelta(t)) = stream.next().await { print!("{t}"); }

Tool calling (streaming + auto loop)

let tool = Tool::simple("get_weather", "Get weather", &[("city", "City")], |args| async move {
    Ok(json!({ "temp": "32C", "condition": "sunny" }))
});
let req = Request::builder(client.provider().chat_model("claude-3-5-sonnet-20241022"))
    .messages(vec![Message::user("Weather in Beijing?")])
    .tools(vec![tool]).max_steps(5).build();
let mut s = sdk::stream_text(req).await?;  // text streams → tool runs → more text

cargo test      # 37 passed
cargo clippy    # zero warnings