colab-cloud-mcp
Runs the MCP server inside Google Colab GPU runtimes, providing tools to inspect runtime/GPU status, execute Python in a persistent namespace, manage background jobs, and export executed code and plots as a notebook.
Provides file management tools for browsing, reading, and writing files in the Colab workspace and mounted Google Drive, including saving logs and model weights to Drive.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
colab-cloud-mcp
让 Claude 云端对话(网页 / 手机 App / Desktop Custom Connectors)、Google Antigravity 以及 Claude Code 通过公网 HTTPS (443) Streamable HTTP 协议直接连接 Google Colab GPU 运行时的远程 MCP Server。
为什么需要这个项目?
本地 stdio 模式的 Colab MCP 依赖本机 CLI 进程与本机浏览器标签页桥接——一旦合上笔记本电脑、休眠或本地网络闪断,连接就会立即断开,且无法在网页版或手机端云对话中使用。
colab-cloud-mcp 将 MCP 服务端直接运行在 Google Colab 虚拟机内部,并通过 HTTPS 隧道(Cloudflare Tunnel 或 ngrok)暴露符合最新规范的 Streamable HTTP 端点:
真正脱离本机:云端对话客户端直接访问 Colab 端服务,本机合盖、断网或切换到手机端对话均不断连。
持久化 Python 命名空间 + 自动捕获图表:
execute_python跨轮次保留所有变量、PyTorch 模型与显存张量;Matplotlib 绘图自动转为 PNG 图像块回传给对话界面。异步后台长任务管理:针对云端单次工具调用超时限制(Claude 240s / Antigravity 默认 180s),内置
start_background_job/get_job_status/cancel_job,支持数小时模型训练或大权重下载至 Google Drive,随时在对话中拉取最新日志。会话导出为
.ipynb:随时调用export_session_notebook将执行过的所有代码与输出图像保存为标准 Jupyter Notebook。三重鉴权兼容所有客户端:
零请求头加密路径模式(
<TUNNEL_URL>/s/<MCP_TOKEN>/mcp):适配 Claude Custom Connectors 的No sign in模式,粘贴单条 URL 即可安全鉴权。标准 Bearer Header 模式(
Authorization: Bearer <MCP_TOKEN>):适配 Google Antigravitymcp_config.json、Claude Code HTTP MCP 与 Claude Request Headers。内置 OAuth 2.1 (PKCE + DCR + CIMD):完整实现
/.well-known/oauth-protected-resource、/.well-known/oauth-authorization-server、动态客户端注册与 PKCE 授权码流程。
Related MCP server: Smart Colab MCP Bridge
快速开始(在 Google Colab 中运行)
打开本仓库的 notebooks/colab_cloud_mcp.ipynb 按顺序执行即可启动服务并打印所有客户端接入配置。
合盖前在最后一个单元格运行保活监控:
mcp_server.keep_alive(status_interval_seconds=60)Colab 实例生命周期提示:MCP 连接本身已完全脱离本机,但 Google Colab 对免费版/Pro 版在关闭全部浏览器标签页后仍有后台空闲回收时间窗口;如需合盖后持续运行数小时,建议配合
keep_alive()或 Colab Pro+ 后台执行,并将日志与权重保存至/content/drive/MyDrive。
固定域名配置(重启 Colab 免改客户端 URL)
在 Colab 左侧 Secrets 中配置以下密钥并以 provider="auto" 启动:
方案 | Colab Secrets 变量名 | 说明 |
固定鉴权密钥 |
| 自定义高强度字符串,重启免改 Token |
Cloudflare Named Tunnel |
| 在 Cloudflare Zero Trust Tunnel 将子域名指向本地端口 |
ngrok 免费静态域名 |
| 使用 ngrok 账号免费静态域名(需安装 |
客户端接入指南
1. Claude 云端(Custom Connectors)
进入 Customize -> Connectors -> + Add -> Add custom connector
最简方式(免配 Header):URL 填入启动器打印的
<TUNNEL_URL>/s/<MCP_TOKEN>/mcp,Authentication 选择No sign inHeader / OAuth 方式:URL 填入
<TUNNEL_URL>/mcp,在 Request headers 添加Authorization: Bearer <MCP_TOKEN>或选择 OAuth 授权
2. Google Antigravity(App / IDE / CLI)
编辑 ~/.gemini/config/mcp_config.json 或工作区 .agents/mcp_config.json:
{
"mcpServers": {
"colab-remote": {
"serverUrl": "<TUNNEL_URL>/mcp",
"headers": {
"Authorization": "Bearer <MCP_TOKEN>"
},
"timeoutSeconds": 600
}
}
}3. Claude Code CLI
claude mcp add --transport http colab-remote <TUNNEL_URL>/mcp \
--header "Authorization: Bearer <MCP_TOKEN>"内置 MCP 工具一览
工具名称 | 功能说明 |
| 查看 GPU 型号与显存 ( |
| 在持久化命名空间执行 Python 代码,自动捕获 |
| 清空持久化变量并触发 |
| 执行 Bash 命令(支持工作目录与超时进程组清理) |
| 通过 |
| 异步启动长时间后台训练或大权重下载任务,日志实时落盘 |
| 查询后台任务状态与末尾日志、列出或终止后台任务 |
| 浏览、按行读取或写入 Colab 工作区与 Google Drive 文件 |
| 将当前会话执行过的所有代码与图表导出为 |
本地测试
python3 tests/test_server.pyLicense
MIT
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Remote streamable-HTTP MCP server running on a single Cloudflare Worker. Your assistant gets live Airbnb, Amazon, Booking.com, Google Flights, Maps and Reddit data, social search on X, Instagram and TikTok, the Meta Ad Library, and image/video generation without any keys. Connect your own accounts to let it send WhatsApp or Telegram messages, work an IMAP inbox, manage Meta Ads campaigns and publish to X and LinkedIn. OAuth 2.1 with PKCE; stored credentials are AES-256-GCM encrypted.
Work on your own computers from Claude or ChatGPT: run commands, edit and search files.
Connect ChatGPT and Claude to your Mac, Windows or Linux with secure MCP remote access.
Deploy websites, apps, agents and MCP servers from Claude Code or Codex. OAuth, runs in Taiwan.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to autonomously manage Google Colab GPU sessions, submit and monitor training jobs, and debug/fix issues via an encrypted tunnel without requiring a browser tab.MIT
- FlicenseNot gradedqualityDmaintenanceConnects Claude Desktop to a Google Colab runtime via ngrok for remote Python execution, enabling GPU/CPU code execution with smart timeouts and file management.1-
- FlicenseNot gradedqualityDmaintenanceEnables MCP clients to execute Python code on Google Colab GPU/TPU runtimes via the official Colab CLI, with session, file, and package management capabilities.-
- FlicenseNot gradedqualityCmaintenanceEnables persistent execution of Python and shell commands on Google Colab's cloud GPU via a single background daemon, offering a web dashboard and REST API for AI agents without opening new tabs.-