AOJ Workbench
Private remote HTTP MCP server that provides a unified endpoint to interact with Google Workspace services (Drive, Sheets, Gmail, Calendar, Docs, Forms) and other toolkits through the Composio provider layer.
README
<div align="center">

AOJ Workbench
One private MCP endpoint for your knowledge-work toolkits.
Give an AI agent one URL and it can work through AOJ Workbench.
Quick Start . Provider Model . Docs
</div>
Remote Only
AOJ Workbench is a private remote HTTP MCP server hosted on Cloudflare Workers.
The supported client experience is:
https://your-worker.workers.dev/mcp
There is no supported local stdio server, local HTTP server, Docker runtime, or local bootstrap flow for MCP clients. Clients connect to the remote /mcp endpoint by URL.
Legacy local commands and Docker instructions from older versions are not part of the supported runtime model.
Why This Exists
AI agents can draft plans, reason over code, and summarize documents, but daily work spans many tools:
- Gmail and Google Workspace
- Outlook
- Notion
- Stripe
- YouTube
- other selected Composio-supported toolkits
AOJ Workbench gives agents one MCP endpoint. The target provider layer is the Composio SDK with managed auth. The current direct Google implementation is legacy migration scaffolding until provider replacement slices remove it.
Quick Start
1. Get The Endpoint
Use the deployed Worker URL:
https://your-worker.workers.dev/mcp
The Worker must already be deployed and configured with its MCP bearer token and the required Composio runtime secrets. Provider accounts are connected in Composio for the configured AOJ Workbench user.
2. Connect Claude Code
Use a descriptive MCP server name so agents understand the full surface area:
claude mcp add --scope user --transport http aoj-workbench https://your-worker.workers.dev/mcp
If bearer auth is enabled for the Worker, configure the same bearer token expected by the server.
3. Connect Codex
export AOJ_WORKBENCH_MCP_TOKEN="replace-with-worker-token"
codex mcp add aoj-workbench \
--url https://your-worker.workers.dev/mcp \
--bearer-token-env-var AOJ_WORKBENCH_MCP_TOKEN
Start a fresh Codex session after changing MCP config.
Provider Model
AOJ Workbench keeps one public MCP connector. Behind that connector, Composio SDK sessions discover and execute selected provider toolkit actions.
The direct Google Workspace modules that remain in the repo are legacy migration scaffolding. The supported runtime routes provider discovery and execution through the Composio native provider layer. Current AOJ Workbench capabilities include:
| Service | Operations | Highlights |
|---|---|---|
| Drive | 7 | Search, enhanced search, read, create, update, batch operations |
| Sheets | 12 | Read/write cells, formulas, formatting, conditional formatting, freeze, column width |
| Gmail | 30+ | List, search, read, draft, send emails, manage labels, and work with attachments |
| Calendar | 9 | Full CRUD, natural language quickAdd, free/busy checks |
| Docs | 5 | Create, insert text, replace, rich text styling, insert tables |
| Forms | 4 | Create forms, add questions, read responses |
| 70+ total |
Live Forms execution requires an active Google Forms-capable connected account in Composio for AOJ_WORKBENCH_USER_ID.
Tools
v4 exposes two MCP tools:
| Tool | Purpose |
|---|---|
search |
Discover provider/toolkit capabilities, schemas, auth status, and execution guidance |
execute |
Run a selected provider toolkit operation |
This keeps the MCP tool list small while avoiding separate client connectors for every provider.
Example direct operation call:
{
"tool": "execute",
"args": {
"service": "gmail",
"operation": "sendMessage",
"args": {
"to": "team@example.com",
"subject": "Q1 report ready",
"body": "The Q1 report is ready for review."
}
}
}
Example discovery call:
{
"tool": "search",
"args": {
"service": "sheets",
"operation": "updateCells"
}
}
Architecture
MCP client
|
| Streamable HTTP
v
Cloudflare Worker /mcp
|
| Composio SDK sessions and managed auth
v
Provider toolkits
Persistent state belongs in Cloudflare services, primarily Workers KV. The user-facing MCP runtime is the Worker URL.
Security
- MCP clients authenticate to the Worker with static bearer auth when configured.
- Provider authorization belongs to Composio managed auth in the target architecture.
- This server is not an OAuth authorization server for MCP clients.
- Do not point MCP client auth metadata at provider OAuth.
Development
Development commands are for contributors changing the Worker implementation, not for running the MCP server locally as a client dependency.
npm test
npm run type-check
npm run build:worker
npm run deploy:worker
Do not add local stdio, local HTTP, or Docker connection instructions back to user-facing docs.
Roadmap
- [ ] Composio native provider — discovery, auth status, schema lookup, execution
- [ ] Google provider replacement — migrate current direct Google capabilities behind Composio
- [ ] Notion toolkit — selected private-workbench workflows
- [ ] Stripe toolkit — selected private-workbench workflows
- [ ] YouTube toolkit — selected private-workbench workflows
- [ ] Outlook toolkit — selected private-workbench workflows
- [x] Cloudflare Workers — remote HTTP MCP runtime
Contributing
This repo is TypeScript and Cloudflare Workers based. Keep changes aligned with the remote-only runtime boundary.
- Fork and create a feature branch.
- Follow existing TypeScript and ESLint conventions.
- Add tests for behavior changes.
- Run the relevant checks.
- Open a PR with a clear description.
License
MIT — see LICENSE for details.
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。