ds_agents

Python package that holds the analytics MCP agent definitions used by both the CLI (chat.py) and the ds-chat backend.
Each agent configures a Model Context Protocol (MCP) server connection plus instructions, safety rails, and tool allow‑lists for a particular workflow.

Package Overview

ds_agents/
├── __init__.py
└── mcp_agents/
    ├── base.py          # BaseMCPAgent scaffold
    └── generic.py       # GenericDatabaseMCPAgent definition

• BaseMCPAgent centralizes MCP server spawning, tool filtering, and agent construction.
• GenericDatabaseMCPAgent configures the shared “Database Explorer (stdio)” experience with provider monitoring instructions and helper tools.

Installation

cd ds-agentic-workflows/ds-agents
python -m venv .venv && source .venv/bin/activate
pip install -U openai-agents
pip install -e .

When working inside the parent repo you can skip installation—agent_core.py injects this directory onto PYTHONPATH automatically.

Using the Agents

from ds_agents.mcp_agents import GenericDatabaseMCPAgent
from agents import Runner

agent = GenericDatabaseMCPAgent(common_tables=["prod.monitoring.provider_combined_audit"])
async with agent.create_mcp_server() as server:
    runner_agent = agent.build(server)
    result = await Runner.run(runner_agent, input="Show top site issues for QL2.")
    print(result.final_output)

The same agent is automatically wired into:

• chat.py (interactive CLI)
• backend/agent_runner.py in the ds-chat FastAPI service

Extending

1. Subclass BaseMCPAgent. Override name, instructions, and get_wrapper_script() for your specific workflow.
2. Expose additional tools. Update allowed_tools or extend ds-mcp to publish new MCP tool definitions, then include them via allowed_tool_names().
3. Tune model settings. Adjust self.model_settings for temperature, reasoning effort, etc.

Ship the new agent by importing it wherever GenericDatabaseMCPAgent is currently referenced.