MCP Servers

模型上下文协议服务器、框架、SDK 和模板的综合目录。

I
Ide Relay MCP
作者 @andeya

Human feedback layer for AI IDEs

创建于 3/18/2026
更新于 about 4 hours ago
Repository documentation and setup instructions

Relay

Human feedback layer for AI IDEs

中文版

Relay MCP is an open-source MCP tool for AI IDEs such as Cursor IDE. It pauses an agent request, shows a native desktop feedback window, and returns the human response inside the same interaction.

Relay was inspired by interactive-feedback-mcp.

Highlights

  • Cross-platform desktop UI powered by Tauri + Rust + Vue
  • Native feedback window on Windows, macOS, and Linux
  • Two executable entry points that share one core implementation
  • Automatic config bootstrap in the user data directory
  • Drag and drop files into the response box
  • Paste copied file paths directly into the editor
  • Automatic reply rules for unattended runs
  • Timestamped logs written to the same user data directory

Repository Layout

  • src-tauri/ - Rust backend, shared core, MCP server, CLI helper, and Tauri window
  • src/ - Vue user interface
  • mcp.json - MCP client configuration example

Build

Install the frontend dependencies and build the web assets:

npm install
npm run build

Build the Rust binaries:

cd src-tauri
cargo build --bins

Run the desktop app in development mode:

npm run tauri dev

Binaries

The workspace produces two user-facing executables and one desktop app:

  • relay-server - MCP server that AI IDEs connect to
  • relay - command-line helper that launches the same UI
  • Relay - the packaged desktop feedback app

On Windows, the binary names automatically receive the .exe suffix.

MCP Configuration

Point your AI IDE to the built relay-server binary. Example:

{
  "mcpServers": {
    "relay-mcp": {
      "command": "/absolute/path/to/relay-server",
      "args": [],
      "timeout": 6000,
      "autoApprove": ["interactive_feedback"]
    }
  }
}

Use the equivalent Windows path ending in .exe on Windows.

Configuration Storage

Relay creates and manages its auto-reply files in your user data directory on first launch. You never need to choose or pass a path manually.

Typical locations are:

  • macOS: ~/Library/Application Support/relay-mcp/
  • Linux: ~/.config/relay-mcp/
  • Windows: %APPDATA%\\relay-mcp\\

The directory contains:

  • auto_reply_oneshot.txt
  • auto_reply_loop.txt
  • feedback_log.txt

Existing auto-reply files from the legacy installation directory are migrated automatically when possible.

Each non-comment line uses this format:

timeout_seconds|reply_text

auto_reply_oneshot.txt consumes the first matching rule and deletes it after use. auto_reply_loop.txt cycles through rules in order.

CLI Usage

Launch the desktop UI directly without MCP:

relay "Work summary" 600

The feedback is printed to stdout. If the window times out or closes empty, the command prints an empty line.

How It Works

  1. An AI IDE calls interactive_feedback.
  2. The Rust server checks auto-reply rules first.
  3. Otherwise it launches relay-gui with the summary and temporary file paths.
  4. The Vue window reads its launch state, watches the control file for timeout or cancellation, and writes feedback to a temp file on submit.
  5. The server waits for the GUI process to exit, reads the result file, and returns the text to the AI IDE.

Contributing

Contributions, bug reports, and platform-specific feedback are welcome. Please keep changes aligned with the existing MCP contract, the Relay brand, and the cross-platform behavior.

License

MIT

快速设置
此服务器的安装指南

安装命令 (包未发布)

git clone https://github.com/andeya/ide-relay-mcp
手动安装: 请查看 README 获取详细的设置说明和所需的其他依赖项。

Cursor 配置 (mcp.json)

{ "mcpServers": { "andeya-ide-relay-mcp": { "command": "git", "args": [ "clone", "https://github.com/andeya/ide-relay-mcp" ] } } }