For the last four months, my working day has run through agent rooms. My AI agents work there, in the open, and nothing they do leaves a room until I say so. Today that room platform is open source. It's called Exec Room.
What it is
Exec Room is one room where AI agents work in view.
You send a request. Coen, the lead agent, takes it. Before he answers, he consults the advisors you've added to the room. When something needs making, he proposes a job for a maker.
Every action lands on a card. Nothing happens until a person looks at it. You approve it, send it back with a note, or dismiss it.
Makers only write drafts. Nothing in the room can post, publish or send on its own. Those actions don't exist there, so no agent can propose them.
Why I built it this way
I wanted agents that do real work without me wondering what they did while I wasn't looking. Most agent setups make you pick one: let them act, or do everything yourself. I wanted a third option: agents do the reading, thinking and drafting, and I make the calls.
So the room is set up for that. You see who Coen asked and what they said. You see what's proposed before anything happens. And afterwards you can see what you decided.
What you can use it for
Exec Room is made for running a business. A room can be for any team: exec, marketing, development. Each room gets its own agents, its own brief about your business and its own limits.
In a room, the agents can:
- Help you think before you decide. Coen asks your advisors, say a technical or a marketing advisor, and you see their answers in the chat.
- Write the first version. Makers write posts, articles and image prompts, and save them as drafts.
- Get the next step ready. An issue or a comment in the GitHub repos you allow. A post saved as a draft in your Notion database, or a longer page in Notion. An email draft that you send yourself. A file saved into a folder you choose. Each one waits on a card until you approve it.
My own example is our marketing room. I ask Coen for something, he checks with my marketing advisor, and the makers write: storyboards for short promo videos, LinkedIn posts, updates for this blog. What comes back are drafts. When I approve, the room saves them in Notion as drafts, or opens a GitHub issue. Posting stays with me. This post started in that room too.
What's in the repo
- Coen, the room lead. Every message goes to him first.
- Tools to build your own advisors and makers, or to import agents you already use in Claude Code. If you've already written agents, you don't have to start over.
- Separate limits for each room: what agents may read, which sites they may open and what they may propose. A room for marketing doesn't need to see your finances, and here it doesn't have to.
- A Decisions tab that lists the proposals you approved, sent back or dismissed. When you wonder why a draft exists or why an issue was never opened, the answer is there.
One thing to be clear about: the open-source version starts with Coen alone. The team around him is yours to build: the advisors you want him to consult and the makers you trust with a job. I think that's the right place to start. The agents that help most are the ones built around how you actually work.
What you need
Exec Room runs locally on macOS or Linux. It needs Node.js 20+ and Claude Code, and it runs on your own Claude plan or an Anthropic API key. Setup is a clone and one script. Then you open the room in your browser.
The README and SECURITY.md say plainly what the room's limits are and aren't. Please read them before you give a room access to anything.
Try it, or have it set up
If you'd like to see it before installing anything, the page has an example under See it in 20 seconds: andrs.nu/exec-room.html
The code is here: github.com/ANDRS-Projects/exec-room
And if you'd rather not set it up yourself: ANDRS can set it up for your team and build rooms and agents around how you work.
I'd love to hear which agent you'd add next to Coen first.