AI Agent SDK (Rust)

English | 中文

One to one Rust translation of claude code source.

Idiomatic agent sdk inspired by the claude code source leak, written in Rust that runs the full agent loop in-process — no subprocess or CLI required. Deploy anywhere: cloud, serverless, Docker, CI/CD.

Quick Demo

use ai_agent_sdk::Agent;

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
let result = agent.prompt("List files in current directory").await?;
println!("{}", result.text);

That's it! The agent automatically uses 20+ built-in tools (Bash, Read, Write, Edit, Glob, Grep, WebFetch, etc.) to accomplish tasks.

Core Features

Quick Start

Install:

cargo add ai-agent-sdk

Configure your API key:

export AI_AUTH_TOKEN=your-api-key
export AI_MODEL=MiniMaxAI/MiniMax-M2.5

Or use .env file:

AI_AUTH_TOKEN=your-token
AI_MODEL=MiniMaxAI/MiniMax-M2.5
AI_BASE_URL=https://api.minimax.chat/v1

Simple Prompt

use ai_agent_sdk::Agent;

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
let result = agent.prompt("What files are in this project?").await?;
println!("{}", result.text);

Multi-turn Conversation

use ai_agent_sdk::Agent;

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 5);
let r1 = agent.prompt("Create file /tmp/hello.txt with 'Hello'").await?;
let r2 = agent.prompt("Read that file back").await?;
println!("Session messages: {}", agent.get_messages().len());

Custom Tools

use ai_agent_sdk::{Agent, ToolDefinition, ToolInputSchema};

fn calculator_impl(input: serde_json::Value, _ctx: ()) -> impl std::future::Future<Output = Result<ai_agent_sdk::ToolResult, ai_agent_sdk::AgentError>> + Send {
    async move {
        let expr = input["expression"].as_str().unwrap_or("0");
        Ok(ai_agent_sdk::ToolResult {
            result_type: "tool_result".to_string(),
            tool_use_id: "".to_string(),
            content: format!("Result of {}", expr),
            is_error: None,
        })
    }
}

let calculator = ai_agent_sdk::Tool {
    name: "Calculator".to_string(),
    description: "Evaluate math expressions".to_string(),
    input_schema: ToolInputSchema::Json(serde_json::json!({
        "type": "object",
        "properties": {"expression": {"type": "string"}},
        "required": ["expression"]
    })),
    executor: Box::new(calculator_impl),
};

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
// Note: Tool registration via agent tools currently uses ToolDefinition

MCP Servers

use ai_agent_sdk::{Agent, McpServerConfig};

let config = McpServerConfig::Stdio(ai_agent_sdk::McpStdioConfig {
    command: "npx".to_string(),
    args: Some(vec!["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]),
    ..Default::default()
});

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
// Note: MCP servers configured via AgentOptions

Subagents

use ai_agent_sdk::{Agent, AgentDefinition};

let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
// Note: Subagents are used via the Agent tool at runtime

Skills

use ai_agent_sdk::{Agent, load_skill_from_dir};

let skill = load_skill_from_dir("skills/debug")?;
let mut agent = Agent::new("MiniMaxAI/MiniMax-M2.5", 10);
// Note: Skills are applied via agent configuration

Plugin System

use ai_agent_sdk::{load_plugin, load_plugins_from_dir, CommandRegistry};

let plugin = load_plugin("plugins/hello-plugin").await?;
let plugins = load_plugins_from_dir("plugins").await?;
let registry = CommandRegistry::new();

Hooks

use ai_agent_sdk::hooks::{HookRegistry, HookDefinition, HookInput};

let mut registry = HookRegistry::new();
registry.register("PreToolUse", HookDefinition {
    command: Some("echo pre-tool".to_string()),
    timeout: Some(5000),
    matcher: Some("Read.*".to_string()),
});

let mut input = HookInput::new("PreToolUse");
input.tool_name = Some("Read".to_string());
let results = registry.execute("PreToolUse", input).await;

Configuration

Options

Environment Variables

Built-in Tools

Architecture

┌─────────────────────────────────────┐
│         Your Application             │
│   use ai_agent_sdk::Agent            │
└──────────────┬──────────────────────┘
               │
    ┌──────────▼──────────┐
    │       Agent         │  Session state, tools
    │    prompt()         │  MCP connections
    └──────────┬──────────┘
               │
    ┌──────────▼──────────┐
    │    QueryEngine      │  Agent loop:
    │   submitMessage()   │  API → tools → repeat
    └──────────┬──────────┘
               │
    ┌──────────┼──────────┐
    │          │          │
┌───▼───┐  ┌───▼───┐  ┌──▼────┐
│  LLM  │  │ 20+   │  │  MCP  │
│  API  │  │Tools  │  │Server │
└───────┘  └───────┘  └───────┘

Examples

Run examples from examples/ directory:

# Configure your .env first
cargo run --example 01_simple_query
cargo run --example 18_plugin
cargo run --example 19_hooks

License

MIT