Model Gateway

model-gateway is a small MCP server and command-line tool that routes named tasks to the language models you choose. Define a tool once in JSON, then use it from an MCP client or the terminal—without writing a separate Python integration for every task.

API keys are read from environment variables and are never stored in the repository.

Quick start

Requirements: Python 3.10 or later and an API key for the provider you choose.

git clone https://github.com/parigihelori352-cyber/model-gateway.git
cd model-gateway
python -m pip install -e .
mg config init

mg config init creates config.json in the directory where you run it. It is ignored by Git, so it is safe to customize locally. The included starter configuration uses the OpenAI-compatible API:

# PowerShell
$env:OPENAI_API_KEY = "your-key"

# macOS / Linux
export OPENAI_API_KEY="your-key"

Then start the MCP server:

python -m model_gateway.mcp_server

For Claude Code, add this entry to your MCP settings. Replace the example path with the folder you cloned:

{
  "mcpServers": {
    "model-gateway": {
      "command": "python",
      "args": ["-m", "model_gateway.mcp_server"],
      "cwd": "C:/path/to/model-gateway",
      "env": {
        "OPENAI_API_KEY": "your-key"
      }
    }
  }
}

Use forward slashes on Windows in this JSON. Alternatively, set MODEL_GATEWAY_CONFIG to an absolute path to use a configuration file stored elsewhere.

Configuration

Start from config.example.json. Each item in capabilities becomes an MCP tool at server startup.

{
  "tool": "summarize_release_notes",
  "description": "Summarize release notes for a non-technical reader.",
  "provider": "openai",
  "model": "gpt-4o-mini",
  "system_prompt": "Write a clear short summary for a non-technical reader.",
  "input_schema": {
    "type": "object",
    "properties": {
      "text": { "type": "string", "description": "Release notes to summarize." }
    },
    "required": ["text"]
  }
}

Providers must expose an OpenAI-compatible chat-completions API. The configuration supports OpenAI and OpenRouter out of the box; add another provider by setting its base_url and the name of its key environment variable. Do not put a real API key in config.json, examples, screenshots, or commits.

Command line

The project includes task-oriented commands for configurations that define the matching capability names:

mg config path
mg config list
mg review path/to/file.py --focus security
mg review placeholder --stdin < path/to/file.py

For a custom configuration, use the dynamic MCP tools. The CLI command names vision, plan, review, decide, workflow, and translate require capabilities named vision_ask, gpt_plan, gpt_review, gpt_decide, gpt_design_workflow, and gpt_translate respectively.

How paths are resolved

Configuration is located in this order: an explicit path supplied by code, MODEL_GATEWAY_CONFIG, config.json in the current working directory, then config.json in the editable project checkout. Run mg config path to see the file that will be used.

Status and limitations

This is an early project. Model identifiers and provider-specific reasoning parameters vary by provider; verify those values in the provider's current documentation before relying on them in production. Keep the server process and its environment private because they contain access to your API keys.

License

MIT