# Google Cloud Storage Plugin
The official plugin for
[Google Cloud Storage](https://cloud.google.com/storage). It bundles a growing
collection of [Agent Skills](https://agentskills.io/home) with a
[Cloud Storage MCP server](#cloud-storage-mcp-server), so your coding agent gets
both vetted GCS expertise and structured tools for buckets and objects.
The skills let you work with your storage resources through natural language in
your preferred CLI or IDE: everyday bucket and object management, file-system
mounts with Cloud Storage FUSE, access-error diagnostics, security assessments,
and infrastructure code generation.
> [!NOTE]
> This plugin is under active development. More skills will be added
> over time.
> [!IMPORTANT]
> **We Want Your Feedback!** Please share your thoughts with us by
> opening an issue on
> [GitHub](https://github.com/GoogleCloudPlatform/google-cloud-storage-plugin/issues).
> Your input is invaluable and helps us improve the project for everyone.
## Contents
- [Installation](#installation)
- [Available Skills](#available-skills)
- [Cloud Storage MCP Server](#cloud-storage-mcp-server)
- [Prerequisites](#prerequisites)
- [Authentication](#authentication)
- [Additional Setup: GCS Security Assessment](#additional-setup-gcs-security-assessment)
- [Example Use Cases](#example-use-cases)
- [Security Reminder: Agent Environment Hardening](#security-reminder-agent-environment-hardening)
- [Support](#support)
- [Contributing](#contributing)
- [License](#license)
## Installation
### Installing in Antigravity CLI
```bash
agy plugin install https://github.com/GoogleCloudPlatform/google-cloud-storage-plugin
```
### Installing in Claude Code
This plugin is listed in Claude Code's official marketplace, which is built in,
so no marketplace needs to be added first:
```bash
claude plugin install google-cloud-storage@claude-plugins-official
```
### Installing in Codex
Add this repository as a marketplace, then install from it:
```bash
codex plugin marketplace add GoogleCloudPlatform/google-cloud-storage-plugin
codex plugin add google-cloud-storage@google-cloud-storage
```
### Installing in Gemini CLI
```bash
gemini extensions install https://github.com/GoogleCloudPlatform/google-cloud-storage-plugin
```
### Installing Skills Only (Skills CLI)
To install the skills without the bundled
[Cloud Storage MCP server](#cloud-storage-mcp-server) or plugin configuration
(prefer a native plugin installation above for the full package), use the
[Skills CLI](https://github.com/vercel-labs/skills):
```bash
npx skills add GoogleCloudPlatform/google-cloud-storage-plugin
```
From the `npx` prompt, select the skills to install, or pass `--skill` to
install a single skill directly:
```bash
npx skills add GoogleCloudPlatform/google-cloud-storage-plugin --skill google-cloud-storage-basics
```
## Available Skills
- [**Google Cloud Storage Basics**](./skills/google-cloud-storage-basics/) —
Everyday GCS expertise: create and configure buckets; upload, download, and
transfer data; control access; manage storage classes, lifecycle, cost, and
data protection — via the gcloud CLI, JSON/XML APIs, client libraries,
Terraform, or Cloud Storage MCP servers.
- [**Google Cloud Storage Bucket Architect**](./skills/google-cloud-storage-bucket-architect/)
— Creates Google Cloud Storage (GCS) buckets: analyzes the workload
(sensitive data, media hosting, ingestion, web hosting, archiving, backup,
logging, analytics, AI/ML, or general-purpose), validates project-level
security settings, and designs a secure-by-default, cost-effective
configuration (location, storage class, security, lifecycle) before creating
it via gcloud, JSON/REST API, Terraform, or SDK client libraries (C++, Java,
Python, and Go).
- [**Google Cloud Storage FUSE**](./skills/google-cloud-storage-fuse/) — Mount
buckets as a POSIX file system with gcsfuse: decide when to use FUSE vs.
direct storage reads, deploy tuned mounts on GKE, Compute Engine, and Cloud
Run, size the file, stat, and list caches, keep file writes and ML
checkpointing safe, and diagnose slow or costly mounts with gcsfuse metrics.
- [**Google Cloud Storage Diagnostic**](./skills/google-cloud-storage-diagnostic/)
— Root-cause access failures: diagnose 403 Permission Denied and other
access errors by analyzing IAM policy bindings, ACLs, Uniform Bucket-Level
Access, Bucket-Level IP Filtering, and VPC Service Controls perimeters, then
walk through a verified fix.
- [**GCS Security Assessment**](./skills/gcs-security-assessment/) — Assess
the security posture of GCS projects and buckets against Google's
[Secure AI Framework (SAIF)](https://saif.google/secure-ai-framework/saif-map):
correlate real telemetry signals to surface **toxic combinations** of
vulnerabilities — scenarios where individually low-risk configurations
combine into a critical exposure — with actionable, verified remediation.
