Crate adk_rust 

Agent Development Kit (ADK) for Rust

A flexible and modular framework for developing and deploying AI agents in Rust. While optimized for Gemini and the Google ecosystem, ADK is model-agnostic, deployment-agnostic, and compatible with other frameworks.

Quick Start

Create your first AI agent in minutes:

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> Result<()> {
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = GeminiModel::new(&api_key, "gemini-2.0-flash-exp")?;

    let agent = LlmAgentBuilder::new("assistant")
        .description("A helpful AI assistant")
        .instruction("You are a friendly assistant. Answer questions concisely.")
        .model(Arc::new(model))
        .build()?;

    // Run in interactive console mode
    Launcher::new(Arc::new(agent)).run().await?;
    Ok(())
}

Installation

Add to your Cargo.toml:

[dependencies]
adk-rust = "0.1"
tokio = { version = "1.40", features = ["full"] }
dotenv = "0.15"  # For loading .env files

Feature Presets

# Full (default) - Everything included
adk-rust = "0.1"

# Minimal - Agents + Gemini + Runner only
adk-rust = { version = "0.1", default-features = false, features = ["minimal"] }

# Custom - Pick exactly what you need
adk-rust = { version = "0.1", default-features = false, features = [
    "agents", "gemini", "tools", "sessions"
] }

Agent Types

ADK-Rust provides several agent types for different use cases:

LlmAgent - AI-Powered Reasoning

The core agent type that uses Large Language Models for intelligent reasoning:

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

let api_key = std::env::var("GOOGLE_API_KEY")?;
let model = GeminiModel::new(&api_key, "gemini-2.0-flash-exp")?;

let agent = LlmAgentBuilder::new("researcher")
    .description("Research assistant with web search")
    .instruction("Search for information and provide detailed summaries.")
    .model(Arc::new(model))
    .tool(Arc::new(GoogleSearchTool::new()))  // Add tools
    .build()?;

Workflow Agents - Deterministic Pipelines

For predictable, multi-step workflows:

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

// Sequential: Execute agents in order
let pipeline = SequentialAgent::new(
    "content_pipeline",
    vec![researcher, writer, reviewer]
);

// Parallel: Execute agents concurrently
let parallel = ParallelAgent::new(
    "multi_analysis",
    vec![analyst1, analyst2]
);

// Loop: Iterate until condition met
let loop_agent = LoopAgent::new("iterative_refiner", refiner, 5);

Multi-Agent Systems

Build hierarchical agent systems with automatic delegation:

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

let coordinator = LlmAgentBuilder::new("coordinator")
    .description("Development team coordinator")
    .instruction("Delegate coding tasks to specialists.")
    .model(model)
    .sub_agent(code_agent)   // Delegate to sub-agents
    .sub_agent(test_agent)
    .build()?;

Tools

Give your agents capabilities beyond conversation:

Function Tools - Custom Operations

Convert any async function into a tool:

use adk_rust::prelude::*;
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};

#[derive(Debug, Deserialize, JsonSchema)]
struct WeatherInput {
    /// City name to get weather for
    city: String,
}

#[derive(Debug, Serialize)]
struct WeatherOutput {
    temperature: f64,
    conditions: String,
}

async fn get_weather(_ctx: ToolContext, input: WeatherInput) -> Result<WeatherOutput> {
    // Your weather API call here
    Ok(WeatherOutput {
        temperature: 72.0,
        conditions: "Sunny".to_string(),
    })
}

let weather_tool = FunctionTool::new(
    "get_weather",
    "Get current weather for a city",
    get_weather,
);

Built-in Tools

Ready-to-use tools included with ADK:

• GoogleSearchTool - Web search via Google
• ExitLoopTool - Control loop termination
• LoadArtifactsTool - Access stored artifacts

MCP Tools - External Integrations

Connect to Model Context Protocol servers:

use adk_rust::prelude::*;

// Connect to an MCP server (e.g., filesystem, database)
let mcp_tools = McpToolset::from_command("npx", &[
    "-y", "@anthropic/mcp-server-filesystem", "/path/to/dir"
]).await?;

// Add all MCP tools to your agent
let agent = builder.toolset(Arc::new(mcp_tools)).build()?;

Sessions & State

Manage conversation context and working memory:

use adk_rust::prelude::*;

// Create a session
let session = session_service.create("user_123", None).await?;

// Store state with scoped prefixes
let state = session.state();
state.set("app:config", "production");      // App-level config
state.set("user:preference", "dark_mode");  // User preferences
state.set("temp:cache", "computed_value");  // Temporary data

// State persists across conversation turns
let config = state.get::<String>("app:config")?;

Callbacks

Intercept and customize agent behavior:

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

let agent = LlmAgentBuilder::new("monitored_agent")
    .model(model)
    // Log all agent invocations
    .before_agent(|ctx| {
        Box::pin(async move {
            println!("Agent starting: {}", ctx.agent_name);
            Ok(None) // Continue execution
        })
    })
    // Modify or cache model responses
    .after_model(|ctx, response| {
        Box::pin(async move {
            println!("Model responded with {} tokens", response.usage.output_tokens);
            Ok(response)
        })
    })
    // Track tool usage
    .before_tool(|ctx, name, args| {
        Box::pin(async move {
            println!("Calling tool: {} with {:?}", name, args);
            Ok(None)
        })
    })
    .build()?;

Artifacts

Store and retrieve binary data (images, files, etc.):

use adk_rust::prelude::*;

// Save an artifact
let image_data = std::fs::read("chart.png")?;
artifact_service.save(
    "reports",           // namespace
    "sales_chart.png",   // filename
    &image_data,
    "image/png",         // MIME type
).await?;

