workspace-admin-mcp
An admin-oriented Model Context Protocol server for Google Workspace that lets LLMs perform directory and user lifecycle operations with safety guardrails and audit logging.
README
workspace-admin-mcp
An admin-oriented Model Context Protocol server for Google Workspace. It lets an LLM (Claude, or any MCP client) perform directory/admin operations — user lifecycle and group management — through natural language, with safety guardrails and full audit logging.
Most Workspace MCP servers are end-user focused (my Gmail, my Drive). This one is built for the admin — the person who onboards, offboards, and manages groups for the whole domain via the Admin SDK.
Why
Running Workspace admin at scale means living in the Admin SDK and tools like GAM: bulk operations, offboarding runbooks, group cleanups. Those are exactly the tasks an LLM is good at describing — but you can't hand an assistant the keys to suspend accounts without guardrails. This server is the safe middle layer:
- Read-only by default. Write tools refuse unless you explicitly opt in.
- Dry-run by default. Even with writes enabled, actions are simulated until you pass
dry_run=false. - Audited. Every call — read, refused, simulated, or applied — emits a structured JSON audit line.
Tools
| Tool | Type | Description |
|---|---|---|
workspace_list_users |
read | List users; supports Admin SDK search (isSuspended=true, orgUnitPath=/Sales, …) |
workspace_get_user |
read | Get one user's directory profile |
workspace_list_groups |
read | List groups, or the groups a user belongs to |
workspace_list_group_members |
read | List a group's members |
workspace_suspend_user |
guarded write | Suspend/restore a user (core offboarding step) |
workspace_add_group_member |
guarded write | Add a member to a group |
workspace_remove_group_member |
guarded write | Remove a member from a group |
Architecture
Claude / MCP client ──MCP──▶ workspace-admin-mcp ──▶ Google Admin SDK ──▶ Workspace
│
guardrails ──▶ audit log (JSON, stderr)
Full diagram and design notes in docs/architecture.md.
Quickstart (local)
git clone https://github.com/pizzacrustreviewer/GWS-MCP.git
cd workspace-admin-mcp
pip install -e ".[dev]"
pytest -q # tests run with no credentials (in-memory fake client)
Run it against a real tenant
You need a Google Cloud service account with domain-wide delegation authorized for the directory scopes, and a Workspace admin to impersonate. Use a personal or developer tenant — never production/corporate.
cp .env.example .env # fill in the values, then:
export $(grep -v '^#' .env | xargs)
workspace-admin-mcp # starts on stdio for a local MCP client
Connect from Claude Desktop
{
"mcpServers": {
"workspace-admin": {
"command": "workspace-admin-mcp",
"env": {
"GOOGLE_APPLICATION_CREDENTIALS": "/absolute/path/service-account.json",
"WORKSPACE_ADMIN_SUBJECT": "admin@yourdomain.com",
"WORKSPACE_READ_ONLY": "true"
}
}
}
}
Configuration
| Env var | Default | Meaning |
|---|---|---|
GOOGLE_APPLICATION_CREDENTIALS |
— | Path to the DWD service-account key JSON |
WORKSPACE_ADMIN_SUBJECT |
— | Admin user to impersonate |
WORKSPACE_CUSTOMER_ID |
my_customer |
Resolves to the admin's own domain |
WORKSPACE_READ_ONLY |
true |
When true, all writes are refused |
WORKSPACE_DRY_RUN |
true |
When true, enabled writes are simulated |
MCP_TRANSPORT |
stdio |
stdio locally, streamable-http on Cloud Run |
Deploy (optional)
The repo ships a Dockerfile and reference Terraform for Cloud Run (private ingress, least-privilege runtime SA, key mounted from Secret Manager). See terraform/. This is reference IaC — it only runs if you apply it against a project you own.
docker build -t workspace-admin-mcp .
Security
- Writes off by default; dry-run by default; explicit opt-in for both.
- Least-privilege scopes (read scopes only unless writes are enabled).
- No credentials in the image or repo;
.gitignoreblocks key files. - Structured audit trail suitable for shipping to Cloud Logging.
Roadmap & limitations
This is v0.1 — the logic is complete and unit-tested; the live Admin SDK layer, pagination, and HTTP-transport auth are the next steps. See ROADMAP.md.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。