Command Center

Run a fleet of coding agents like a team. One board where every agent — across every project, repo and session — reports what it's doing, what's blocked, and what needs you. You glance at it when you're back at your machine and see, in one screen, the two things that actually matter: what needs a decision, and what's ready to review.

If you drive more than one agent at a time, you know the failure mode: five terminals, three repos, and no idea which one is stuck waiting on you. Command Center fixes that. Agents write short status reports to a local store; the board generates itself from that store and groups everything by product. No dashboards to wire up, no service to run — it's a ~800-line Node CLI with zero dependencies.

Install

npx skills add chakkyy/agent-command-center

That drops the command-center skill into your agent.

This is agent-first. The ccmd CLI ships inside the skill — it's not on your PATH, and you're not meant to type it into a shell yourself. Your agent runs it for you. Inside Claude Code the skill is a slash command, so just:

/command-center            → set it up, or open the board
/command-center serve      → run the live dashboard

…or ask in plain language ("set up the command-center skill", "open the board"). On first use the skill's SKILL.md walks the agent through interviewing you (products, theme, features) and running ccmd init — so ccmd: command not found never happens, because you never type ccmd. The agent does, by path.

Want a real /ccmd command? ccmd init offers it, or run it any time — it writes a Claude Code slash command with the binary's path baked in:

ccmd install          # → ~/.claude/commands/ccmd.md
/ccmd serve           # then, in Claude Code: run the board
/ccmd goal --status ready --title "…"

Prefer a plain shell ccmd? (optional)

Link the bundled binary onto your PATH once — the agent can do this for you:

# from wherever the skill was installed:
skills/command-center/bin/ccmd link      # symlinks ccmd → ~/.local/bin
ccmd init && ccmd serve                  # now plain `ccmd` works

ccmd link prints a PATH hint if ~/.local/bin isn't on your PATH yet. The hooks call the binary by absolute path (ACC_CCMD), so they work with or without the link.

Codex & other agents

ccmd is a plain Node CLI — it works with any agent that can run a shell.

• Codex: npx skills add chakkyy/agent-command-center --agent codex (add --global for a global install). Codex invokes the skill as $command-center (or $command-center:command-center when installed as a plugin). It has /goal but no /loop.
• The /ccmd slash command (ccmd install) is Claude-Code-specific — it writes ~/.claude/commands/ccmd.md. On other agents, run ccmd by path or ccmd link it onto your PATH.
• Hooks are separate from the skill. npx skills add installs the skill (the reporting protocol); the enforcement hooks are opt-in and installed manually (they need jq, git, bash, and /hooks authorization in your agent). Requires Node ≥ 18. The hooks assume Bash + a POSIX /tmp; on Windows use the CLI without them.

How it works

1. Agents report themselves. At the end of a turn, or the moment they're blocked, they run one command:

   ccmd report --status IN_PROGRESS --summary "Rebuilding checkout as a 3-step flow; running the Stripe suite now" --task "PROJ-412 / PR #88" --next "ship review step"
   ccmd signal --kind question --msg "Guest checkout, or force account creation? UI is built both ways behind a flag — I just need the call."

2. The board generates itself from the store ($ACC_HOME, default ~/.agents-command-center). project and branch auto-derive from git. Nobody hand-edits HTML.

3. You read it and clear items. On ccmd serve, "Mark read" persists instantly. Anywhere else, "Copy what I read" hands you a line you paste back into any agent chat to drop those items.

The Overview is deliberately a triage screen: one line per pending item, each jumping to the product tab that holds the full detail. Blockers and questions stay pinned, highlighted, until an agent resolves them.

Goals — work that runs while you sleep

Some work can run for hours without you toward a clear finish: a big migration, a deep refactor, a backlog burn-down. Instead of losing it in a tracker, an agent curates the full paste-ready prompt (with a completion condition) and shelves it as a goal. You pick one any night and fire it with Claude Code's /goal — it runs until the goal is met.

ccmd goal --title "Migrate the test suite to Vitest" --hours 5 --status ready \
  --prompt-path goals/web/vitest-migration.md --project web

Renamed from "loops" — the shelf is about what done looks like, so it pairs with /goal (completion) rather than /loop (recurrence). ccmd loop still works as a deprecated alias.

Themes

Three built-in themes plus a custom accent, injected as CSS variables at render time. swiss (Apple-ish, light/dark auto), terminal (dense mono, ops vibe), and soft (Linear-like).

Make it automatic (optional)

The hooks wire reporting into Claude Code so agents can't forget: a Stop hook blocks a turn that changed a repo without reporting (once, never loops), and re-renders the board; a UserPromptSubmit hook archives what you've read when you come back.

Why local-first

This started life publishing an HTML mirror to a hosted page on every turn. That was pure overhead — re-uploading the whole board on every report. The live ccmd serve at localhost is the default now; the hosted mirror survives only as an optional adapter (features.mirror) for teams that want a shared URL.

Reference

• The agent protocol → — what agents should report, when, and how (this is the heart of it).
• Full CLI reference →
• Hooks →

Works on mobile too:

MIT licensed.