AI REPLs

This project provides a set of bash functions I use to enhance the capabilities of AI agents, like the Cursor agent. Inspired by https://github.com/nickgnd/tmux-mcp.

I think the best feature is the py_run function which makes the cursor chat feel like an psuedocode repl. Just type psudecode in the chat. I didn't use mcp because I wanted to stream the output from the repl.

Right now the 2 parts are the tmux part and the advice part.

The tmux part helps the agent manage tmux sessions. Tmux is great for helping the AI manage long running commands and repls.

Some examples of how the AI can use this:

• manage multiple long running experiments in parallel. Check on those experiments periodicially with tmux capture-pane
• have an ssh session
• have a python repl
• have a python repl within an ssh session
• make plots in a python repl which show up in the matplotlib UI.
• Have the agent iterate quickly on an idea in the python repl. Tell it save plots to image files. Have it write a report in markdown. Have the agent link to the image file in markdown. Have the agent convert markdown to pdf with pandoc and then you have a research paper with nice plots

Since it's tmux you can use tmux attach -t to work closely with your agent.

The advice part lets your agent get advice from multiple reasoning models in parallel.

Setup

in your ~/.bashrc add source <path to this repo>/tmux_run_funcs.sh

Copy the rules files to cursor or whatever agentic tool you use.

The tmux functions depend on having tmux installed. The ask_advice function depends on having github.com/simonw/llm setup with the gemini plugin.

Functions

• ask_advice: Agent can get help from both Gemini and O3 at once
• tmux_run: Run shell commands in a tmux pane and get the output synchronously.
• start_tmux_repl: Start a tmux session with a Python REPL.
• py_run: Execute Python code within a Python REPL in a tmux session.