Hi all! I’ve been building chatgpt-cli for a while now (800+ stars), a command-line tool for working with LLMs. It supports things like prompt files, thread-based context, streaming, images and audio, and MCP (Model Context Protocol) tool calls.

I just shipped a new experimental feature I’m really excited about: an agent mode that follows the ReAct loop (think → act → observe), similar to what tools like Claude Code and Cursor do. It can plan and execute multi-step tasks using shell commands, file operations, and LLM reasoning, all behind budget limits and safety policies (allowed tools, denied commands, workdir sandboxing).

It’s still early, for example, you can’t yet resume workflows. I think it’s ready for people to try and break though!

Quick silly test:

chatgpt create a text file in this directory with a short weather report for where i am right now --agent

Next, I want to iterate on the current version based on feedback before adding bigger ideas like parallel execution or sub-agents. If you try it, I’d love to hear what’s confusing, broken, or missing.