Build AI agents faster with GCS (Google Cloud Storage) MCP server

Connecting an agent to storage requires robust security and governance. GCS MCP server is built on Google Cloud’s standard identity, observability, and security frameworks:

• Identity-first security: Authentication is handled entirely through Identity and Access Management (IAM) rather than shared keys. This ensures agents can only access data (buckets and objects) explicitly authorized by the user.

• Full observability: To track agent activity, every request and action taken via these MCP servers is logged in Cloud Audit Logs. This provides security teams with a record of every interaction, maintaining visibility alongside ease of access.

• MCP security – content scanning: You can optionally configure the MCP endpoint with Google’s content security service, Google Cloud Model Armor. This allows you to implement security controls against common MCP attack vectors—such as direct and indirect prompt injection attacks, MCP Tool poisoning attacks, and malicious URL/SQL injections—as well as prevent the leakage of sensitive data.

Cloud Storage MCP servers are perfect for most production use cases; however, as with all remote servers, you lose the capability to fully customize your MCP tools.

2. Local MCP Server: Self-managed for controlled customization
While the Remote server handles standard data access, Local MCP is the right choice when you need to build custom tools specific to your business logic. For example, if your agent needs to perform specialized data transformations—such as redacting PII or adding context from another internal system—whenever it reads a file from GCS, a Local MCP server allows you to define those unique capabilities

The GCS Local MCP server is an open-source GitHub repository of Google-maintained tools that provides you with a reliable bridge to your data. Here are a few tips to keep in mind while designing custom tools:

• Provide precise, clear descriptions to minimize incorrect invocations by the models

• Implement model-friendly error handling for models to understand their mistakes and self-correct

The GCS Local MCP is now also a part of the MCP Toolbox for Databases, a single open-source repository containing connectors for major data services such as GCS, BigQuery, AlloyDB, Spanner, and Cloud SQL, making it easier to monitor and manage your data ecosystem. The Toolbox offers simplified development with reduced boilerplate code, enhanced security through OAuth2 and OIDC, and end-to-end observability with OpenTelemetry integration.

Get started

Whether you are optimizing an existing process like Snap or automating workflow creations like Airwallex, your unstructured data is one of your agent’s greatest assets.