Needs [additional setup](#additional-setup-gcs-security-assessment) for a
complete assessment.
- [**Google Cloud Storage OKF**](./skills/google-cloud-storage-okf/) —
Token-efficient search, semantic navigation, and progressive disclosure of
unstructured or semi-structured objects (PDFs, logs, JSON/CSV dumps, media,
large documents) in GCS using the Open Knowledge Format (OKF) and Model
Context Protocol (MCP).
## Cloud Storage MCP Server
Installing the plugin also configures a local Cloud Storage MCP server, so your
agent can call structured storage tools instead of only shelling out to the CLI.
The server is [MCP Toolbox](https://github.com/googleapis/mcp-toolbox) running
its prebuilt `cloud-storage` tools over stdio. It starts on demand through
`npx`, so there is no binary to download, but Node.js must be installed.
It authenticates with the same Application Default Credentials as the skills
(see [Authentication](#authentication)), and needs the `roles/storage.*` roles
for the operations you call: `roles/storage.objectViewer` to read,
`roles/storage.objectAdmin` to copy, move, and delete objects, and
`roles/storage.admin` to create or delete buckets.
### Configuration
The server needs one setting:
* `CLOUD_STORAGE_PROJECT`: the Google Cloud project it operates on.
How you supply it depends on the harness:
* **Antigravity** and **Codex**: export it before starting your agent:
```bash
export CLOUD_STORAGE_PROJECT="<your-project-id>"
```
* **Claude Code**: pass `--config cloud_storage_project=<project-id>` on
install, or run `/plugin` inside Claude Code.
* **Gemini CLI**: prompted on install. View or update later with `gemini
extensions config google-cloud-storage` (restart the CLI to apply).
### Available Tools
Tool | Description
:---------------------- | :----------------------------------------------
`list_buckets` | List buckets in the project.
`get_bucket_metadata` | Get a bucket's metadata.
`get_bucket_iam_policy` | Get a bucket's IAM policy.
`create_bucket` | Create a bucket.
`delete_bucket` | Delete an empty bucket.
`list_objects` | List objects in a bucket.
`get_object_metadata` | Get an object's metadata.
`read_object` | Read UTF-8 text content (up to 8 MiB).
`download_object` | Download an object (binary included) to a file.
`write_object` | Write text content to an object (overwrites).
`upload_object` | Upload a local file (binary included).
`copy_object` | Copy an object.
`move_object` | Rename an object; deletes the source.
`delete_object` | Delete an object.
> [!CAUTION]
> Five of these tools destroy or overwrite data: `delete_bucket`,
> `delete_object`, `move_object` (deletes the source), `write_object`, and
> `upload_object`. Your agent must ask for explicit permission before calling
> them.
For a comparison with the Google-hosted remote Cloud Storage MCP server, which
supports Model Armor screening and IAM deny policies, see the
[MCP usage reference](./skills/google-cloud-storage-basics/references/mcp-usage.md).
If your agent reports `spawn npx ENOENT`, Node.js is not installed or `npx` is
not on your `PATH`.
### Cloud Storage OKF MCP Server
Installing the plugin also registers a second local server,
`google-cloud-storage-okf`, which serves the `peek`, `list_regions`, `expand`,
and `write_okf` tools used by the
[Google Cloud Storage OKF skill](./skills/google-cloud-storage-okf/). It is a
Python script bundled with the skill and launched through
[`uv`](https://docs.astral.sh/uv/), which installs its dependencies (`mcp`,
`google-cloud-storage`) into an isolated environment on first start, so there is
nothing to `pip install`, but `uv` must be installed.
It uses the same Application Default Credentials and `CLOUD_STORAGE_PROJECT`
setting as the Toolbox server, and needs `storage.objects.get` to read objects
and `storage.objects.update` to write OKF indices.
If your agent reports `spawn uv ENOENT`, `uv` is not installed or not on your
`PATH`.
## Prerequisites
Ensure you have the following:
* **A Google Cloud project** with the resources you want to work with.
* **Google Cloud SDK (gcloud CLI):**
[Install and initialize](https://cloud.google.com/sdk/docs/install) the
gcloud CLI and ensure
[Application Default Credentials (ADC)](https://cloud.google.com/docs/authentication/provide-credentials-adc)
are configured.
* **A compatible coding agent**, such as Antigravity CLI, Claude Code, Codex,
or Gemini CLI.
* **[Node.js](https://nodejs.org/)**: the
[Cloud Storage MCP server](#cloud-storage-mcp-server) runs via `npx`.
* **[uv](https://docs.astral.sh/uv/getting-started/installation/)**: the
[Cloud Storage OKF MCP server](#cloud-storage-okf-mcp-server) runs via `uv
run`.
