For Rust devs

Native Rust. No FFI tax.

The Rust crate is the engine. Python and TS bindings are wrappers; Rust users get the runtime direct, with zero abstraction cost.

Install

cargo add agentmatic
# or in Cargo.toml:
# agentmatic = "0.1"

Five-line agent

use agentmatic::prelude::*;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let agent = Agent::builder("math-helper")
        .llm(OpenAI::from_env()?)
        .tools(vec![calculator_tool()])
        .build()?;
    println!("{}", agent.invoke("What is 42 * 17?").await?);
    Ok(())
}

What's different from other Rust LLM crates

• Graph-first, not LLM-first. StateGraph, channels, supersteps — the same primitives as the Python SDK, with Rust ergonomics.
• Tokio-native. Runs on any Tokio runtime; no synchronous bridge.
• Zero-copy state. Arc<Frame> snapshots; clone is cheap, mutation is COW.
• Trait-based LLM providers. Implement LlmProvider for any backend; OpenAI / Anthropic / Bedrock are just trait impls.
• No serde Value dancing. Strong types end-to-end.

Workspace structure

agentmatic = { version = "0.1", features = ["openai", "anthropic", "mcp", "postgres"] }
# Or pull individual crates:
# agentmatic-core         — Pregel runtime
# agentmatic-agent        — Builder + prebuilt patterns
# agentmatic-llm          — Provider abstraction
# agentmatic-tools        — @tool + MCP client
# agentmatic-memory       — Conversation memory
# agentmatic-vectorstore  — Vector store abstraction

What we ship for Rust specifically

• Axum extractor — turn any agent into a route handler.
• Tower service impl — compose with retry, timeout, rate-limit middleware.
• OpenTelemetry tracing wired through tracing.
• Native FFI exports (agentmatic-ffi) if you want to embed in Go or Java.