It reads before it proposes
Every round starts in the code. The manager knows what exists, what shipped last and what is half-finished.
Give an AI project manager a repository or one blunt sentence. It decides what to build, in what order and how, then keeps handing work to a designer and up to twelve AI developers. You are the CEO: approve, reject, or step out.
A deleted invoice is gone for good today. A 10-second undo saves the support ticket.
Most agent tools wait for your next prompt. This one has a manager whose job is to have the ideas. You set the direction once. The manager owns the roadmap from there.
A tool for freelancers to send invoices. Make it good.
An empty repository gets a running core first: one client, one invoice, a PDF.
Tax offices expect gapless numbers. Nobody asked for this. The manager raised it.
Proposed after reading what round 1 shipped and seeing due dates had no follow-up.
You rejected a dashboard with the note "too early". The manager changed course.
Every round starts in the code. The manager knows what exists, what shipped last and what is half-finished.
Each idea arrives as a work order with the architecture decisions already made: data shapes, files to touch, how to verify.
When the board runs low, it drafts the next round. Your rejections, and anything you tell it in conversation, shape what comes next.
It reads the repository and your brief, then writes work orders with the architecture already decided.
Anything a person will see gets a design spec first: layout, every state, exact wording, the tokens to use.
Approve or reject from your Desk. A rejection can carry a reason, and the manager reads it before the next round.
Free developers pick up tasks in parallel, edit files, run the tests and commit. A blocked command pauses and asks you.
It checks the diff against the work order and the design. Approved work is pushed to GitHub.
On autopilot the manager's proposals go straight to the developers. Left alone, agents would run until your Claude limit. Four brakes stop them first.
Claude Code reports how full your subscription window is after each run. Booardroom reads that number and holds new work once the agents have used their share.
Those are the defaults. Each one is a setting.
Tell it a rule, a wish or a change of mind. It answers, briefs the developers and the designer, and queues the work.
Past Claude Code sessions on your machine show up on their own. claude.ai chats import from a data export.
Hand an old chat to a project. Claude compresses it into a brief the team reads before each run.
Developers may run what you already allow Claude Code. Anything else pauses the task and asks you, with one press to allow.
The server listens on localhost only. Projects, tasks and logs live in one SQLite file in your user folder.
One window for macOS, Windows and Linux. The browser version runs from the same code.
The attack is called slopsquatting: a model suggests a package that sounds right, someone has already published malware under that name, and an unattended agent installs it. Booardroom checks every package before any agent installs it.
$ npm install express react-pretty-state slopsquat-guard blocked this install. BLOCK react-pretty-state: does not exist on the registry (likely a hallucinated name) Do not retry with another guessed name. Find the real package in the project docs or on npmjs.com.
The check is slopsquat-guard, our own open-source plugin for Claude Code. It is one layer of defence. It does not read package source code.
$ git clone https://github.com/GracjanFilipek/Agent-Corp.git booardroom $ cd booardroom $ npm install $ npm run app # or try it with canned agents, no Claude calls $ npm run app:demo
claude -p under your own login, on a Pro or Max plan or an API key.