Gcs

com.pulsemcp由 pulsemcp.com 发布v0.1.14更新于 Oct 24, 2026

MCP server for Google Cloud Storage with fine-grained tool access control.

概览

AI 生成的概览

面向 Google Cloud Storage 的本地 MCP 服务器,让助手列出、读取、上传、复制和删除存储桶与对象,并可用工具组限制权限。

可以做什么
列出并查看存储桶和对象,检查存储桶是否存在及可访问,读取对象文本内容,或将对象及整个前缀以原始字节下载到本地磁盘。以内联文本、单个本地文件或整个目录树上传或更新对象,并在同一或不同存储桶间复制对象。创建存储桶,删除对象或空存储桶,具体取决于启用的工具组。通过工具组(readonly、readwrite、delete)或单独启用/禁用列表控制可用工具,并可将所有操作限制在单个存储桶内。
适用场景
适合助手需要处理 Google Cloud Storage 中文件的场景:浏览和读取已存数据、把本地文件或文件夹移入存储桶、复制对象,或清理对象和存储桶。只读组适合探索与分析,单存储桶模式适合最小权限配置。
运行要求
通过 npx 在本地运行(需要 Node.js)。需要 GCS_PROJECT_ID 和 Google Cloud 凭据:通过 GCS_SERVICE_ACCOUNT_KEY_FILE 提供服务账号密钥文件、通过 GCS_SERVICE_ACCOUNT_KEY_JSON 提供内联密钥 JSON,或使用 gcloud 的 Application Default Credentials。服务账号需具备所启用工具对应的存储 IAM 权限。需要访问 Google Cloud Storage 的网络。
安装前请注意
服务账号密钥是敏感凭据,内联 JSON 方式会把它写入配置。写入和删除工具会更改或移除真实云端数据,delete_bucket 会删除存储桶。默认启用全部工具组(含删除),建议用 GCS_ENABLED_TOOLGROUPS 或启用/禁用列表加以限制。SKIP_HEALTH_CHECKS=true 会关闭所有凭据校验。上传工具会跟随文件符号链接,可能把本地文件发送到 Google Cloud。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Gcs,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

GCS MCP Server

MCP server for Google Cloud Storage operations with fine-grained tool access control. Supports read/write operations on GCS buckets and objects, with configurable tool groups for security and permission management.

Highlights

  • Full GCS bucket and object management (list, get, put, copy, delete)
  • Fine-grained access control with tool groups (readonly, readwrite, delete)
  • Individual tool enable/disable via environment variables
  • Multiple authentication methods (service account key file, inline JSON, ADC)
  • GCS credential validation with health checks
  • Single bucket constraint mode
  • TypeScript with strict type checking
  • Comprehensive testing (functional, integration, manual)

Capabilities

Tools

ToolGroupDescription
list_bucketsreadonlyList all GCS buckets in the Google Cloud project
list_objectsreadonlyList objects in a bucket with prefix and pagination
get_objectreadonlyGet object contents as text
download_objectreadonlyDownload a single object to a local file (binary-safe)
download_prefixreadonlyRecursively download a prefix to a local directory (binary-safe)
head_bucketreadonlyCheck if a bucket exists and is accessible
put_objectreadwriteUpload or update an object from inline string content
put_object_from_pathreadwriteUpload a single local file by streaming from disk (binary-safe, context-free)
upload_prefixreadwriteRecursively upload a local directory tree from disk (binary-safe, context-free)
copy_objectreadwriteCopy an object within or across buckets
create_bucketreadwriteCreate a new GCS bucket
delete_objectdeleteDelete an object from a bucket
delete_bucketdeleteDelete an empty GCS bucket

Downloading to Local Disk

get_object returns an object's contents inline as UTF-8 text, which is lossy for binary data and impractical for large files or many files at once. For local data-wrangling, use the download tools instead — both stream raw bytes to disk and belong to the readonly toolgroup:

  • download_object — download a single object to a local file path. Defaults to a unique file under the OS temp directory.

  • download_prefix — recursively download every object under a prefix to a local directory, preserving the key path structure as subdirectories. Paginates through the full listing, skips directory-placeholder objects (keys ending in /), collects per-object errors without aborting the batch, and returns a manifest:

    json
    {  "destinationDir": "/tmp/gcs-download-my-bucket-1700000000000",  "objectCount": 1234,  "totalBytes": 5678901,  "files": [    { "key": "logs/2024/01/data.json", "localPath": "/tmp/.../01/data.json", "size": 1234 }  ],  "filesTruncated": true,  "errors": []}

    The inline files list is capped (maxInlineEntries, default 100), but objectCount and totalBytes always reflect the full download. When destinationDir is omitted it defaults to a unique folder under the OS temp directory.

