MTA: Minimal Terminal Agent
Alongside this blog post, we also release the Minimal Terminal Agent on GitHub. It is a minimal framework for terminal agent research and study.
What is a terminal agent?
Humans use their eyes to see and their hands to manipulate, so graphical user interfaces (GUIs) and touchscreens are natural, ergonomic choices for us. LLMs, in contrast, are much better at processing text, so a terminal interface is their native channel to the external environment. In fact, since early operating systems (e.g., Unix/Linux) were built primarily around terminal control, using the terminal as a general interface for LLMs to interact with our computing devices is very accessible.
Why do we need another (terminal) agent framework?
The Minimal Terminal Agent (MTA) framework aims to provide an extensible, flexible, and easy-to-start foundation for AI agent research. It is not a product-level framework that competes with Codex, Claude Code, or DeepSeek Harness. The philosophy we want to uphold is simplicity: the core of MTA is fewer than 500 lines so far, and we are still working to make it more concise without losing flexibility.
How can this framework facilitate agentic research?
MTA is specifically designed for retrospective research analysis and further development. Here are some features that we find extremely useful, and that even the most capable harnesses (e.g., Codex) do not provide natively:
- A minimal sandbox environment that ensures every trace is reproducible;
- Serializable trajectory management that can replay an LLM's thinking, interactions, or failures for post-training and optimization;
- An integration layer that plugs in various academic benchmarks without changing the source code.
Citation
If you find MTA useful in your research, please cite:
@misc{xiong2026mta,
author = {Zhen Xiong},
title = {{MTA}: Minimal Terminal Agent},
year = {2026},
howpublished = {\url{https://github.com/Eric2i/minimal-terminal-agent}},
note = {GitHub repository}
}