This project is a proof of concept (POC) demonstrating how to implement a Model Context Protocol (MCP) with a custom-built host to play with agentic systems. The code is primarily written from scratch to provide a clear understanding of the underlying mechanisms.

See the experimental website for documentation (auto-generated) at https://owulveryck.github.io/gomcptest/

The primary goal of this project is to enable easy testing of agentic systems through the Model Context Protocol. For example:

• The dispatch_agent could be specialized to scan codebases for security vulnerabilities
• Create code review agents that can analyze pull requests for potential issues
• Build data analysis agents that process and visualize complex datasets
• Develop automated documentation agents that can generate comprehensive docs from code

These specialized agents can be easily tested and iterated upon using the tools provided in this repository.

• Go >= 1.21
• Access to the Vertex AI API on Google Cloud Platform
• github.com/mark3labs/mcp-go

The tools use the default GCP login credentials configured by gcloud auth login.

• host/openaiserver: Implements a custom host that mimics the OpenAI API, using Google Gemini and function calling. This is the core of the POC.
• tools: Contains various MCP-compatible tools that can be used with the host:
  • Bash: Execute bash commands
  • Edit: Edit file contents
  • GlobTool: Find files matching glob patterns
  • GrepTool: Search file contents with regular expressions
  • LS: List directory contents
  • Replace: Replace entire file contents
  • View: View file contents

• OpenAI Compatibility: The API is designed to be compatible with the OpenAI v1 chat completion format.
• Google Gemini Integration: It utilizes the VertexAI API to interact with Google Gemini models.
• Streaming Support: The server supports streaming responses.
• Function Calling: Allows Gemini to call external functions and incorporate their results into chat responses.
• MCP Server Interaction: Demonstrates interaction with a hypothetical MCP (Model Control Plane) server for tool execution.
• Single Chat Session: The application uses single chat session, and new conversation will not trigger a new session.

You can build all the tools using the included Makefile:

# Build all tools
make all

# Or build individual tools
make Bash
make Edit
make GlobTool
make GrepTool
make LS
make Replace
make View

Read the .envrc file in the bin directory to set up the required environment variables:

export GCP_PROJECT=your-project-id
export GCP_REGION=your-region
export GEMINI_MODELS=gemini-2.0-flash
export IMAGEN_MODELS=imagen-3.0-generate-002
export IMAGE_DIR=/tmp/images

You can test the CLI (a tool similar to Claude Code) from the bin directory with:

./cliGCP -mcpservers "./GlobTool;./GrepTool;./LS;./View;./dispatch_agent -glob-path .GlobTool -grep-path ./GrepTool -ls-path ./LS -view-path ./View;./Bash;./Replace"

⚠️ WARNING: These tools have the ability to execute commands and modify files on your system. They should preferably be used in a chroot or container environment to prevent potential damage to your system.

This guide will help you quickly run the openaiserver located in the host/openaiserver directory.

• Go installed and configured.
• Environment variables properly set.

1. Navigate to the host/openaiserver directory:

   cd host/openaiserver

2. Set the required environment variables. Refer to the Configuration section for details on the environment variables. A minimal example:

   export IMAGE_DIR=/path/to/your/image/directory
   export GCP_PROJECT=your-gcp-project-id
   export IMAGE_DIR=/tmp/images # Directory must exist

3. Run the server:

   go run .

   or

   go run main.go

The server will start and listen on the configured port (default: 8080).

The openaiserver application is configured using environment variables. The following variables are supported:

• This is a POC and has limitations.
• The code is provided as is for educational purposes to understand how to implement MCP with a custom host.