Oleksandr Kuzmenko, EditorPublished 5 min readAI-assisted · editor-reviewed
TL;DR
- Isolates each Claude Code instance within its own git worktree and tmux session to prevent state corruption.
- Halts execution and surfaces visible permission requests indefinitely across restarts until explicitly answered.
- Rotates between multiple configured Claude accounts automatically when rate limits or quotas are reached.
Key facts
- Required Runtime
- Node.js 22.18+, tmux, git, pnpm
- Agent Execution Context
- Dedicated tmux session per process + Git worktree
- Inactivity Sleep Timeout
- 30 minutes idle
- Supported Permission Modes
- Standard, Auto, Unrestricted
Source: Key facts — Inactivity Sleep Timeout
Source: Key facts. Inactivity Sleep Timeout: 30 minutes idle. Each figure is copied from the story.
- Inactivity Sleep Timeout30 minutes idle
Architecture and Isolation
Orchestrator runs on a single host (Node.js 22.18+, tmux, git, and pnpm) and maps each agent directly to an OS-level environment. Every assigned task spawns a clean git worktree and branch. Each worker process executes inside an isolated tmux session with full visibility over its tool invocations and terminal output.
Tasks can be driven by a persistent supervisor or by setting continuous goals via /goal <command>. The agent continues iterating until verification commands, such as automated test suites, return exit code 0.
Permission Guardrails and Account Cycling
The orchestrator exposes three permission tiers:
- Standard: Routine file edits and terminal commands proceed, but all significant actions require explicit human approval.
- Auto: Delegates routine decisions to Claude Code's native auto-mode, prompting only on critical forks.
- Unrestricted: Enables Claude Code bypass-permissions mode, while retaining human-in-the-loop gates for high-risk system commands.
To address quota depletion, the server allows registering multiple Claude accounts per machine. When one plan is exhausted, sessions transition seamlessly to alternate accounts or fallback models such as Opus or Haiku without manual context rebuilding.
Try it in 2 minutes
pnpm build && pnpm start --openWhen to use
- Managing multiple Claude Code agents collaborating simultaneously across large repositories.
- Enforcing strict automated test validation and code review prior to merging agent contributions.
When NOT to use
- Single-file quick fixes where a basic Claude Code terminal command is sufficient.
- Environments lacking tmux support or where local process spawning is strictly restricted by security policies.
What to do today
- Install dependencies: ensure Node.js 22.18+, tmux, git, and Claude Code CLI are authenticated.
- Clone the orchestrator repository and run pnpm build to package the office server and agent layers.
- Launch the workspace via pnpm start --open or package a local desktop app using pnpm desktop.
Step names are taken from What to do today.
Step names are taken from What to do today. Install dependencies: ensure Node.js 22.18+, tmux, git, and Claude Code CLI are authenticated.. Clone the orchestrator repository and run pnpm build to package the office server and agent layers.. Launch the workspace via pnpm start --open or package a local desktop app using pnpm desktop.
- Install dependencies: ensure Node.js 22.18+, tmux, git, and Claude Code CLI are authenticated.
- Clone the orchestrator repository and run pnpm build to package the office server and agent layers.
- Launch the workspace via pnpm start --open or package a local desktop app using pnpm desktop.