Uploading from Local Disk

put_object takes the object content inline as a string argument, which routes every byte through the model context and requires base64 for binary data — impractical for large files or bulk uploads. For uploading files or folders without consuming context, use the local-path upload tools instead. Both stream raw bytes directly from disk to GCS server-side, returning only metadata. They belong to the readwrite toolgroup, so read-only deployments never gain them.

  • put_object_from_path — upload a single local file by streaming it from a filesystem path. Content type is auto-detected from the file extension (override with contentType). Binary files (images, archives) work without base64. Returns { bucket, key, sourcePath, size, etag, generation }.

  • upload_prefix — recursively upload every file under a local directory, preserving the directory structure as key paths beneath a destination prefix. Symbolic links to files are followed and uploaded (links to directories are skipped to avoid cycles), per-file failures are collected without aborting the batch, and it returns a manifest:

    json
    {  "bucket": "my-bucket",  "sourceDir": "/tmp/exports",  "destPrefix": "uploads/",  "objectCount": 1234,  "totalBytes": 5678901,  "files": [    { "localPath": "/tmp/exports/01/data.json", "key": "uploads/01/data.json", "size": 1234 }  ],  "filesTruncated": true,  "errors": []}

    The inline files list is capped (maxInlineEntries, default 100), but objectCount and totalBytes always reflect the full upload. When destPrefix is omitted, files upload to the bucket root.

Resources

ResourceDescription
gcs://configServer configuration and status (for debugging)

Tool Groups

Control which tools are available via the GCS_ENABLED_TOOLGROUPS environment variable:

GroupDescription
readonlyRead-only operations (list, get, head)
readwriteNon-destructive write operations (put, copy, create)
deleteDelete operations (delete_object, delete_bucket)

Examples:

  • GCS_ENABLED_TOOLGROUPS="readonly" - Only read operations
  • GCS_ENABLED_TOOLGROUPS="readonly,readwrite" - Read and write, no deletes
  • GCS_ENABLED_TOOLGROUPS="readonly,readwrite,delete" - All operations
  • Not set - All tools enabled (default)

Individual Tool Control

Fine-grained control over specific tools:

  • GCS_ENABLED_TOOLS="list_buckets,get_object" - Only enable these tools
  • GCS_DISABLED_TOOLS="delete_bucket,delete_object" - Disable these tools

Single Bucket Mode

Constrain all operations to a specific bucket using GCS_BUCKET:

bash
GCS_BUCKET="my-bucket"

When set:

  • All object operations are automatically scoped to this bucket
  • Bucket-level tools (list_buckets, create_bucket, delete_bucket, head_bucket) are hidden
  • The bucket parameter is automatically injected and hidden from tool inputs
  • For copy_object, both source and destination are constrained to the specified bucket

This lets you grant a least-privilege service account that has only object-level access to one bucket, without project-level bucket-listing permission. The startup healthcheck honors this: when GCS_BUCKET is set it validates credentials with an object-scoped probe (storage.objects.list on that bucket) and never calls the project-level bucket list or a bucket-metadata probe. See Required IAM permissions for the exact permissions per mode.

Required IAM permissions

The minimal IAM permissions depend on whether the server is constrained to a single bucket (GCS_BUCKET) and which tool groups you enable. The startup healthcheck validates credentials within these bounds — it does not require any permission beyond what the table below lists for your mode.

Startup healthcheck

ModePermission the healthcheck needsNotes
Single bucket (GCS_BUCKET set)storage.objects.list on the constrained bucketProbes list_objects with maxResults: 1. Does not require project-level storage.buckets.list or bucket-metadata storage.buckets.get, so a correctly least-privilege, object-only service account can start.
Unconstrained (no GCS_BUCKET)project-level storage.buckets.listValidates by listing buckets in the project.

Set SKIP_HEALTH_CHECKS=true to bypass startup validation entirely (disables all credential checks — not recommended).

