Google ADK
Open-source AI agent framework for self-hosted workflows, with local model adapters, Gemini support, and an Apache 2.0 license.
Google ADK is an open-source AI agent framework for developers building applications that carry out multi-step tasks. You can run agents locally in Docker or on your own infrastructure, and connect them to locally running models through adapters. The framework is optimized for Gemini but supports other models and providers. Its Python repository uses the Apache 2.0 license.
ADK combines model reasoning with explicit workflow logic. Graph workflows let developers define execution paths, run tasks in parallel, repeat steps, retry failures, and pause for human input. Multiple specialized agents can delegate work to one another, with support for both ongoing exchanges and single-turn tasks with controlled output. Tool confirmation can require approval before an agent acts.
Agents can use custom functions, MCP tools, and tools described by OpenAPI specifications. A built-in development UI supports testing, debugging, and evaluation, so developers can assess agent behavior before deployment. The framework is available in Python, TypeScript, Go, Java, and Kotlin.
Context management handles sessions, memory, tool results, and artifacts together. It filters irrelevant events, summarizes older conversation turns, loads artifacts as needed, and tracks token usage. Developers can customize that behavior for more complex tasks.
Self-hosting and model hosting are separate choices: agents running on your infrastructure can still call hosted models. ADK also supports deployment to Google Cloud through Agent Runtime, Cloud Run, or GKE.
