Machi

Machi is a Rust-based Web3/Web4.0-native AI Agent Framework (v0.8.1) under Apache-2.0 and MIT dual license.

Core Architecture#

The framework is built around these core concepts:

• Agent: Self-contained unit with LLM provider, instructions, tools, and optional sub-agents
• Runner: Stateless execution engine driving the ReAct loop (think → act → observe → repeat)
• Tool / DynTool: Callable capability units for agents
• ChatProvider: Trait abstracting LLM backends

Key Features#

LLM Backend Support

• OpenAI API (GPT-4o, etc.)
• Ollama local models
• Extensible via ChatProvider trait

Web3 Capabilities

• EVM wallet integration (via alloy/kobe-evm)
• x402 protocol support (402-based HTTP & EVM payment flows)

Protocol Support

• MCP (Model Context Protocol) server integration
• A2A (Agent-to-Agent) inter-agent communication

Tools & Extensions

• Built-in toolkit: filesystem, shell, web search
• #[tool] procedural macro for custom tool development
• JSON Schema structured output
• SQLite session persistence

Quick Start#

[dependencies]
machi = "0.8"

use std::sync::Arc;
use machi::prelude::*;

#[tokio::main]
async fn main() -> Result<()> {
    let provider: SharedChatProvider = Arc::new(OpenAI::from_env()?);
    
    let agent = Agent::new("assistant")
        .instructions("You are a helpful assistant.")
        .model("gpt-4o-mini")
        .provider(provider);
    
    let result = agent
        .run("What is the capital of France?", RunConfig::default())
        .await?;
    println!("{}", result.output);
    Ok(())
}

Feature Flags#

• openai / ollama: LLM backends
• wallet: EVM wallet functionality
• mcp / a2a: Protocol support
• x402: Payment protocol
• toolkit: Built-in tool collection
• memory-sqlite: Session persistence
• schema: Structured output
• full: All features (default)

Repository Structure#

Workspace contains: machi (core library), machi-auto (CLI), machi-derive (procedural macros)