claude-team_skill

Overview

This skill orchestrates multiple Claude Code workers through iTerm2 using a local MCP server. It spawns isolated worker sessions (each in its own iTerm pane and optional git worktree), assigns beads issues or custom prompts, and provides tools to monitor, message, and close workers. The workflow emphasizes parallel development, context isolation, and visible agent sessions you can interact with.

How this skill works

The MCP server manages worker sessions and exposes mcporter commands to spawn, list, message, inspect, and close workers. Each spawned worker can create a git worktree on an ephemeral branch, run beads commands to pick up an issue, and commit changes while the coordinator stays clean. The server can run as a persistent HTTP service for session recovery, cron integration, and faster CLI responses.

When to use it

• Fan out parallel tasks across multiple Claude Code workers
• Isolate feature work or experiments using per-worker git worktrees
• Run beads-driven issue implementations end-to-end without manual context switching
• Monitor and interact with live Claude Code sessions in iTerm2
• Automate background monitoring and notifications via cron or launchd

Best practices

• Always spawn workers for code changes—do not edit files directly in the coordinator
• Assign bead IDs so workers follow the beads workflow and mark issues in progress/completed
• Use worktrees to keep each worker’s commits isolated and parallelizable
• Enable skip_permissions only when workers must write files; review permissions risk
• Monitor workers and review their branches before merging or cherry-picking
• Always close workers to commit and clean up ephemeral worktrees

Example use cases

• Parallelize a large refactor by spawning workers for submodules and waiting for all to finish
• Assign a beads issue to a single worker that creates a feature branch in a worktree, implements changes, runs tests, and commits
• Run a short-lived worker to generate unit tests or documentation snippets and then extract commits to main
• Use the HTTP mode with launchd to keep the MCP server persistent and let scripts spawn workers from CI or cron
• Set up a cron job that notifies via Telegram when workers become idle or complete their tasks