We built the thing that runs AI agents.
Cardloop is our own open-source cockpit for running Claude agents from a kanban board. You write a card, and an agent does the work on its own branch, away from your main code. We designed and built it, released it under the MIT license, and use it every day to run our own projects.
- Project
- Our own product
- Our work
- Design, code and every release
- Timeline
- Public since June 2026, in daily use here
- Stack
- Python, React, Claude Agent SDK, SQLite search
- Status
- Open source (MIT), self-hosted, on GitHub
What the user gets
From a written task to changes you approved, in four steps.
-
Write a card
Add a task to a project's board. The card text becomes the instruction for the agent.
-
The agent works on a copy
Each card runs on its own branch and working folder, so the main code stays as it was. You can watch it live from a laptop or a phone.
-
Review the result
The card lands in Review with a summary and the exact changes. Check runs the project's tests and marks the result safe or risky before you decide.
-
Apply or discard
Apply merges the work into your branch. Discard deletes the card's branch and folder.
What comes out
Every finished card leaves a review page you can read in a minute: what the agent says it did, and what it actually changed. Shown on a demo project.
- The agent's own summary of the run. In this demo it is scripted: one new function and one test.
- The changed files and the full diff, taken from git against the branch the card started from.
- A Check button that runs the project's tests before anyone decides.
- Apply and Discard: the main branch stays untouched until a person presses Apply.
How it stays safe to use
An agent with full access to a machine needs hard rules. Four of ours.
Nothing merges by itself
The agent works on its own branch and folder. Its changes reach the main branch only when a person presses Apply.
A few commands are blocked outright
The agent has full access to the machine it runs on, and we say so plainly. A short list of irreversible commands, such as wiping the home folder or force-pushing to the main branch, is stopped before it runs. Plan mode and ask mode add approval steps when you want them.
A usage limit does not lose the work
When the five-hour usage limit interrupts a run, Cardloop queues it to resume after the limit resets, so a long task can finish overnight.
Updates can undo themselves
An update waits until the agents are idle, restarts the service and checks it for about a minute. If the check fails, Cardloop goes back to the previous release on its own.
Want a workflow like this running your own business?
We build AI workflows the same way: every change isolated, read by a person and easy to undo. We also install Cardloop on a server you own, see local AI setup. Tell us which process you would hand over first.
Tell us what's slowing you down.
You get a free written review by email within two business days.