Build intelligent agents in pure code

Open-source Python framework for AI agents. Define, test, debug, and deploy — with tools that actually work.

agent.py

from google.adk import Agent

agent = Agent(
    name="assistant",
    model="gemini-2.0-flash",
    instruction="You are a helpful assistant.",
    tools=[search, code_exec],
)

Python pip install google-adk

Java com.google.adk:google-adk

TS npm install @google/adk

Go go get google.golang.org/adk

Your coding agent already knows ADK

ADK ships as a Skill and MCP server — coding agents like Claude Code and Gemini CLI can scaffold, extend, and test your agents automatically.

        
$ claude "Build me a weather agent with ADK"
⠋ Using ADK skill + MCP server...

fromgoogle.adkimport Agent
fromgoogle.adk.toolsimport google_search

agent = Agent(

        name="weather",

        model="gemini-2.5-flash",

        tools=[google_search],

      )

✓Agent created with tool bindings

Any Model, Your Infra

Use Gemini, GPT, Claude, or open models. Run on your laptop, Cloud Run, or your own infrastructure.

Single Agent → Multi-Agent → Graph Workflows

ADK 2.0 — Graphs + Code

Graph-based agents combine AI reasoning with deterministic logic. Mix LLM calls with regular code for workflows you can actually trust.

Graphs + Code = Reliable Agents

Built-in Dev UI

Interactive web UI for testing, debugging, and tracing your agents. See every tool call, inspect state, and iterate fast.

Go beyond vibes

ADK's eval framework tests both the final response AND the agent's trajectory — the steps it took to get there.