Skip to content

README.md

Latest commit

 

History

History
105 lines (79 loc) · 5.64 KB

README.md

File metadata and controls

105 lines (79 loc) · 5.64 KB
Raw

Predactiv MCP — AI Agent Samples in Python

Sample code demonstrating how to connect AI agents to the Predactiv MCP server (https://mcp.predactiv.com) from Python.

The Predactiv Data Platform is exposed as a set of tools over the Model Context Protocol (MCP). These samples show how to authenticate, discover those tools at runtime, and let a large language model use them to work with your audiences, datasets, datasources, destinations, and more.

What the Predactiv MCP server can do

Predactiv MCP server (https://mcp.predactiv.com) turns the Predactiv Data Platform into an AI agent-native toolset over MCP, letting an LLM go from raw data to modeled, deliverable audiences without leaving the conversation. The toolset spans a few broad areas (and grows over time):

  • Audience intelligence — build, preview, and refine audience segments with live size estimates, then deliver them to downstream platforms. Segments can be assembled from natural-language intent, predictive signals, and rich filters.
  • Data + ML modeling — datasets aren't just stored, they're activated. First-party data feeds a library of machine-learning models that enrich and expand reach: audience enrichment, lookalike modeling, URL-similarity / contextual models, and more — with the catalog of model types expanding over time.
  • Signals & discovery — explore thousands of predictor signals, filter metadata, data sources, and delivery destinations (e.g. LiveRamp, The Trade Desk) to target precisely.
  • Insights — surface demographic, behavioral, and geographic breakdowns of an audience with share and index analytics.

Tools are discovered dynamically at runtime (tools/list), so agents always see the latest capabilities, and every call is OAuth2 bearer-authenticated — see oauth-token-lib.

What's in this repo

Each directory is a self-contained sample built on a different framework, plus a shared authentication helper:

Directory What it shows LLM (for demo)
langchain/ Agent built with LangChain + langchain-mcp-adapters, client-side tool-calling loop OpenAI
claude/ Agent using Anthropic Claude via the native server-side MCP connector Anthropic Claude
openai/ Agent using OpenAI GPT via the native server-side MCP connector OpenAI
mcp-sdk/ Direct use of the MCP Python SDK, client-side tool-calling loop with no framework OpenAI
oauth-token-lib/ Shared OAuth2 authentication helper used by all samples —

The LLM column is just what each sample happens to use for demonstration — not a requirement. The framework-based samples are model-agnostic: langchain/ and mcp-sdk/ can run on any LLM that supports tool calling (OpenAI, Anthropic, or others) by swapping the client and the *_MODEL constant at the top of agent.py. The two connector samples are tied to their provider by design — claude/ demonstrates Anthropic's native MCP connector and openai/ demonstrates OpenAI's, so each only makes sense with that provider. Predactiv MCP itself is LLM-agnostic; the choice of model is entirely yours.

How the samples work

Every sample follows the same three steps:

  1. Authenticate — oauth-token-lib exchanges your CLIENT_ID and CLIENT_SECRET for an OAuth2 bearer token (client-credentials flow).
  2. Connect & discover tools — the sample connects to the Predactiv MCP server over HTTP and fetches the available tools dynamically (nothing is hardcoded).
  3. Run the agent — the MCP tools are bound to an LLM, which calls them to accomplish a natural-language goal.

The claude/ and openai/ samples are a variation on steps 2–3: instead of fetching tools and running the loop client-side, they hand the MCP server definition to the provider's API (Anthropic's Messages API / OpenAI's Responses API), and the provider connects to the server and runs the tool-calling loop server-side. See each sample's README for details.

Getting started

  1. Get credentials. You'll need a Predactiv CLIENT_ID and CLIENT_SECRET. Contact Predactiv if you don't have them.
  2. Pick a sample. langchain/ is a good starting point.
  3. Follow that sample's README for setup, configuration, and run instructions.
cd langchain

# (recommended) create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate        # on Windows: .venv\Scripts\activate

pip install -r requirements.txt
export CLIENT_ID=your-predactiv-client-id
export CLIENT_SECRET=your-predactiv-client-secret
python agent.py

Prerequisites

  • Python 3.9+ (the mcp-sdk/ sample requires Python 3.10+, per the MCP Python SDK)

  • Predactiv OAuth2 API credentials (CLIENT_ID / CLIENT_SECRET)

  • An LLM provider API key, depending on the sample:

    See each sample's README for details.

Learn more

License

See LICENSE.