Google ADK

ADK is an open-source, code-first framework for building, evaluating and deploying agents — not a hosted product. You write agents in Python, Java, Go, TypeScript or Kotlin, run them locally or on your own infrastructure (Cloud Run, GKE, or elsewhere), and optionally deploy to Google's managed runtime, currently rebranded 'Agent Runtime' under the new Gemini Enterprise Agent Platform umbrella. There is no ADK-specific pricing to evaluate because the framework itself costs nothing; what costs money is whichever compute and model tokens the agents you build consume.

The Python SDK (google-adk) has 21,146 GitHub stars and is shipping roughly weekly point releases (v2.6.1 to v2.7.0 in under two weeks), which for a one-year-old framework is unusually fast, sustained iteration backed by a large team rather than a side project. It's also the only major agent framework with multi-language parity: Go (8,660 stars), Java (1,683), TypeScript (1,349) and a newer Kotlin SDK all ship from the same google org, so a team building in a JVM or Go shop isn't stuck bridging to a Python-only tool. ADK ships built-in multi-turn evaluation tooling (CLI and dev UI) for regression-testing agent behavior, native bidirectional audio/text streaming via the Gemini Live API, and hierarchical multi-agent orchestration as first-class primitives rather than a pattern you build yourself. Its biggest structural strength is the Agent2Agent (A2A) protocol for cross-framework agent interoperability: Google donated it to the Linux Foundation in June 2025, and by its one-year mark more than 150 organizations — including AWS, Microsoft, Cisco, Salesforce and SAP — were listed as backers, which means an agent built in ADK can in principle talk to one built in a competing framework without either vendor controlling the wire protocol.

ADK's own docs describe non-Gemini models (Claude, OpenAI, and others via LiteLLM, plus local models via Ollama/vLLM) as fully supported for the core agent loop. But the built-in tools that are otherwise a selling point — Google Search grounding and code execution via the Gemini API — are explicitly restricted: per ADK's own 'Tool limitations' documentation, using either tool excludes using any other tool in that same agent, and built-in tools generally cannot be used inside sub-agents at all (with a narrow, Python-only workaround for the two search tools). These restrictions aren't about which LLM you choose so much as which built-in capabilities you can combine — but they matter most to teams trying to build a Gemini-tool-heavy agent with a non-Gemini root model. A buyer planning a Gemini-first agent won't hit this; a buyer building a mostly-Claude or mostly-GPT agent that also wants Google Search grounding should test this combination early rather than assume parity.

ADK is Apache-2.0 and free to self-host. If you deploy to Google's managed Agent Runtime instead, billing is metered: $0.085 per vCPU-hour and $0.009 per GiB-hour of memory, with a monthly free tier of 50 vCPU-hours and 100 GiB-hours per account, idle time between turns not billed. The managed Memory Bank service (long-term agent memory storage and retrieval) is billed separately — $0.30/GiB-month storage plus per-operation charges — and only starts billing under this structure on September 1, 2026. None of this is ADK pricing per se; it's Google Cloud infrastructure pricing that applies only if you choose the managed runtime over self-hosting, but it's easy to miss because ADK's own docs and GitHub don't mention it at all.

adk-python was created April 2025 and has had continuous, near-daily commit activity through at least mid-August 2026 (last push was yesterday relative to today's check). The visible top contributors on the Python repo are overwhelmingly Google-affiliated usernames and Google's internal sync bot (copybara-github), which is typical for corporate-led open source but means this isn't a community-governed project the way some competing frameworks are — Google decides the roadmap. Open issue counts (561 on Python, 305 on Go, 104 on Java) are proportionate to star counts and dominated by feature requests rather than a backlog of unresolved bugs, based on a sample of the highest-reaction issues.