## Authentication
Before using the skills, authenticate with Google Cloud so your agent can read
your storage resources and run any changes you approve. It is recommended to run
**both** of the following commands:
```bash
gcloud auth login
gcloud auth application-default login
```
* **`gcloud auth application-default login`** is **required**: skill scripts
use Application Default Credentials (ADC) to generate access tokens for GCP
API calls.
* **`gcloud auth login`** allows the agent (or you) to run standard `gcloud`
commands to explore configurations or dig deeper into specific resources
beyond what the skill scripts cover.
## Additional Setup: GCS Security Assessment
The GCS Security Assessment skill runs with nothing more than working
Application Default Credentials (see [Authentication](#authentication)) — there
is no required IAM permission. However, signals the skill cannot read are
reported as `UNKNOWN`, so for a complete assessment grant the recommended
**read-only** roles covering Storage Insights telemetry (bucket/object analysis)
and project-level posture (IAM and audit config, org policies, VPC Service
Controls, and Model Armor). See **[PERMISSIONS.md](./PERMISSIONS.md)** for the
full permission tables and a ready-to-apply custom IAM role
([`gcs-security-assessment-role.yaml`](./gcs-security-assessment-role.yaml)).
> [!TIP]
> For the best analysis, we highly recommend being a
> [Storage Intelligence](https://docs.cloud.google.com/storage/docs/storage-intelligence/overview)
> customer. When Storage Intelligence is enabled, the skill can query your
> Storage Insights datasets to perform deep, bucket-level and object-level
> assessments. Without it, the skill falls back to a project-level assessment
> only.
The other skills need no permissions beyond the [prerequisites](#prerequisites)
and whatever IAM access your identity already has to the buckets you work with.
## Example Use Cases
The skills cover the full storage lifecycle — provisioning, data movement,
file-system access, access control, troubleshooting, protection and compliance,
cost, security, and automation. Interact with Google Cloud Storage using natural
language, right from your coding agent:
### Design and provision storage for any workload
* **Quick start:** "Create a new GCS bucket named 'audio-video-assets' in the
'my-gcp-project' project"
* **Sensitive data:** "Create a secure GCS bucket to store PII and other
sensitive data. Make sure the data is protected against exfiltration and
unauthorized public exposure"
* **Media serving:** "I am building a high-performance media streaming service
that delivers millions of high-definition images and videos to a global
audience. Set up a Cloud Storage bucket as the origin, paired with a global
Content Delivery Network (CDN), to minimize latency and ensure optimal
streaming performance at scale"
* **AI/ML workloads:** "I have a large-scale model training and checkpointing
use case. Help me set up GCS to optimize performance"
### Mount buckets as a file system
* **Workload fit:** "Should my ML training workload use Cloud Storage FUSE,
native gs:// reads, or Filestore? It reads millions of small files every
epoch"
* **Tuned mounts:** "Help me mount the 'ml-datasets' bucket as a local file
system with gcsfuse, with mount options tuned for high-throughput model
training"
* **GKE deployment:** "Deploy a gcsfuse mount on my GKE training cluster with
the CSI driver, with caches sized for repeated reads of the training
dataset"
* **Performance diagnosis:** "My training job reads from a gcsfuse mount and
GPU utilization is low. Diagnose whether the mount is the bottleneck and
tune it"
* **Cost diagnosis:** "My GCS bill spiked after we moved to gcsfuse. Figure
out which mount options are causing the excess operations"
### Search and navigate large objects with Open Knowledge Format (OKF)
* **Token-efficient search:** "Find the root cause of the outage in the 500MB
server log gs://my-bucket/logs/incident-2026-08.log without downloading the
entire file"
* **Semantic document navigation:** "Navigate the regulatory filing
gs://reports-bucket/annual-filing-2025.pdf using OKF to extract Section 4.2"
* **Generate OKF Table of Contents:** "Generate an OKF Table of Contents index
with semantic regions and write it to the GCS object context for
gs://archive-bucket/quarterly-dump.json"
### Move, replicate, and migrate data at scale
* **Cloud migration:** "Migrate the data in my S3 bucket 'legacy-exports' into
a new GCS bucket"
* **Disaster recovery:** "Set up continuous replication of bucket 'ops-bucket'
to bucket 'vault-bucket-isolated', and ensure all the existing historical
data is copied as well"
* **Zero-downtime moves:** "Relocate my 'analytics-archive' bucket from
us-east1 to us-central1 without downtime"
### Control who can access your data
* **Temporary sharing:** "How can I temporarily give one of my users access to
upload a large video to my bucket?"
