Prediction Guard × MCP Demo

A minimal CLI demo showing how to connect Prediction Guard's Responses API to a remote Model Context Protocol (MCP) server. The agent uses a mock / demo Salesforce MCP server as its tool backend and maintains a running conversation in the terminal.

What this shows

• How to pass MCP servers as tools via the responses.create API
• How to parse the Responses API output format (output[].content[].text)
• How to maintain multi-turn conversation history with the Responses API

Prerequisites

• Python 3.12+
• uv (package manager)
• A Prediction Guard API key

Setup

# 1. Clone the repo
git clone <repo-url>
cd pg-demo-mcp

# 2. Copy the env template and add your API key
cp .env_example .env
# Edit .env and set: PREDICTIONGUARD_API_KEY=<your key>

# 3. Install dependencies
uv sync

Run

uv run python main.py

You'll see an interactive prompt:

=== GTM Agent | PredictionGuard + Salesforce ===
Type 'quit' to exit.

You: What does our renewal pipeline look like?

Agent: Here's a summary of upcoming renewals...

Type quit or exit (or press Ctrl+C) to stop.

How it works

client.responses.create(
    model="gpt-oss-120b",
    input=conversation,       # full conversation history
    tools=sf_tools_json,      # MCP server config
)

The tools list declares the MCP server — its URL, label, and which tools the model is allowed to call. Prediction Guard handles the MCP protocol; your code only sees the final text response.

MCP tool config shape

{
    "type": "mcp",
    "server_url": "https://your-mcp-server/mcp",
    "server_label": "my-mcp-server",
    "allowed_tools": ["tool_one", "tool_two"],
    "server_description": "What this server does"
}

Project structure

main.py          # All demo logic
.env_example     # API key template
pyproject.toml   # Dependencies (managed by uv)