Per-tool permissions

Beyond the healthcheck, each tool needs the corresponding GCS permission at runtime. Grant only what the tool groups you enable require:

Tool / groupPermission
list_objects (readonly)storage.objects.list
get_object, download_object (readonly)storage.objects.get
download_prefix (readonly)storage.objects.list, storage.objects.get
list_buckets (readonly)storage.buckets.list
head_bucket (readonly)storage.buckets.get
put_object, put_object_from_path, upload_prefix, copy_object (readwrite)storage.objects.create, storage.objects.get (these tools read each object's metadata back after writing to return its etag/generation; copy_object additionally reads the copy source)
create_bucket (readwrite)storage.buckets.create
delete_object (delete)storage.objects.delete
delete_bucket (delete)storage.buckets.delete

For a read-only, single-bucket service account, the roles/storage.objectViewer role on the bucket (or a custom role with storage.objects.list + storage.objects.get) is sufficient — no project-level or bucket-metadata permissions are needed.

Quick Start

Configuration

Environment Variables
VariableRequiredDescriptionDefault
GCS_PROJECT_IDYesGoogle Cloud project ID-
GCS_SERVICE_ACCOUNT_KEY_FILENoPath to service account key JSON file-
GCS_SERVICE_ACCOUNT_KEY_JSONNoService account key JSON contents (inline)-
GCS_BUCKETNoConstrain operations to single bucket-
GCS_ENABLED_TOOLGROUPSNoComma-separated tool groupsAll enabled
GCS_ENABLED_TOOLSNoSpecific tools to enable-
GCS_DISABLED_TOOLSNoSpecific tools to disable-
SKIP_HEALTH_CHECKSNoSkip credential validationfalse
Authentication

The server supports three authentication methods (in order of precedence):

  1. Service Account Key File: Set GCS_SERVICE_ACCOUNT_KEY_FILE to the path of a JSON key file
  2. Inline Service Account Key: Set GCS_SERVICE_ACCOUNT_KEY_JSON with the JSON contents directly
  3. Application Default Credentials (ADC): No additional env vars needed - uses gcloud auth application-default login

Claude Desktop Configuration

If this is your first time using MCP Servers, make sure you have the Claude Desktop application and follow the official MCP setup instructions.

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

json
{  "mcpServers": {    "gcs": {      "command": "npx",      "args": ["-y", "gcs-google-mcp-server"],      "env": {        "GCS_PROJECT_ID": "your-project-id",        "GCS_SERVICE_ACCOUNT_KEY_FILE": "/path/to/service-account-key.json",        "GCS_ENABLED_TOOLGROUPS": "readonly"      }    }  }}

Restart Claude Desktop and you should be ready to go!

Read-Only Mode

For safer exploration, enable only read operations:

json
{  "env": {    "GCS_ENABLED_TOOLGROUPS": "readonly"  }}

Using Application Default Credentials

If you have gcloud CLI configured locally:

bash
gcloud auth application-default login

Then configure without key file:

json
{  "env": {    "GCS_PROJECT_ID": "your-project-id"  }}

Development

Install Dependencies

bash
npm run install-all

Build

bash
npm run build

Running in Development Mode

bash
npm run dev

Testing

bash
# Run functional testsnpm run test:run
# Run integration tests (full MCP protocol)npm run test:integration
# Run manual tests (real GCS - requires .env)npm run test:manual:setup  # First time onlynpm run test:manual
# Run all automated testsnpm run test:all

Manual Testing Setup

Create a .env file in the gcs directory:

bash
GCS_PROJECT_ID=your-project-idGCS_SERVICE_ACCOUNT_KEY_FILE=/path/to/service-account-key.json

Then run:

bash
npm run test:manual

Project Structure

gcs/├── local/                 # Local server implementation│   ├── src/│   │   ├── index.ts      # Main entry point with env validation│   │   └── index.integration-with-mock.ts│   └── package.json├── shared/               # Shared business logic│   ├── src/│   │   ├── server.ts     # Server factory with DI│   │   ├── tools.ts      # Tool registration with grouping│   │   ├── tools/        # Individual tool implementations│   │   ├── resources.ts  # Resource implementations│   │   ├── gcs-client/   # Google Cloud Storage client wrapper│   │   └── logging.ts│   └── package.json├── tests/                # Test suite│   ├── functional/       # Unit tests with mocks│   ├── integration/      # MCP protocol tests│   ├── manual/          # Real API tests│   └── mocks/           # Mock implementations└── package.json         # Root workspace config

License

MIT

来源:experimental/gcs/README.md,提交 0ed34de

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v0.1.14最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分