Agent Andon: A Status Board for AI Coding Agents
When running multiple AI coding agents, attention runs out long before compute. Compute is easy to scale, but tracking multiple terminal windows at once is impractical. Agents pause constantly: one pauses for permission, another finishes a turn, and nothing announces the change. Without a watcher, the agent sits idle.
Agent Andon is a status board built for this. It indicates which agent needs attention right now. Each session gets a row. Runs requiring intervention stand out; the rest stay dim. The name comes from the manufacturing “andon,” a status lamp signaling where assistance is needed on a production line.
Dashboard states
Sessions have several states, but only two matter: a turn that ended in an error, or one waiting for input or permission. Running, finished, and idle states require no action, so the board dims them. The active alerts use color and subtle motion, and can be forwarded to a phone when away from the screen.
A finished turn does not mean the underlying work is done. An agent might yield control but leave background tasks running, so the state is marked as returned rather than finished. Andon also waits a few seconds before updating, giving the agent time to resume if it has consecutive steps.
Remote notifications
By default, the board stays on the local network. Accessing it remotely from a phone requires a relay. Most hosted services require sending raw logs to their servers and trusting the operator. Andon takes a different approach, encrypting everything locally. The relay only passes encrypted ciphertext, keeping project names and logs private. The relay sees metadata, including session counts, rough timestamps, and origin addresses, but zero content.
There is a design catch: the relay serves the dashboard web page, which is where the decryption key is entered. Consequently, encryption cannot fully protect against a malicious relay. Running the relay locally avoids this. For the hosted version, a verification check compares the served assets against the public Git repository. This check is a snapshot, not a perfect guarantee, but it makes silent tampering much harder.
Quick start
Self-hosting is the default. The codebase is tiny and easy to audit. For remote notifications, spin up a private relay or connect to the public instance. The project is early.
AGPL-3.0.
EN / ZH