Agent Development Kit (ADK) for Go

What it is

ADK Go is the Go version of Agent Development Kit, Google’s open-source framework for building AI agents in code. You write the agent, its model, its tools and the way several agents work together as ordinary Go, so they are tested and versioned like the rest of your service. The README says ADK is optimized for Gemini but model-agnostic, and the repository has samples for Claude and OpenAI too. Version 2.0, generally available since June 30, 2026, adds graph workflows that mix model steps with plain Go functions, and tool confirmation by a person. A launcher built into your program runs it as a terminal chat, a web chat page, a REST API or an A2A server.

1. You write an agent in Go: a model, instructions and tools
2. go get adds google.golang.org/adk/v2 to your module
3. go run agent.go starts a chat in your terminal
4. web api webui serves a chat page on port 8080
5. adkgo deploys the same program to Cloud Run

Who it suits

Go developers who want agents inside their own services, with the types, tests and single binary they already use. Teams on Google Cloud get the most help, since the README names Cloud Run and the deploy tool targets it. Python, Java, Kotlin and TypeScript teams have their own ADK versions.

Skip it if you want to build agents without writing code. The Visual Builder in the web interface is Python only, and so is the docs’ page on evaluating agents.

Where to use it

The README calls ADK deployment-agnostic. The Quickstart gives steps for macOS and Linux, and for Windows in PowerShell or Command Prompt.

• macOS and Linux: the Quickstart’s shell lines, with a .env file for the key
• Windows: the same Go commands, with env.bat in place of .env
• Google Cloud Run: the adkgo tool builds and deploys the program, with no Dockerfile to write
• Claude Code, Cursor and Antigravity: an MCP server that lets them read the ADK docs while they write your code
• Any other coding agent: llms.txt and llms-full.txt, the docs as plain files

Models: Gemini with an API key or through Google Cloud, Claude on Vertex AI or the Anthropic API, and OpenAI or a compatible provider that serves its Responses or Chat Completions API.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with each platform’s steps from the ADK docs. The main route is the Go Quickstart: one folder, one Go file and a Gemini key.

1. Make a folder with agent.go and .env. On Windows the Quickstart uses env.bat in place of .env.
2. Copy the Quickstart code into agent.go. Its agent runs on gemini-flash-latest with the built-in Google Search tool and reads your key from GOOGLE_API_KEY.
3. Run the three module lines. They create the module, add ADK Go 2.0 and fetch what the imports need.
4. Save your Gemini key in .env. The second block shows the macOS and Linux line, and the kit has both Windows forms.
5. Start the agent with go run agent.go. Load your .env first, as the Quickstart’s comment says.

go mod init my-agent/main
go get google.golang.org/adk/v2
go mod tidy

echo 'export GOOGLE_API_KEY="YOUR_API_KEY"' > .env
# Remember to load keys and settings: source .env OR env.bat
go run agent.go

How to use it

The keywords after go run agent.go pick how the agent runs. With none, it chats in the terminal.

In the chat page, pick your agent at the upper left and type a request. The page lists your sessions, their state and every event behind an answer, so you can follow each step. To call the agent from your own code, start the API. The API Server page gives a Go line that moves it to port 8000 with no prefix, so the page’s curl examples work as written.

go run agent.go web -port 8000 api -path_prefix ""

From a second terminal, create a session with a POST to /apps/<app>/users/<user>/sessions/<session>. Then send each message as a POST to /run with the app, user and session IDs. /run returns the turn’s events as one list, and /run_sse streams them.

Example

You want one agent to use another that runs as its own service. The A2A quickstart’s a2a_basic sample has a root agent with a local dice agent, and a prime checker served over A2A on port 8001. With both running, you type to the root agent:

roll a die and check if it's a prime

The root agent hands the roll to roll_agent, which calls roll_die with six sides. Then transfer_to_agent hands the number to prime_agent, the agent on the other server, which calls prime_checking. In the docs’ run the die landed on 3, and the reply was “3 is a prime number.”

For newcomers and experts

If you are new. An agent is a language model with instructions and tools it may call, such as a web search or one of your functions. In ADK Go you build one with llmagent.New, give it a name, a model and an instruction, and hand it to the launcher. Start with the Quickstart agent and its chat page, where the events show what the agent did for each answer. Then add a tool of your own with functiontool.New. If a coding agent writes the code with you, point it at llms-full.txt. The README says that file holds the Go API reference and samples.

For experts. The 2.0 graph engine treats agents, tools and functions as nodes. workflow.NewFunctionNode wraps a Go function, workflow.Chain wires nodes in order, and each node’s typed return value reaches the next through event.Output, with no session state writes. A router node sets event.Routes, and edges with StringRoute, IntRoute or BoolRoute pick the branch. NewJoinNode waits for parallel branches, and workflowagent.New turns the graph into a standard agent.Agent. The workflow samples add pauses for a person’s input and orchestration written as plain Go with NewDynamicNode.

State lives in code, not flags. The Go console has no save, resume or replay options, and the Go API server has no Swagger page or PATCH route for sessions. You pass a persistent session.Service such as session/database in launcher.Config. Upgrading from 1.x means the /v2 import path, a context as the first argument of session.NewEvent, and five new event fields. Versions before v2.2.0 stored three of them under their Go names, so fixed schemas must accept both forms.

Good to know

ADK Go is a library, so it runs where you run it. What leaves your machine depends on the model and tools you give the agent.

• Model calls: your prompts and tool results go to the provider you configure, with your own key or Google Cloud account. The Quickstart agent also uses Gemini’s built-in Google Search tool
• Usage data: the web interface page says the ADK Web UI’s anonymous telemetry is off until you turn it on in User Settings. The choice is kept in ~/.adk/config.json
• Development only: the Quickstart says ADK Web is not meant for production. adkgo deploys to Cloud Run with --no-allow-unauthenticated by default and reaches the service through a local proxy
• Go version: the README and go.mod ask for Go 1.26.10 or newer, while the Quickstart and Installation pages still say 1.25
• Experimental: both A2A quickstarts mark A2A in ADK Go as experimental
• Safety: the safety page names prompt injection through tool results as a risk. It says to escape model output shown in a web page and to run model-written code in a sandbox. The MCP page says to expose only the tools an agent needs
• Removal: there is nothing to uninstall. ADK Go is a module in your go.mod, and the kit says how to take it out
• License: Apache-2.0. The README says internal/httprr has its own LICENSE file