open-ptc-agent

Why we included this project

Most agent frameworks have the model issue one JSON tool call at a time, which gets expensive when the tools return large payloads. This project flips that: the agent writes Python that runs inside a Daytona sandbox, MCP tools are exposed as callable functions, and only the final result returns to the model's context. That can cut token usage by 85-98% when the task involves filtering and aggregating structured or time-series data before the model ever sees it. The codebase is built on LangGraph and borrows from LangChain's DeepAgents, so teams already working with those frameworks can adapt it fairly quickly. It also ships with an interactive CLI, background subagent execution, and demo MCP servers for financial data, which makes it a handy reference for anyone exploring code-execution-based tool calling.