Nest Llm Aigent

🚀 nest-llm-aigent

With the introduction of the MCP (Model Context Protocol), while unifying the protocol, it provides conditions for large model extension programs (MCP servers) to achieve one-time development and run anywhere. However, due to the natural focus of MCP on the CS architecture, it is difficult for us to conveniently integrate it into the company's existing business. Based on this background, this forwarding solution is designed, which not only solves this problem but also can be understood as an AI agent (Agent), supporting the configuration of multiple MCP servers and providing a unified HTTP service for large models externally.

🚀 Quick Start

Background

With the launch of the MCP (Model Context Protocol), while standardizing the protocol, it enables large model extension programs (MCP servers) to be developed once and run anywhere. However, since MCP inherently leans towards the CS architecture, integrating it into our company's existing business has become a challenge. To address this, I've designed this forwarding solution. This project can be seen as an AI agent (Agent), supporting the configuration of multiple MCP servers and offering a unified HTTP service for large models.

The specific architecture diagram is as follows:

Goals

• Easy Integration: Quickly connect to existing Web services (such as those built on NestJS) through an adaptation layer.
• Seamless Expansion: Package MCP servers as private NPM packages for rapid integration and unified management, ensuring high portability.
• Convenient Deployment: Centralize management and version control of MCP servers via private NPM packages, and integrate them seamlessly into the existing release process.
• Large Model Support: Almost all types of large model calls are supported.

Interface Definitions

To meet the functional requirements of different scenarios, the following standardized interfaces are designed:

Usage Examples

Basic Usage

.....

Advanced Usage

.....

MCP Server Deployment

It is recommended to integrate and deploy MCP servers to enhance flexibility and unified management. The following solutions are suggested:

• Private NPM Package: Package the MCP server as an NPM package and distribute it through the enterprise private NPM repository.
• Installation and Configuration: Install the MCP server package directly via npm install and achieve quick integration through a configuration file (e.g., mcp.config.json).

Example configuration file mcp.config.json:

{
  "mcpServers": {
    "server1": {
      "name": "example-server",
      "args": ["server.js"],
      "path": "./servers/server1/"
    }
  },
  "mcpClient": {
    "name": "mcp-client",
    "version": "1.0.0"
  }
}

Future Plans

• Currently, the MCP server only implements stdio based on the SSE protocol.
• The context protocol and global variables during requests can be shared across the interface layer, MCP client layer, and MCP server layer.
• Currently, HTTP interfaces are supported. In the future, these interfaces are planned to be encapsulated as sockets, which is more suitable for chat scenarios.
• After completing this project, I have an idea: Is it possible to write a common class based on the original MCP protocol, so that MCP servers written based on this class can support not only MCP client calls but also function calls after importing modules? This might be another independent project.
• The development of the entire project took about one day. Due to time constraints, it will be further improved in the future.