ADK-Rust

Rust Agent Development Kit (ADK-Rust) - Build AI agents in Rust with modular components for models, tools, memory, RAG, security, realtime voice, and more.

A flexible framework for developing AI agents with simplicity and power. Model-agnostic, deployment-agnostic, optimized for frontier AI models. Includes support for realtime voice agents, RAG pipelines, graph workflows, declarative security, and 120+ working examples.

Supported Providers

Quick Start

1. Create a new project:

cargo new my_agent && cd my_agent

2. Add dependencies:

[dependencies]
adk-rust = "2.0.0"
tokio = { version = "1.40", features = ["full"] }
dotenvy = "0.15"

3. Set your API key:

echo 'GOOGLE_API_KEY=your-key' > .env

4. Write src/main.rs:

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

#[tokio::main]
async fn main() -> AnyhowResult<()> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = GeminiModel::new(&api_key, "gemini-2.5-flash")?;

    let agent = LlmAgentBuilder::new("assistant")
        .instruction("You are a helpful assistant.")
        .model(Arc::new(model))
        .build()?;

    Launcher::new(Arc::new(agent)).run().await?;
    Ok(())
}

5. Run:

cargo run

Even Faster — adk::run()

Auto-detect your provider and run an agent in one call:

use adk_rust::run;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();
    // Set ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY
    let response = run("You are a helpful assistant.", "What is Rust?").await?;
    println!("{response}");
    Ok(())
}

provider_from_env() checks ANTHROPIC_API_KEY → OPENAI_API_KEY → GOOGLE_API_KEY in order.

Adding Tools

let agent = LlmAgentBuilder::new("researcher")
    .instruction("Search the web and summarize findings.")
    .model(Arc::new(model))
    .tool(Arc::new(GoogleSearchTool::new()))
    .build()?;

Workflow Agents

// Sequential execution
let pipeline = SequentialAgent::new("pipeline", vec![agent1, agent2, agent3]);

// Parallel execution
let parallel = ParallelAgent::new("analysis", vec![analyst1, analyst2]);

// Loop until condition (max 5 iterations)
let loop_agent = LoopAgent::new("refiner", vec![agent])
    .with_max_iterations(5);

Multi-Agent Systems

let coordinator = LlmAgentBuilder::new("coordinator")
    .instruction("Delegate tasks to specialists.")
    .model(model)
    .sub_agent(code_agent)
    .sub_agent(test_agent)
    .build()?;

Realtime Voice Agents

Build voice-enabled AI assistants with bidirectional audio streaming:

use adk_realtime::{RealtimeAgent, openai::OpenAIRealtimeModel, RealtimeModel};

let model: Arc<dyn RealtimeModel> = Arc::new(
    OpenAIRealtimeModel::new(&api_key, "gpt-realtime")
);

let agent = RealtimeAgent::builder("voice_assistant")
    .model(model)
    .instruction("You are a helpful voice assistant.")
    .voice("alloy")
    .server_vad()  // Voice activity detection
    .build()?;

Features:

• OpenAI Realtime API & Gemini Live API
• Vertex AI Live with Application Default Credentials
• LiveKit WebRTC bridge for production audio routing
• Bidirectional audio (PCM16, G711, Opus)
• Server-side VAD
• Mid-session context mutation (swap instructions/tools without dropping the call)
• Real-time tool calling
• Multi-agent handoffs

Graph-Based Workflows

Build complex workflows using LangGraph-style graph agents:

use adk_graph::prelude::*;

let agent = GraphAgent::builder("processor")
    .node_fn("fetch", |ctx| async move { /* ... */ })
    .node_fn("transform", |ctx| async move { /* ... */ })
    .edge(START, "fetch")
    .edge("fetch", "transform")
    .edge("transform", END)
    .checkpointer(SqliteCheckpointer::new("sqlite:state.db?mode=rwc").await?)
    .build()?;

Features:

• Cyclic graphs for ReAct patterns
• Conditional routing, and with_goto for a node that picks its own successor
• State management with reducers
• Checkpointing (memory, SQLite), delta checkpoints, and retention policies
• Human-in-the-loop interrupts, before or after a node
• Subgraphs — a compiled graph as a node, nested, with its own checkpoint thread
• Per-node retry with capped backoff, node timeouts, and a concurrency bound
• Time travel — step, fork, and state history

The SqliteCheckpointer above needs the graph-sqlite feature. adk-graph has no default features, so each capability is forwarded from this crate:

