My to-do list kept growing faster than I could clear it, and most of it wasn't hard work. "Compare three suppliers." "Draft the follow-up email." "Check the site still loads." Any decent AI agent could do these. The annoying part was everything around them: open a chat, paste the context, wait, copy the answer somewhere, then forget what I'd asked two days later.
I wanted the cards to just get done. So I built OpenLeo, a to-do board where agents, running on my own harness built on Pi, pick up the cards. It's free and open source under AGPL-3.0.
Here's the 30-second version:
How it works
It looks and feels like Trello: lists, with cards in them. The one difference is that a list can have an agent.
On the board above, a card dropped into Research goes to the researcher agent and a card in Writing goes to the writer. The Checks list has a monitor that looks at my site every weekday at 9:00. Inbox has no agent, so those cards wait for me.
When an agent finishes, the answer stays on the card. Here's a real one. I asked for the price of a Hario V60 at three shops, and the researcher came back with a little table and links:
A card can start because I pressed run, because its schedule came round, or because it landed in a list with an agent. Each run gets noted on the card, so when I look the next morning I can see what happened and when.
The same board also works as a table, with whatever fields you add (I used priority, due date and budget):
Every board gets its own computer
I didn't want agents poking around my real laptop. In OpenLeo each board has its own Linux computer, a cua sandbox, and the agents do their work in there. They open a browser, click around, fill in forms and save files. If I'm curious I open the computer and watch, and if an agent gets stuck I can take over the mouse.
On a Mac it uses Apple's container runtime when it can and falls back to Docker otherwise, and on Windows and Linux it uses Docker. You can also run the computers in cua's cloud. On first launch OpenLeo walks you through setting one up.
The board's notes, its memory and the skills you teach an agent (how you file receipts, say) live on that computer too, right next to the work.
It runs on your ChatGPT plan
I really didn't want people fiddling with API keys. OpenLeo uses OpenAI's Sign in with ChatGPT for open-source apps, the same kind of login Codex uses. You sign in with your ChatGPT account and the agents use the plan you already pay for. The tokens sit in a local vault file with owner-only permissions and never get sent to the web page.
What it isn't
It's for one person and their agents. I'm deliberately not turning it into a team project-management tool. It's also early. Expect rough edges, and back up your data before updating.
Try it
The site is leooi.com and the code is at github.com/ancs21/openleo.
git clone https://github.com/ancs21/openleo
cd openleo
bun install
bun run dev # http://127.0.0.1:3000
Or skip the terminal and get the desktop app. There are builds for macOS (Apple Silicon), Windows (x64) and Linux (x64 and arm64) on leooi.com/download and the releases page.
So here's my question for you: if a to-do card could do itself, which one would you hand off first? And which one would you never trust an agent with? I'll read every answer, and if something breaks, open an issue.
Top comments (3)
I'd definitely hand off simple, repetitive tasks that don't require much visual analysis or human context.
But here's something I wouldn't trust an agent with just yet. :)
Recently, I had to create three presentations for a company event celebrating 25 years in business. I was given around 3,000 photos and 700 videos.
The AI agent did a pretty good job planning the presentations. But then came the real problem.
Who's going to sort through 3,000 photos and 700 videos, pick the best moments, and put everything into the right context?
And how the hell is an AI supposed to know which guy is Sławek and which one is Radek? Or whether a particular photo was taken 20 years ago, 10 years ago, or 7?
That's where things get complicated. It's not just about analyzing images. It's about understanding people, relationships, history, and context that might not exist anywhere in the data.
For now, that's definitely a job I wouldn't hand over to an autonomous agent.
Anyway, really cool project! I like the direction you're taking. Respect for building it and making it open source. Recommended! :)
Thanks! Yeah, no agent will know which one is Sławek 😄 It could still do the dull first pass, like cutting duplicates and blurry shots.
This is close to what I run for my own work: tickets on a board, agents pull them, each ticket gets its own stack that is torn down when it is done, and every run is written into the ticket's thread. To your question: I hand off the code change on a dev stack, and QA and anything near production stay with me. With one computer per board, do two cards in Research ever step on each other's files?