Every AI session on one board, and which ones are waiting on you

When you run a dozen AI sessions at once, the hard part isn't the work. It is knowing which session is busy, which has stalled, and which finished ten minutes ago and is waiting for you to answer a question.

Agent Commander shows every session on the machine as a character in a live city, one building per department. It is a map of attention: one look tells you where you are needed.

• What it is: A live board of AI coding sessions
• For: Anyone running many AI agents in parallel
• Runs: Locally, on Mac and Windows
• My role: Designed and built it

The bottleneck is the person, so show the person where to look

AI agents work in parallel. A person does not. Every minute a session spends waiting for a decision is wasted, and with many terminals open it is easy to miss the one that needs you.

A list of processes doesn't solve that. A picture does.

Three views of the same sessions

A 3D campus, a network view, and a pixel-art office, over a table that says what each session last said.

• Campus: The default view. Each department is a building, and each AI session is a character at a desk inside it: typing when it is working, on its feet with a hand up when it needs you.
• Nexus: The same sessions as a network: one spine per project, one branch per agent, data flowing out when working and in when waiting.
• Office: A pixel-art floor plan. Sessions that are working sit at their desks. Everyone else is in the kitchen, flagged if they need you.
• Sessions: The table under every view: what each session last said, how long it has been in that state, and how many things it is asking of you.
• Docs: A shelf of every project's documents, readable in place: Markdown, HTML, PDF, text, and images.
• Session drawer: Open any session to chat with it, see exactly what it needs from you, or shut it down properly after it runs its closing routine.

Agent Commander in seven short videos

1. Every AI session on one board. Every Claude Code session on the machine, drawn as a city you can read at a glance. Watch the video
2. Which one is waiting on you?. A session that has stopped and needs an answer stands up with its hand raised. Watch the video
3. See what each session is doing. Open any session to read what it last said and what it is asking for. Watch the video
4. A stalled session is not a surprise. A session that has gone quiet is marked, so it is found before the afternoon is lost. Watch the video
5. Three ways to look. The city, a network with the traffic visible, and an office floor: the same sessions in each. Watch the video
6. The whole fleet in one table. Every session in rows, for when you want a list and not a picture. Watch the video
7. Walk the floor. A walk through the city from building to building, the way you would check on a team. Watch the video

What makes it work

• It works out each session's state: Working, waiting on you, stalled, or idle, inferred from what the session last did and how long ago.
• It reads what each session is asking: Requests are pulled out of each session's closing list and tagged as a decision, a task, a notification, or a note.
• Three ways to look: A city where each department is a building, a network of projects and agents, and an office floor. The same data, three moods.
• Chat without switching windows: Open a session's drawer to read its last message and reply to it, with files attached if needed.
• It shuts sessions down properly: Turning a session off first runs that project's closing routine, waits for it to finish, and only then ends the process.
• It doesn't guess: When it cannot prove which process belongs to which session, it says unproven instead of picking one.
• A document shelf: Every project's documents in one panel, readable in place.
• A character for every session: Each session is a named character with a role and a look, in its project's colour, so you recognise it at a glance.

Guardrails

• Keys and tokens are removed from what sessions say before it reaches the screen.
• It only ever shows prose, never commands or code changes.
• It will not answer a security prompt on your behalf.
• It runs only on your own machine and is not reachable from the network.

Under the hood

• Plain HTML, CSS, and JavaScript
• Three.js for the 3D views
• Python standard library server
• tmux
• No framework, no build step

If your team is starting to run AI agents in parallel, the tooling around them matters as much as the agents. I build that too.