Browser Automation

Give agents web browsing capabilities with 46 tools:

use adk_browser::{BrowserSession, BrowserToolset, BrowserConfig};
use std::sync::Arc;

let config = BrowserConfig::new().webdriver_url("http://localhost:4444");
let session = Arc::new(BrowserSession::new(config));
let toolset = BrowserToolset::new(session);
let tools = toolset.all_tools();  // 46 browser tools

let mut builder = LlmAgentBuilder::new("web_agent")
    .model(model);

for tool in tools {
    builder = builder.tool(tool);
}

let agent = builder.build()?;

Tools include navigation, extraction, forms, screenshots, JavaScript execution, and more.

Agent Evaluation

Test and validate agent behavior:

use adk_eval::{Evaluator, EvaluationConfig, EvaluationCriteria};

let evaluator = Evaluator::new(EvaluationConfig::with_criteria(
    EvaluationCriteria::exact_tools().with_response_similarity(0.8)
));

let report = evaluator.evaluate_file(agent, "tests/agent.test.json").await?;
assert!(report.all_passed());

RAG Pipeline

Build retrieval-augmented generation pipelines with adk-rag:

use adk_rag::{RagPipelineBuilder, InMemoryVectorStore, FixedSizeChunker};

let pipeline = RagPipelineBuilder::new()
    .chunker(FixedSizeChunker::new(512, 50))
    .vector_store(InMemoryVectorStore::new())
    .embedding_provider(embedding_model)
    .build()?;

// Ingest documents
pipeline.ingest("doc1", "Your document text here...").await?;

// Search
let results = pipeline.search("query", 5).await?;

Features:

• 3 chunking strategies (fixed-size, recursive, markdown-aware)
• 6 vector store backends (in-memory, Qdrant, Milvus, Weaviate, Pinecone, SurrealDB)
• Pluggable embedding providers and rerankers
• Ready-made RagTool for agent integration

Declarative Security

Tools declare their required scopes — the framework enforces automatically:

use adk_tool::FunctionTool;
use adk_auth::{ScopeGuard, ContextScopeResolver};

// Tool declares what scopes it needs
let transfer = FunctionTool::new("transfer", "Transfer funds", handler)
    .with_scopes(&["finance:write", "verified"]);

// Framework enforces before execution — no imperative checks in handlers
let guard = ScopeGuard::new(ContextScopeResolver);
let protected = guard.protect(transfer);

Features:

• Declarative required_scopes() on the Tool trait
• ScopeGuard with pluggable resolvers (context, static, custom)
• Role-based access control with allow/deny rules
• SSO/OAuth integration (Auth0, Okta, Azure AD, Google OIDC)
• Audit logging for all access decisions

Deployment

# Console mode (default)
cargo run

# Server mode
cargo run -- serve --port 8080

Installation Options

# Minimal (default) — agents, Gemini, runner, sessions (fastest build)
adk-rust = "2.0.0"

# Standard — minimal + tools, memory, OpenAI, Anthropic, server, auth,
# graph, eval, guardrails, skills, plugins, artifacts, telemetry
adk-rust = { version = "2.0.0", features = ["standard"] }

# Enterprise — standard + realtime, browser, rag, payments, awp
adk-rust = { version = "2.0.0", features = ["enterprise"] }

# Full — enterprise + experimental crates (audio, code, sandbox, code-tools)
adk-rust = { version = "2.0.0", features = ["full"] }

# Gemini Enterprise Agent Platform — every Vertex/EAP integration except
# realtime transports; composable with any tier. The right default for
# ReasoningEngine BYOC deployments. Deploy-time tooling is host-side and
# not included.
adk-rust = { version = "2.0.0", features = ["standard", "gemini-agent-platform"] }

# gemini-agent-platform + Vertex AI Live API (pulls in the realtime WebSocket/audio stack)
adk-rust = { version = "2.0.0", features = ["standard", "gemini-agent-platform-full"] }

# Custom
adk-rust = { version = "2.0.0", default-features = false, features = ["agents", "gemini", "tools"] }

# With new providers (forwarded to adk-model)
adk-model = { version = "2.0.0", features = ["fireworks", "together", "mistral", "perplexity", "cerebras", "sambanova", "bedrock", "azure-ai"] }

Documentation

• API Reference
• Official Guides
• Examples — 75+ working examples in-repo, 120+ in the playground
• Wiki — Comprehensive guides and tutorials

License

Apache 2.0