Together Introduces Code Interpreter API for Seamless LLM Code Execution
Caroline Bishop May 21, 2025 16:44
Together.ai launches the Together Code Interpreter (TCI), an API enabling developers to execute LLM-generated code securely and efficiently, enhancing agentic workflows and reinforcement learning operations.
Together.ai has unveiled a groundbreaking tool, the Together Code Interpreter (TCI), which provides an API designed to seamlessly execute code generated by Large Language Models (LLMs). This development is poised to enhance the capabilities of developers and businesses employing LLMs for code generation and agentic workflows, according to together.ai.
Streamlining Code Execution
While LLMs are adept at generating code, they traditionally lack the ability to execute it, necessitating manual testing and debugging by developers. TCI addresses this limitation by offering a straightforward approach to securely execute LLM-generated code at scale. This innovation simplifies agentic workflow development and paves the way for more advanced reinforced learning operations.
Key Features and Applications
The Together Code Interpreter operates by taking LLM-generated code as input, executing it in a secure sandbox environment, and outputting the results. This output can then be reintroduced into the LLM for continuous improvement in a closed-loop system. This process allows for richer, more dynamic responses from LLMs.
For instance, when an LLM like Qwen Coder 32B generates code to create a chart, TCI can execute the code and produce a visual output, overcoming the LLM's inherent execution limitations.
Enhancing Reinforcement Learning
TCI's rapid code execution capabilities have attracted significant interest from machine learning teams focusing on reinforcement learning (RL). It enables automated evaluation through comprehensive unit testing, facilitating efficient RL training cycles. TCI can handle hundreds of concurrent sandbox executions, providing the secure environments necessary for rigorous testing and evaluation.
Notably, the open-source initiative Agentica, from Berkeley AI Research and Sky Computing Lab, has integrated TCI into their RL operations. This integration has accelerated their training cycles and improved model accuracy while maintaining cost efficiency.
Scalability and Accessibility
Together.ai has introduced the concept of “sessions” as a unit of measurement for TCI usage, priced at $0.03 per session. Each session represents an active code execution environment, lasting 60 minutes and supporting multiple execution jobs. This model facilitates scalable, efficient use of TCI across various applications.
Getting Started with TCI
Developers can begin leveraging TCI through the available Python SDK or API, with comprehensive documentation and resources provided by Together.ai. This launch includes support for MCP, allowing the integration of code interpreting abilities into any MCP client, expanding the tool's accessibility and utility.
The Together Code Interpreter is set to transform how developers approach LLM-generated code, offering a streamlined, scalable solution for executing complex workflows and enhancing machine learning operations.
Image source: Shutterstock