What changed
Google added retrieving-developer-knowledge to its public google/skills repository on September 25. The skill teaches AI coding assistants how to use Google's Developer Knowledge MCP server: choose answer_query for broad how-to questions, use search_documents for focused lookups, verify that retrieval succeeded, and fall back to the REST API when MCP is unavailable.
The underlying Developer Knowledge MCP server exposes three useful primitives: search relevant excerpts, retrieve full Markdown documents, and generate answers grounded in Google's public developer documentation. Google also documents installation for Antigravity, Claude Code, Cursor, Codex, and other clients that support the open Agent Skills standard.
Why it matters
This is a product shift from “give the agent a documentation endpoint” to “give the agent a repeatable retrieval policy.” That distinction matters in real developer workflows. A coding assistant needs to know when a short search result is enough, when it should fetch the surrounding document, and what to do when an MCP connection or API key fails. Those decisions affect correctness, latency, token cost, and whether the assistant quietly falls back to stale model memory.
It also makes official documentation a more portable context layer. The same skill can travel with a project or agent setup while the transport stays flexible: remote MCP for clients that support it, HTTPS REST for everything else.
Builder takeaway
For teams building documentation-aware agents, the reusable unit is no longer only a tool schema. It is the combination of a trusted corpus, a routing policy, an error check, and a fallback path. Start with narrow retrieval, keep citations or source references attached to the result, and make degraded behavior explicit instead of allowing an agent to guess.
Google's release is a useful template for turning retrieval quality into a product feature that can be installed, reviewed, and reused across agent clients.