claude

Predactiv MCP — Anthropic Claude AI Agent

This sample shows how to connect an AI agent to the Predactiv MCP server using Anthropic Claude directly.

Unlike a client-side agent loop, this sample uses Claude's native MCP connector: the MCP server definition is handed to the Messages API, and Anthropic's infrastructure connects to the server, discovers its tools, and runs the tool-calling loop server-side. Your code makes a single messages.create call and reads back the final answer.

Contrast with the mcp-sdk/ sample, which uses the raw MCP Python SDK and runs the tool-calling loop client-side. Here, Anthropic owns the connection and the loop.

How it works

1. Authenticate — oauth_token_lib.get_bearer_token() performs the Predactiv OAuth2 client-credentials exchange and returns a bearer token, passed to the MCP server as authorization_token.
2. Connect & discover tools — the mcp_servers parameter declares the Predactiv server (type, name, url, authorization_token). The matching mcp_toolset entry in tools (linked by mcp_server_name) enables those tools for the model.
3. Run the agent — Claude connects to the MCP server, lists its tools, calls them as needed, and returns a final response. The mcp-client-2025-11-20 beta flag enables the connector.

The Predactiv MCP tools expose the data platform: audiences, datasets, datasources, destinations, filters/predictors, and insights (predactiv_* tool families).

Prerequisites

• Python 3.9+
• A Predactiv account with API credentials (CLIENT_ID and CLIENT_SECRET). Contact Predactiv if you don't have these.
• An Anthropic API key — this sample uses claude-haiku-4-5 as the agent's LLM. Set it as the ANTHROPIC_API_KEY environment variable.

Two separate credentials. ANTHROPIC_API_KEY authenticates your calls to Claude. CLIENT_ID / CLIENT_SECRET authenticate to the Predactiv MCP server via oauth_token_lib. Both must be set.

Setup

From this claude/ directory:

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

# install dependencies (also installs the shared oauth-token-lib package)
pip install -r requirements.txt

Configuration

Set the following environment variables:

export CLIENT_ID=your-predactiv-client-id
export CLIENT_SECRET=your-predactiv-client-secret
export ANTHROPIC_API_KEY=your-anthropic-api-key

Run

python agent.py

The sample's default goal is:

Get list of my audiences from the Predactiv MCP server and return it to me.

Edit the GOAL constant in agent.py to ask the agent to do something else with your Predactiv data. The model is set in the messages.create call (model="claude-haiku-4-5") — swap it for a more capable model (e.g. claude-sonnet-4-6, claude-opus-4-8) if your task needs it.

Troubleshooting

• CLIENT_ID and CLIENT_SECRET environment variables must be set. — Export both variables (see Configuration) before running.
• 401 Unauthorized from the MCP server — Your credentials are invalid or expired. Verify them, or run python ../oauth-token-lib/oauth_token_lib.py to confirm a token can be obtained.
• 429 rate_limit_error from Anthropic — The MCP connector injects the server's full tool catalog into each request as input tokens, so a single call can be large. If you hit a 429, wait ~60 seconds between runs, use a model with more headroom, or request an increase at https://console.anthropic.com/settings/limits.
• AttributeError reading the response — response.content is a mixed list of blocks (MCP tool-use, tool-result, and text). The sample joins only the text blocks to build the final answer — don't assume content[0] is text.
• Anthropic authentication errors — Make sure ANTHROPIC_API_KEY is set and valid.

Learn more

• Predactiv MCP server: https://mcp.predactiv.com
• Anthropic MCP connector: https://docs.claude.com/en/docs/agents-and-tools/mcp-connector
• Model Context Protocol: https://modelcontextprotocol.io