A model-free ADK MCP check: discover tools, search docs, then read a page #8030

When connecting ADK to a remote MCP server, I find it useful to check discovery and a real tool call before bringing a model into the loop. I added a small runnable example to the BlockVectra agent examples I maintain:

uv run --script https://raw.githubusercontent.com/blockvectra/agent-quickstart/3c705a97ee6abd8040a473045bbb5d9a1c0d8a6a/examples/adk_docs.py

Python 3.11+ and uv are required. The script pins google-adk[mcp]==2.14.0 and mcp==2.3.0. It uses McpToolset and an in-memory ADK tool context to:

1. Connect to the remote Streamable HTTP documentation endpoint.
2. Verify that it exposes search_docs and read_doc.
3. Search for quickstart, read that page, and print the Markdown in JSON.

No model call, Google Cloud project, wallet or API key is needed. The server exposes only the two documentation tools; the example does not execute the blockchain or account steps described in the page.

I ran this once with Python 3.12: five MCP HTTP requests returned 200/202/200/200/200, and both tool calls succeeded. That checks client compatibility, not a model-backed agent run or a performance comparison.

The example closes the toolset in finally, has 15-second HTTP/read timeouts and a 60-second cooperative total timeout (dependency installation is separate). ADK's own session recovery remains in place. For an agent that uses a model, the same toolset can go in its tools list with model setup handled separately, as in the ADK MCP guide.

Fixed source and bilingual usage instructions.