Google AI Updates: April 21, 2026

1. Google Cloud Promotes Wave of Remote MCP Servers to GA Across Data and Compute Stack

Google. Google Cloud moved a broad set of remote Model Context Protocol servers to general availability on April 20, covering Cloud Storage, Compute Engine, BigQuery, Spanner, Firestore, Bigtable, AlloyDB, Cloud SQL across all three engines, Memorystore (Redis and Valkey), Managed Service for Apache Kafka, and Pub/Sub. The servers let LLM-based agents create resources, run read-only SQL, manage instances, and read or write data through a uniform interface. The release effectively makes most of Google Cloud’s data and compute services agent-addressable by default. Source

2. Database Insights MCP Server Hits GA for Spanner, AlloyDB, Bigtable, and Cloud SQL

Google. Alongside the broader MCP wave, Google Cloud took the Database Insights remote MCP server to GA for AlloyDB, Bigtable, Spanner, and all three Cloud SQL engines. The server exposes performance and system metrics so AI agents can analyze database health and diagnose issues conversationally instead of through dashboards. It complements each database’s resource-management MCP server with a dedicated observability surface. Source

3. BigQuery Conversational Analytics Agents Can Now Be Published to Gemini Enterprise

Google. A new Preview capability lets teams publish BigQuery Conversational Analytics agents directly into Gemini Enterprise, where business users can query datasets in natural language without standing up a separate analytics app. The change tightens the loop between Google’s data warehouse and its enterprise agent platform, betting that the natural place for ad-hoc data questions is inside the same surface employees already use for chat and search. Source

4. BigQuery Adds Vectorized Python UDFs in Preview Backed by Apache Arrow

Google. BigQuery launched vectorized Python UDFs in Preview, executing batches via Apache Arrow instead of one row at a time. The release also adds Cloud Monitoring metrics for UDF runs, container request concurrency controls, and per-job cost visibility through external_service_costs. The change targets ML and data-engineering workloads where Python UDFs were previously a noticeable performance bottleneck. Source