* **Least privilege:** "Give the analytics team read-only access to the
'reports' bucket without granting them anything else in the project"
### Diagnose and fix access errors
* **403 Permission Denied:** "User alice@example.com is getting a 403
Permission Denied when trying to list objects in gs://my-team-bucket. Help
me diagnose and fix it"
* **Confusing denials:** "Diagnose why reading gs://data-bucket/object.txt
fails even though I have object viewer permissions"
* **IP filtering lockout:** "I am getting a 403 error on gs://my-secure-bucket
due to IP filtering restrictions"
* **Service agents:** "Pub/Sub notifications on my bucket stopped working
after we enabled CMEK. Check whether the service agents have the right
permissions"
### Protect data and meet compliance requirements
* **Recovery:** "I accidentally deleted objects from the 'prod-reports'
bucket. Can I get them back?"
* **Immutability:** "Configure my 'audit-logs' bucket so objects cannot be
deleted or modified for 7 years"
### Optimize storage costs
* **Cost analysis:** "Analyze my buckets and recommend storage classes and
lifecycle rules to reduce storage costs"
* **Usage insight:** "Find my largest and least-accessed datasets across all
buckets in the project"
### Assess and harden your security posture
* **Targeted assessment:** "Assess the security posture of buckets [BUCKET_1],
[BUCKET_2] in project [PROJECT_ID]"
* **Project-wide assessment:** "Run a security assessment of project
[PROJECT_ID] and show me the exact commands to remediate any toxic
combinations you find"
* **Follow-up investigation:** "Explain why the 'ml-training-data' bucket is
flagged as a toxic combination, and show me the exact command to remediate
the public access finding"
### Generate infrastructure and application code
* **Terraform:** "Generate a Terraform configuration to provision a GCS bucket
in us-central1 for application logs. Make sure public access is prevented
and Uniform Bucket-Level Access is enabled, and add a lifecycle rule to
transition logs to Nearline storage after 30 days and delete them after 365
days"
* **Client libraries:** "Generate Java code to upload a local directory to my
'app-backups' bucket in parallel using the Cloud Storage client library"
### Set up and secure Cloud Storage MCP servers
* **Guarded setup:** "Set up the Cloud Storage MCP server for my coding agent,
and integrate Model Armor with it to screen tool calls for prompt injection"
* **Authentication and tools:** "How do I authenticate and authorize with the
remote Cloud Storage MCP server, and what tools are available on it?"
* **Choosing a server:** "For downloading large files from my buckets, which
Cloud Storage MCP server should I use?"
* **Read-only enforcement:** "Lock down the Cloud Storage remote MCP server
with an IAM deny policy so my agent can only call read-only tools"
### Build event-driven and AI-powered workflows
* **Event notifications:** "Send a Pub/Sub notification whenever new objects
land in my 'ingest' bucket so my pipeline can process them"
* **Agentic workflows:** "Scan the 'retail-raw-products' bucket for assets
related to 'ProductX', draft a promotional social media campaign listing,
and write the draft output file to bucket 'retail-campaigns'"
## Security Reminder: Agent Environment Hardening
Your agent can execute tools and commands on your behalf. Protect your Google
Cloud resources by enforcing **The Principle of Least Privilege** across all
CLIs, MCP servers and other resources available to your agents.
* **Service Accounts:** Use
[service accounts](https://docs.cloud.google.com/docs/authentication/use-service-account-impersonation)
instead of end user credentials to access Google Cloud resources.
* **Limited Permissions:** Assign roles with
[limited permissions](https://docs.cloud.google.com/iam/docs/roles-overview)
to the service account that you're using for authentication.
* **Principal Access Boundaries:** Prevent unwanted cross-org agent access by
using
[Principal Access Boundary policies](https://docs.cloud.google.com/iam/docs/principal-access-boundary-policies#use-case-one-project)
to scope your agent to projects you intend it to access.
* [Include a condition in the policy binding](https://docs.cloud.google.com/iam/docs/principal-access-boundary-policies#use-case-one-project)
to ensure that the policy only applies to the service accounts that you
intend to restrict.
You can read more
[here](https://docs.cloud.google.com/data-cloud-extension/vs-code/prompt-injection-risk)
on how to mitigate prompt injection attacks with Google Cloud MCP.
## Support
If you need help or encounter issues with this plugin, search for existing
issues or open a new one in the
[GitHub Issue Tracker](https://github.com/GoogleCloudPlatform/google-cloud-storage-plugin/issues).
## Contributing
We welcome contributions to improve this plugin. You can help by:
* [Reporting bugs or inaccuracies](https://github.com/GoogleCloudPlatform/google-cloud-storage-plugin/issues)
in the skill files.
* Suggesting new skills to add to this plugin by filing a feature request.
## License
You are free to copy, modify, and distribute this plugin under the terms of the
Apache 2.0 license. See the `LICENSE` file for details.