Agent Board Kit — Git skill for Claude Code

agent-board-kit

Run a fleet of Claude Code and Codex agents on an always-on Linux machine and manage them from your inbox, a Mac menu-bar board and a GitHub-Issues to-do list.

• Email your agents. Every agent mails you from one shared Gmail account. Your reply lands back in the exact session that wrote to you, even hours later. Mail that no agent owns goes to an orchestrator (one fresh claude -p run per email), which starts, relays to, and closes agents for you.
• Mail you can act on. Before an agent's mail leaves, a cheap model with no context reads it as you will. If it would leave you asking "what does this mean?" or "where's the plot?", it goes back to the agent once, with concrete fixes. Plots go inline, and a mail containing an API key or token is never sent. Mail to collaborators and other people is checked for clarity too.
• One worktree per agent. agent-spawn makes a branch off origin/main, a git worktree, a tmux session and a memory-capped systemd scope, and launches Claude Code or Codex in it with a brief. agent-close removes it, but only when its work is landed.
• See everything at a glance. The agent board (a Mac menu-bar app) shows every session as a tile coloured by state (busy, waiting on you, blocked, idle, offline), with git progress toward main, plan usage, and right-click actions (message, close, resume, switch model/login).
• A to-do list agents can work on. Your tasks are issues in a private GitHub repo. A web app and a Mac app edit them. Assign to agent starts an agent on a task, or hands it to one that is already running; it reports back on the issue and by email.
• Free status notes. Hooks log every turn; agents add one-line milestones with agent-status note. A status report email costs one small model call.

Everything personal (names, addresses, GitHub names, paths, hosts) lives in one file, `~/.config/agent-kit/kit.env`. The kit contains no one's data or credentia