ModelRelay Rust SDK

The ModelRelay Rust SDK is a responses-first, streaming-first client for building cross-provider LLM features without committing to any single vendor API.

It’s designed to feel great in Rust:

• One fluent builder (ResponseBuilder) for sync/async, streaming/non-streaming, text/structured, and customer-attributed requests.
• Structured outputs powered by real Rust types (schemars::JsonSchema + serde::Deserialize) with schema generation, validation, and retry.
• A practical tool-use toolkit (registry, typed arg parsing, retry loops, streaming tool deltas) for “LLM + tools” apps.

[dependencies]
modelrelay = "0.44.1"

Quick Start (Async)

use modelrelay::{Client, ResponseBuilder};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let client = Client::with_key(std::env::var("MODELRELAY_API_KEY")?).build()?;

    let response = ResponseBuilder::new()
        .model("claude-sonnet-4-20250514")
        .system("Answer concisely.")
        .user("Write one line about Rust.")
        .send(&client.responses())
        .await?;

    // The response is structured: output items, tool calls, citations, usage, etc.
    println!("tokens: {}", response.usage.total());
    Ok(())
}

Why This SDK Feels Good

Fluent request building (value-style)

ResponseBuilder is a small, clonable value. You can compose “base requests” and reuse them:

use modelrelay::ResponseBuilder;

let base = ResponseBuilder::new()
    .model("gpt-4.1")
    .system("You are a careful reviewer.");

let a = base.clone().user("Summarize this changelog…");
let b = base.clone().user("Extract 3 risks…");

Streaming you can actually use

If you only want text, stream just deltas:

use futures_util::StreamExt;
use modelrelay::ResponseBuilder;

let mut deltas = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Write a haiku about type systems.")
    .stream_deltas(&client.responses())
    .await?;

while let Some(delta) = deltas.next().await {
    print!("{}", delta?);
}

If you want full control, stream typed events (message start/delta/stop, tool deltas, ping/custom):

use futures_util::StreamExt;
use modelrelay::{ResponseBuilder, StreamEventKind};

let mut stream = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Think step by step, but only output the final answer.")
    .stream(&client.responses())
    .await?;

while let Some(evt) = stream.next().await {
    let evt = evt?;
    if evt.kind == StreamEventKind::MessageDelta {
        if let Some(text) = evt.text_delta {
            print!("{}", text);
        }
    }
}

Structured outputs from Rust types (with retry)

Structured outputs are the “Rust-native” path: you describe a type, and you get a typed value back.

use modelrelay::{Client, ResponseBuilder};
use schemars::JsonSchema;
use serde::Deserialize;

#[derive(Debug, Deserialize, JsonSchema)]
struct Person {
    name: String,
    age: u32,
    email: Option<String>,
}

let client = Client::with_key(std::env::var("MODELRELAY_API_KEY")?).build()?;

let result = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Extract: John Doe is 30 years old, john@example.com")
    .structured::<Person>()
    .max_retries(2)
    .send(&client.responses())
    .await?;

println!("{:?}", result.value);

And you can stream typed JSON with field-level completion for progressive UIs:

use futures_util::StreamExt;
use schemars::JsonSchema;
use serde::Deserialize;
use modelrelay::ResponseBuilder;

#[derive(Debug, Deserialize, JsonSchema)]
struct Article {
    title: String,
    summary: String,
    body: String,
}

let mut stream = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Write an article about Rust's ownership model.")
    .structured::<Article>()
    .stream(&client.responses())
    .await?;

while let Some(evt) = stream.next().await {
    let evt = evt?;
    for field in &evt.complete_fields {
        if field == "title" {
            println!("Title: {}", evt.payload.title);
        }
    }
}

Tool use is end-to-end (not just a schema)

The SDK ships the pieces you need to build a complete tool loop:

• create tool schemas from types
• parse/validate tool args into typed structs
• execute tool calls via a registry
• feed results back as tool result messages
• retry tool calls when args are malformed (with model-facing error formatting)

use modelrelay::{
    function_tool_from_type, parse_tool_args, respond_to_tool_call_json, ResponseBuilder, Tool,
    ToolChoice, ToolRegistry, ResponseExt,
};
use schemars::JsonSchema;
use serde::Deserialize;

#[derive(Debug, Deserialize, JsonSchema)]
struct WeatherArgs {
    location: String,
}

let registry = ToolRegistry::new().register(
    "get_weather",
    modelrelay::sync_handler(|_args_json, call| {
        let args: WeatherArgs = parse_tool_args(call)?;
        Ok(serde_json::json!({ "location": args.location, "temp_f": 72 }))
    }),
);

let schema = function_tool_from_type::<WeatherArgs>()?;
let tool = Tool::function(
    "get_weather",
    Some("Get current weather for a location".into()),
    Some(schema.parameters),
);

let response = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Use the tool to get the weather in San Francisco.")
    .tools(vec![tool])
    .tool_choice(ToolChoice::auto())
    .send(&client.responses())
    .await?;

if response.has_tool_calls() {
    let call = response.first_tool_call().unwrap();
    let result = registry.execute(call).await;
    let tool_result = respond_to_tool_call_json(call, &result.result)?;

    // Feed the tool result back as an input item and continue the conversation.
    let followup = ResponseBuilder::new()
        .model("claude-sonnet-4-20250514")
        .user("Great—now summarize it in one sentence.")
        .item(tool_result)
        .send(&client.responses())
        .await?;

    println!("followup tokens: {}", followup.usage.total());
}

Customer-Attributed Requests

For metered billing, set customer_id(...). The customer’s tier can determine the model (so model(...) can be omitted):

use modelrelay::ResponseBuilder;

let response = ResponseBuilder::new()
    .customer_id("customer-123")
    .user("Hello!")
    .send(&client.responses())
    .await?;

Blocking API (No Tokio)

Enable the blocking feature and use the same builder ergonomics:

use modelrelay::{BlockingClient, BlockingConfig, ResponseBuilder};

let client = BlockingClient::new(BlockingConfig {
    api_key: Some(std::env::var("MODELRELAY_API_KEY")?),
    ..Default::default()
})?;

let response = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Hello!")
    .send_blocking(&client.responses())?;

Feature Flags

Errors

Errors are typed so callers can branch cleanly:

use modelrelay::{Error, ResponseBuilder};

let result = ResponseBuilder::new()
    .model("claude-sonnet-4-20250514")
    .user("Hello!")
    .send(&client.responses())
    .await;

match result {
    Ok(_response) => {}
    Err(Error::Api(e)) if e.is_rate_limit() => {}
    Err(Error::Api(e)) if e.is_unauthorized() => {}
    Err(Error::Transport(_)) => {}
    Err(e) => return Err(e.into()),
}