// Load an artifact
let artifact = artifact_service.load("reports", "sales_chart.png", None).await?;

Deployment Options

Console Mode (Interactive CLI)

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

// Interactive chat in terminal
Launcher::new(agent).run().await?;

Server Mode (REST API)

# Run your agent as a web server
cargo run -- serve --port 8080

Provides endpoints:

• POST /chat - Send messages
• GET /sessions - List sessions
• GET /health - Health check

Agent-to-Agent (A2A) Protocol

Expose your agent for inter-agent communication:

use adk_rust::server::{A2AServer, AgentCard};

let card = AgentCard::new("my_agent", "https://my-agent.example.com")
    .with_description("A helpful assistant")
    .with_skill("research", "Can search and summarize information");

let server = A2AServer::new(agent, card, session_service, artifact_service);
server.serve(8080).await?;

Observability

Built-in OpenTelemetry support for production monitoring:

use adk_rust::telemetry::{TelemetryConfig, init_telemetry};

let config = TelemetryConfig::new("my-agent-service")
    .with_otlp_endpoint("http://localhost:4317");

init_telemetry(config)?;
// All agent operations now emit traces and metrics

Architecture

ADK-Rust uses a layered architecture for modularity:

┌─────────────────────────────────────────────────────────────┐
│                    Application Layer                        │
│              Launcher • REST Server • A2A                   │
├─────────────────────────────────────────────────────────────┤
│                      Runner Layer                           │
│           Agent Execution • Event Streaming                 │
├─────────────────────────────────────────────────────────────┤
│                      Agent Layer                            │
│    LlmAgent • CustomAgent • Sequential • Parallel • Loop    │
├─────────────────────────────────────────────────────────────┤
│                     Service Layer                           │
│      Models • Tools • Sessions • Artifacts • Memory         │
└─────────────────────────────────────────────────────────────┘

Feature Flags

Feature	Description	Default
agents	Agent implementations	✅
models	Model integrations	✅
gemini	Gemini model support	✅
tools	Tool system	✅
mcp	MCP integration	✅
sessions	Session management	✅
artifacts	Artifact storage	✅
memory	Semantic memory	✅
runner	Execution runtime	✅
server	HTTP server	✅
telemetry	OpenTelemetry	✅
cli	CLI launcher	✅

Examples

The examples directory contains working examples for every feature:

• Agents: LLM agent, workflow agents, multi-agent systems
• Tools: Function tools, Google Search, MCP integration
• Sessions: State management, conversation history
• Callbacks: Logging, guardrails, caching
• Deployment: Console, server, A2A protocol

Related Crates

ADK-Rust is composed of modular crates that can be used independently:

• adk-core - Core traits and types
• adk-agent - Agent implementations
• adk-model - LLM integrations
• adk-tool - Tool system
• adk-session - Session management
• adk-artifact - Artifact storage
• adk-runner - Execution runtime
• adk-server - HTTP server
• adk-telemetry - Observability

Re-exports

pub use anyhow;

pub use futures;

pub use serde;

pub use serde_json;

pub use tokio;

Modules

agentagents

Agent implementations (LLM, Custom, Workflow agents).

agent_loader

Core traits and types.

artifactartifacts

Artifact storage.

callbacks

Core traits and types.

context

Core traits and types.

error

Core traits and types.

event

Core traits and types.

instruction_template

Core traits and types.

memorymemory

Memory system with semantic search.

modelmodels

Model integrations (Gemini, etc.).

prelude

Convenience prelude for common imports.

runnerrunner

Agent execution runtime.

serverserver

HTTP server (REST + A2A).

sessionsessions

Session management.

telemetrytelemetry

Telemetry (OpenTelemetry integration).

tooltools

Tool system and built-in tools.

types

Core traits and types.

Structs

Content

Core traits and types.

Event

Core traits and types.

EventActions

Core traits and types.

GenerateContentConfig

Core traits and types.

Launchercli

CLI launcher for running agents.

LlmRequest

Core traits and types.

LlmResponse

Core traits and types.

MemoryEntry

Core traits and types.

MultiAgentLoader

Core traits and types.

RunConfig

Core traits and types.

SingleAgentLoadercli

CLI launcher for running agents.

UsageMetadata

Core traits and types.

Enums

AdkError

Core traits and types.

BeforeModelResult

Core traits and types.

FinishReason

Core traits and types.

IncludeContents

Core traits and types.

Part

Core traits and types.

StreamingMode

Core traits and types.

Constants

KEY_PREFIX_APP

Core traits and types.

KEY_PREFIX_TEMP

Core traits and types.

KEY_PREFIX_USER

Core traits and types.

Traits

Agent

Core traits and types.

AgentLoader

Core traits and types.

Artifacts

Core traits and types.

CallbackContext

Core traits and types.

InvocationContext

Core traits and types.

Llm

Core traits and types.

Memory

Core traits and types.

ReadonlyContext

Core traits and types.

ReadonlyState

Core traits and types.

Session

Core traits and types.

State

Core traits and types.

Tool

Core traits and types.

ToolContext

Core traits and types.

Toolset

Core traits and types.

Functions

inject_session_state

Core traits and types.

Type Aliases

AfterAgentCallback

Core traits and types.

AfterModelCallback

Core traits and types.

AfterToolCallback

Core traits and types.

BeforeAgentCallback

Core traits and types.

BeforeModelCallback

Core traits and types.

BeforeToolCallback

Core traits and types.

EventStream

Core traits and types.

GlobalInstructionProvider

Core traits and types.

InstructionProvider

Core traits and types.

LlmResponseStream

Core traits and types.

Result

Core traits and types.

ToolPredicate

Core traits and types.

Attribute Macros

async_trait