Hackeroom — give Claude Code a dev team you hire yourself, and watch them work in an office

Dev House I got tired of telling one Claude Code session to plan, code, review, test, and fix its own mistakes. So I built a software team around Claude Code. You literally hire the team. Give each member a name, role, model, permissions, and skills — then put them into seats like: Planner → Coder → Reviewer → Tester → Auditor Each member runs in its own real Claude Code session. And you can watch the whole thing happen in a pixel-art office. The weird part is that the team isn't hardcoded. Want 3 coders and no reviewer? Do it. Want an auditor that can only inspect? Do it. Leave a seat empty? That phase simply doesn't run. A few things I went surprisingly deep on: Permissions actually enforce themselves. A PreToolUse hook checks every tool call against a capability manifest. .claude/ is unwritable, so an agent can't modify its own permissions. Skills are isolated per agent. A huge skill document doesn't get dumped into everyone's context. It's loaded only when that member needs it. Memory is shared without sharing conversations. Agents write down what they learn, and other agents get a tiny indexed summary before reading the full memory on demand. Token usage is tracked per member. You can see exactly who's burning tokens and which model they're using. So instead of: You → Claude → everything it's more like: You → Team → Planner → Coder → Reviewer → Tester And yes, there are caveats. The part I'd actually love people to tear apart is the permission/hook architecture. I'd especially like feedback on whether the capability-manifest + PreToolUse approach is actually robust enough, and where it could be bypassed or improved. Repo: https://github.com/SinghAbhinav04/dev-house If you find it interesting, give the repo a star. It'd mean a lot.