Google's Developer Knowledge API and MCP Server: Official Docs for Your Coding Agent, and the Limits Google Lists

Coding agents often write against stale docs. Google's answer for its own products is the Developer Knowledge API and a matching MCP server that serve Google's public developer documentation as Markdown.

The API itself is not new. Google announced a public preview on February 4, 2026, and on October 7 published a post adding a gcloud surface, an agent skill and client libraries. This article reads that post plus the API page, the MCP server page and the original announcement. I have not used it.

What it offers

Google describes the API as the programmatic source of truth for its public developer documentation, covering products such as Firebase, Google Cloud, Android and Google Maps Platform. It exposes three functions:

• SearchDocumentChunks: find relevant page URIs and snippets for a query.
• GetDocument and BatchGetDocuments: fetch the full Markdown of results.
• AnswerQuery: generate answers drawn from the documentation corpus.

The MCP server maps these to three tools: search_documents, get_documents and answer_query. Google recommends a search-then-fetch pattern, so the agent reads short chunks first and pulls full pages only when needed, which it says keeps token use down.

Ways to use it

Per Google's docs:

• REST and RPC endpoints, plus client libraries for C#, Go, Java, Node.js and TypeScript, PHP, Python and Ruby.
• gcloud CLI, for example gcloud developer-knowledge answer-query --query="...", documents search-chunks and documents describe. Google's post says it comes pre-installed in Cloud Shell.
• MCP server, for coding assistants. Google names Claude Code, Cursor, GitHub Copilot, Codex and Antigravity as tools it works with.
• An agent skill, installed with npx skills add google/skills --skill retrieving-developer-knowledge. It tells the assistant which tool to pick, how to handle errors, and how to fall back to the REST API with curl if the MCP server is unavailable.

Google's guidance on choosing: use the API directly if you want custom tools, need to filter results by data_source or update_time, or want to process results before they reach a model. Use the MCP server and skill if you want to connect an assistant without writing tool definitions and want automatic updates as the server adds capabilities.

Authentication

The MCP page describes two approaches:

• API key, passed in the X-Goog-Api-Key header over HTTPS. Google recommends it for third-party IDEs and CLI agents. You need a Google Cloud project, and you enable the Developer Knowledge API there. The page suggests restricting the key to the Developer Knowledge API.
• OAuth and Application Default Credentials, suited to Antigravity or enterprise setups that need an OAuth 2.0 bearer token.

The API page lists two accepted OAuth scopes: cloud-platform and devprofiles.full_control. It says no specific IAM roles are needed to enable or use the API.

One detail from the MCP page: if you plan to reuse the same key for model calls such as GEMINI_API_KEY, you also have to allow the Generative Language API on it. Think about that before sharing one key across uses.

Limits Google lists

• Scope. Only public pages in the published corpus are included. Content from GitHub, open-source sites, blogs and YouTube is not.
• Markdown quality. The Markdown is generated from HTML, so there may be discrepancies or formatting issues.
• Freshness. The API page states a goal of re-indexing within 48 hours of publication, so updates are available within 2 business days. The February announcement said documentation was re-indexed within 24 hours during public preview, and the October post says content is indexed frequently. The three statements differ, so do not assume a specific lag.
• Status. The only status I found is the public preview announcement from February. I did not find a general availability statement in the pages I read.

What I would check

These are my own untested suggestions.

1. Ask the agent a question about a recently changed Google API and compare its answer with the live docs page.

2. Check whether your use case needs content outside the corpus, such as GitHub samples, since those are excluded.

3. Treat answer_query output as generated text. Spot-check it against the page it cites.

4. Decide where the API key lives and who can read it.

What I would not conclude

• That agents with this server stop hallucinating about Google APIs. Google claims it reduces them, but I have no measurement.
• That documentation older than two business days is always missing. Google states a goal, not a guarantee.
• That it covers non-Google libraries. The scope is Google's public docs.

Frequently asked questions

What is the Developer Knowledge API?

Google's programmatic access to its public developer documentation: search, fetch as Markdown, and grounded answers.

What tools does the MCP server expose?

search_documents, get_documents and answer_query.

How do I authenticate?

With an API key in the X-Goog-Api-Key header, or with OAuth and Application Default Credentials.

Does it cover GitHub repos or blogs?

No. Google says only public pages in its corpus reference are included.

How fresh is the content?

Google's API page states a goal of re-indexing within 48 hours of publication. Earlier statements said 24 hours during preview.

Is it generally available?

I found only the February 2026 public preview announcement and no later status statement in the pages I read.

Is there a command-line option?

Yes. gcloud developer-knowledge has answer-query, documents search-chunks and documents describe.

Sources

• Google Developers Blog, Supercharge your development with the Google Developer Knowledge API ecosystem (October 7, 2026): gcloud, agent skill, client libraries.
• Developer Knowledge API: functions, auth, scopes, limitations, freshness goal.
• Connect to the Developer Knowledge MCP server: tools, authentication options, key setup. I read the first part of the page.
• Google Developers Blog, Introducing the Developer Knowledge API and MCP Server (February 4, 2026): public preview and 24-hour freshness statement. I read only parts of it.