This project implements Programmatic Tool Calling (PTC) for LLM agents, addressing the inefficiency and token bloat of traditional tool-use methods. By enabling LLMs to generate and execute Python code directly within a sandboxed environment, it significantly reduces token consumption (85-98%) and enhances agent capabilities for complex data processing tasks. It's designed for developers building sophisticated LLM-powered applications that handle large datasets or require intricate, multi-step operations.

How It Works

PTC leverages the LLM's strength in code generation. Instead of making discrete JSON tool calls, the LLM writes Python code that orchestrates workflows. This code is then executed in a secure Daytona sandbox environment. Data processing, filtering, aggregation, and transformations occur locally within the sandbox, with only the final, concise output returned to the LLM's context. This approach, built upon langchain-ai's deep-agent and Daytona for sandboxing, drastically cuts down on token usage, especially when dealing with large structured or time-series data.

Quick Start & Requirements

• Installation: Clone the repository, navigate into the directory, and run uv sync.
• Prerequisites: Python 3.12+, Node.js.
• Configuration: Requires API keys for at least one LLM provider (Anthropic, OpenAI, etc.) and Daytona, configured via a .env file. Additional keys for services like Tavily or cloud storage are optional but recommended for full functionality.
• Demo: A Jupyter notebook (PTC_Agent.ipynb) is provided for a quick demonstration.

Highlighted Details

• Universal MCP Support: Automatically converts existing MCP server tools into Python functions.
• Progressive Tool Discovery: Tools are discovered on-demand, minimizing prompt token usage.
• Secure Sandbox Execution: Utilizes Daytona for isolated code execution with filesystem isolation and snapshot support.
• Enhanced File Operations: Includes refined glob, grep, and other file manipulation tools.
• Multi-LLM Support: Compatible with Anthropic, OpenAI, and other LLM providers.
• Auto Image Upload: Supports automatic uploading of generated charts and images to cloud storage (R2, S3, OSS).
• LangGraph Ready: Compatible with LangGraph Cloud/Studio deployments.
• Native Tools: Provides core functionalities like execute_code, Bash, Read, Write, Edit, Glob, and Grep.

Maintenance & Community

The provided README does not detail specific maintenance contributors, community channels (like Discord/Slack), or roadmap information.

Licensing & Compatibility

The project is released under the MIT License, which is permissive and generally allows for commercial use, modification, and distribution.

Limitations & Caveats

Setup requires obtaining and configuring API keys for multiple external services, including Daytona, which is essential for the core sandboxing functionality. The project's reliance on these external services and specific LLM configurations may introduce dependencies and potential points of failure.