Google Cloud: developer AI tooling

Google spent the past week pushing two tools at developers: one to feed coding agents fresher documentation, and one to run AI agents on a single device. (developers.googleblog.com) The first release is the Developer Knowledge API, which Google announced in public preview on February 4, 2026. It lets software pull Google documentation as Markdown instead of relying on older model training or web scraping. (developers.googleblog.com) Google paired that application programming interface with a Model Context Protocol server, a standard connector that lets an AI tool query outside systems. Google said the server can search official docs for Firebase, Google Cloud, Android, Maps, and other products, and the preview re-indexes docs within 24 hours of an update. (developers.googleblog.com) In plain terms, the setup works like a live reference desk for coding assistants. Instead of guessing from stale memory, the assistant can search document snippets with `search_documents` and pull full pages with `get_documents` from Google’s own library. (developers.google.com) Google has been building out that connector layer across its stack since late 2025. In February and March 2026, the company added managed Model Context Protocol servers for databases, BigQuery, Google Maps, Google Kubernetes Engine, Cloud Run, and other services, all hosted by Google rather than by the developer. (cloud.google.com 1) (cloud.google.com 2) The second push came on April 2, 2026, when Google published a Gemma 4 demo aimed at “edge” devices, meaning phones, laptops, desktops, and other local hardware. Google said Gemma 4 supports multi-step planning, visual input, offline code generation, and more than 140 languages under an Apache 2.0 license. (developers.googleblog.com) Google’s demo centered on “Agent Skills” in the Google Artificial Intelligence Edge Gallery app for iOS and Android. The company said those on-device workflows can query sources like Wikipedia, summarize text or video into flashcards, build charts from speech input, and connect to other models for tasks like text-to-speech. (developers.googleblog.com) The hardware pitch is that smaller teams do not always need a rack of servers to test an agent. Google said developers can run Gemma 4 models through Android’s new Artificial Intelligence Core developer preview or through LiteRT-LM on mobile, desktop, and internet-of-things devices, while separate Google Cloud posts positioned larger Gemma 4 variants for Google Tensor Processing Units and Vertex Artificial Intelligence. (developers.googleblog.com) (cloud.google.com) Taken together, the two releases point at the same problem: an agent is only useful if it can reach current information and run in the places developers actually have. Google’s answer is a managed document gateway on one end and lighter-weight local models on the other. (developers.googleblog.com 1) (developers.googleblog.com 2)