Google ADK Java adds Gemini caching and a switch to stop configs launching local MCP servers

Reuse the shared instruction, control the local process

Google’s Java Agent Development Kit has two practical additions for developers running tool-using agents: explicit Gemini context caching and a switch that can reject local MCP servers supplied by an agent configuration file. Both appear in version 1.12.0, published October 14.

The caching sample illustrates a weather assistant answering several questions against the same large reference instruction. Once the cacheable prefix is stable, the app creates a cache on the Gemini backend and later turns reference it. The runner reports cache creation, reuse and recreation after expiry, together with prompt and cached-token counts.

This gives developers a concrete way to inspect repeated-context handling. It does not establish a universal latency or cost saving: the sample warns that output depends on model latency and minimum cache sizes. AITrending reviewed the documentation and implementation notes but did not run a performance comparison.

Shared agent configs can start programs before a model responds

The MCP change addresses a separate boundary. The original commit explains that stdioServerParams and stdioConnectionParams contain a command and arguments that launch a local process when tools are resolved on the first turn, before the model is contacted. Agent configs may arrive through shared bundles, templates or registry entries.

Applications can now reject those configuration-supplied local servers with McpToolset.setAllowConfigStdioServers(false), or operators can set -Dadk.mcp.allowConfigStdioServers=false at launch. The default preserves existing behavior and permits them. Upgrading alone therefore does not activate the restriction.

Remote SSE transports and MCP toolsets constructed directly in Java are unaffected. This is a targeted configuration-loading control, not a sandbox for every tool or a claim that all local execution is blocked.

A Java framework upgrade, with experimental graphs too

The release also introduces experimental workflow graph types, including nodes, edges and joins, and fixes approved tool confirmations under ParallelAgent. These are framework changes rather than a new Gemini model.

The documented installation route uses Maven’s com.google.adk:google-adk:1.12.0, with a separate development-interface package. The cache sample requires Java 17+, Maven 3.9+ and either Gemini API credentials or Vertex AI configuration. Backend usage remains subject to its pricing; the Apache 2.0 framework license does not remove inference or cache charges. The useful upgrade decision is specific: Java teams can evaluate the new cache lifecycle and choose explicitly whether imported agent configs should be allowed to start local MCP processes.