Jo is a personal project I started in March 2025. It's a simple tool to run AI agents on a local PC without the need of a GPU. This article is a timeline I update every time I do a bigger release of Jo.
It's still in development and is not public yet.
As I keep improving it, I'm writing LinkedIn posts about it:
October 2026, a desktop app, built on top of pi agent harness
Jo is now a Tauri + Vue desktop app, still with no GPU requirement: it runs through pi as engine, which support any model.
- Notes allow me to keep all my notes in markdown, like Obsidian. It's the extension of my brain.
- The agent-per-prompt is now Bots: each bot has a specific role and use its own tools, powered by pi.
- The Knowledge bot answers to my questions from my Notes
- The Video producer bot help me to make and edit videos with AI
- The Settings is where i can set the models, the notes folder and the videos folder.
September 2026, a GUI as a productivity tool
Jo stopped being "a prompt picker in a terminal" and got a window: a grid of agent cards: Thinking, Writing, Sales, Planning, Code, Translator.
The translator agent has a proper chat window with a language picker instead of typing language codes.
It also stopped being just an agent runner and became a tool I actually use every day:
- PDF split/merge, downloads, etc.
- translations
- a knowledge app (RAG on my own notes)
March 2025, a CLI script with a prompt folder
The first version was a Python script. One config file, one prompt folder, one idea: each agent is just a prompt file, and the script picks the right one and starts a chat with the LLM.
jo.conf
model=claude api_key=azerty prompts_path=prompts
uv run jo.py [--args]
> jo translate
> jo debug --prompt=odoo-debug
> jo answer prompt=general
> jo code prompt=odoo-code
Only one agent actually worked at that point: the translator. Running uv run jo.py translate opened an interactive session in the terminal.
