Google Cloud Storage Joins MCP Toolbox: Bucket Tools for AI Agents

Google Cloud Storage is a core piece of the modern agentic stack and a preferred home for unstructured data at scale. As teams move from prototyping to running autonomous agents in production, though, a shift in architecture shows up: storing data is no longer the hard part. The challenge is converting petabytes of raw, unstructured files into context an agent can reach on demand. To address that, GCS is now supported natively and generally available in the MCP Toolbox.

Unstructured data in agent applications

Databases hold structured facts, while object stores hold the day-to-day reality of a business: PDF contracts, architectural diagrams, financial and compliance papers, raw application logs and multi-modal media. Giving agents a standard way to talk to GCS enables several kinds of applications.

Multi-modal analysis is the first. Agents can pair the deterministic strength of structured data in BigQuery or AlloyDB with the richer context stored in GCS. An automated insurance claims agent, for instance, could read a customer's policy limits and deductibles from an AlloyDB database while retrieving crash-scene photos and scanned police reports from GCS, then assess the damage and recommend a payout.

Automated document pipelines are the second. A background agent can watch an intake bucket for newly uploaded vendor invoices or legal contracts, extract the key entities and trigger downstream ERP workflows.

Codebase and log forensics come third. Site reliability agents can dig into buckets full of historical server logs or crash dumps for root-cause analysis, with no developer manually downloading gigabytes of text.

The fourth is dynamic, on-the-fly RAG, a pattern emerging among customers who are moving beyond rigid vector-embedding pipelines. Rather than chunking and indexing everything in advance, you grant an agent secure access to a particular GCS directory. When a user asks a complex question, the agent lists the bucket contents, finds the relevant 200-page manual, reads the raw file and builds the answer straight from the source.

Companies such as Snap, Palo Alto Networks and Airwallex are already using GCS with MCP, according to Google's developer blog.

What shipped

Since the MCP Toolbox launched, native support for large object stores like GCS has been among the most requested features. Like the other Toolbox sources, the GCS integration is easy to use, and the quick setup instructions in the Cloud Storage documentation connect a client right away.

It comes with pre-built tools covering most core GCS functions: creating, listing, reading and updating buckets, objects and their metadata. For easier management there are also two built-in toolsets.

• cloud-storage-buckets equips an agent to manage storage infrastructure, for example listing buckets and fetching bucket metadata.
• cloud-storage-objects lets an agent work with the unstructured data itself, for example listing files, reading object contents, uploading and deleting.

Additional tools are planned over the next few months.

Custom tools for stricter business logic

Pre-built tools suffice for most cases, but some patterns need custom business logic, and the Toolbox already offers a declarative, extensible framework of custom tools. The native GCS integration can serve as a building block for tools tailored to a specific business need.

A typical example is limiting an agent's access to a handful of buckets so that nothing leaks in the future. Wrapping the GCS integration in a custom tool allows a hardcoded allow-list, which removes the risk of prompt injection leading to traversal across buckets. A tool that lists only allow-listed buckets looks like this:

kind: tool
name: list_allowlisted_buckets
type: cloud-storage-list-buckets
source: my-gcs-source
description: Use this tool to list Cloud Storage buckets allowlisted for the vendor risk assement agent
project: my_safe_project"   # hard-coded GCP project
prefix: ai_buckets          # allowlisted bucket prefix

In the same way, reading objects can be restricted to an allow-listed, safe bucket:

kind: tool
name: read_allowlisted_objects
type: cloud-storage-read-object
source: my-gcs-source
description: Use this tool to read contents of available, allowlisted Cloud Storage objects
bucket: bucket_for_agents   # allowlisted GCS bucket

Getting started

Connecting the largest unstructured data stores directly to AI agents opens many possibilities. The official documentation for GCS with the Toolbox, the reference page for the pre-built tools and toolsets, and the MCP Toolbox repository on GitHub are the places to begin.