Claude Orgtree


claude-orgtree

A persistent, visual **organization of Claude Code agents** — a tree of real, addressable Claude Code sessions with a credit budget, an office-room canvas, and full agent-to-agent delegation. You sit at the root as the overseer; you hire top-level agents, they hire their own reports, and the whole org runs on your existing Claude Code installation and subscription.

Design document: [PLAN.md](PLAN.md) · UI manual: [docs/ui-guide.md](docs/ui-guide.md) · How much infrastructure to run, and what each tier buys: [docs/infrastructure-tiers.md](docs/infrastructure-tiers.md)

**Design motto:** one thing, done very very well. orgtree is a simple idea — a persistent, visual organization of Claude agents — refined meticulously and taken to its logical conclusion, not a feature jamboree. And within that one thing: permit as much as possible; close gaps with minimal-friction shortcuts; step out of the way. The tree is not a rigid structure you're confined to — it's a **sandbox of capabilities** with a *suggested* organization that improves efficiency on the margin. Asking for what's already true is a no-op, not an error; where a refusal would just tell you which other command to run, orgtree runs it for you and tells you what it did. Hard "no"s are reserved for real resource limits, true impossibilities, and protecting the user's data. Even messaging reach is opt-in structure: siblings always talk directly, so a flat org (or the coordinator charter's open-office floorplan) is a complete graph — nobody is ever out of reach unless you chose the nesting that makes them so.

The model in one breath

**You are the root.** You hire top-level agents; agents hire their own reports with the `orgtree_*` MCP tools every node is given. **Credits are occupancy, not spend** — a live node *holds* its seat (haiku 1 · sonnet 3 · opus 5 · fable 10) plus its grant; retiring releases everything back; tokens are unlimited (real dollars are