CC Commander
A guided AI project manager for Claude Code. /ccc reads your project, recommends one next step and offers the alternatives as a click-first menu. Specialist agents do the work; lifecycle hooks keep state, guard secrets and check results.
| What ships in v8.1.5 | Count |
|---|---|
Plugin skills (/ccc-*) |
93: 32 model-visible, 61 on demand |
| Specialist agents | 22 |
| Lifecycle hooks | 30 hook events, 51 handlers |
| Bundled MCP servers | 2 bundled MCP (context7, sequential-thinking), +16 opt-in via /ccc-connect |
| Plugin workflows / evals | 6 / 8 |
| Ecosystem catalog | 467 skills across 11 CCC domains |
Counts come from the filesystem: node scripts/check-parity.js fails the build when any doc disagrees with disk.
Install
Claude Code CLI 2.1.292 or later (local hooks need Node.js 20.19+ or 22.12+):
/plugin install commander --marketplace KevinZai/commander
That installs commander@commander-hub. On older builds:
/plugin marketplace add KevinZai/commander
/plugin install commander
Claude Code Desktop: Settings → Plugin Marketplace → Add from GitHub → KevinZai/commander → install commander, then quit and reopen.
Codex CLI:
codex plugin marketplace add KevinZai/commander
codex plugin add commander@commander-hub
Codex treats plugin hooks as untrusted until you review them with /hooks inside Codex; the one-time steps are in dual-host hooks.
Once the Anthropic plugin directory lists it, the same plugin installs as cc-commander@claude-community.
Update with claude plugin marketplace update commander-hub && claude plugin update commander, then restart. /ccc-update walks through it and /ccc-doctor diagnoses stale installs.
Example prompts
/ccc what should I do first in this repo?
/ccc-review review this branch against main before I open the PR
/ccc-plan turn this Linear issue into a plan with acceptance criteria
/ccc-doctor
Load on demand
Claude sees 32 skills: the 12 core skills (ccc, ccc-build, ccc-review, ccc-research, ccc-security, ccc-plan, ccc-testing, ccc-suggest, ccc-systematic-debugging, ccc-design, ccc-qa, ccc-save-session) plus the skills their menus open. The other 61 set disable-model-invocation: true: their descriptions stay out of Claude's context, and you open any of them by typing /ccc-<name>. When a prompt matches an on-demand skill, the router offers it as a menu choice instead of invoking it.
Session-start context, measured in CC-2112 with js-tiktoken cl100k as a proxy:
| Before | After | |
|---|---|---|
| Skill listing | 22,643 chars, 283% of the 8,000-char budget, truncated | 7,266 chars, 90.8%, no truncation (npm run measure:listing) |
| Session-start total | about 6,700 tokens | about 2,900 tokens |
The Commander Method
Twelve enforceable gates that any model can be checked against: plan before build, verifier separate from worker, prove before alarm, loops need gates, leave durable state, and seven more. /ccc-fable on arms them for a session; the doctrine is in rules/fable-method.md.
What's inside
- Skills: the full generated table of all 93, with which ones Claude sees, is in SKILLS-INDEX.md.
- Agents: architect, reviewer, builder, security-auditor, debugger, designer, qa-engineer and more, on model aliases (
fable,opus,sonnet). Your selected session model is kept; only delegated agents use their own alias. - Hooks: session state, intent routing, secret-leak guard, config protection, permission gate, knowledge capture. The cost ceiling is off unless you set
CCC_COST_CEILING. The voice layer is off unless you setCCC_VOICE=1. - Workflows and evals:
/commander:ccc-deep-review,ccc-audit,ccc-release-auditand three more;claude plugin evalcases for the core flows. - Commander HUD mod: an observe-only band above the prompt;
/ccc-hudtoggles it.
PR workflows (adapted from tstack)
Nine pull-request skills are adapted from t3dotgg/tstack (MIT, pinned at 92e38e2c). All load on demand: type the command, or the router offers it when your prompt matches. They also appear in the CLI menu under PR workflows and in Codex as $ccc-<name>.
| Skill | Say | What it does |
|---|---|---|
/ccc-full-send |
"full send" | Takes a change to a merged PR unattended, only through the full local gate and an independent reviewer. |
/ccc-full-stop |
"full stop" | The same work, stopping with the PR ready for you to merge. |
/ccc-council |
"council" | Read-only review by models from three providers, with a score table. |
/ccc-sol-loop |
"sol loop" | A bounded review loop with an OpenAI reviewer through the Codex CLI. |
/ccc-file-pr |
"file a PR" | Files a short draft PR that leads with the problem. |
/ccc-takeover |
"take over this PR" | Finishes someone else's PR or branch and keeps their credit. |
/ccc-babysit-pr |
"babysit this PR" | Watches a PR through CI and review bots and fixes real findings. |
/ccc-modernize |
"modernize this PR" | Updates an old PR from its base and judges whether it should still exist. |
/ccc-pr-comments |
"reply to the review comments" | Formats comments an agent posts on your behalf. |
CC Commander changes to upstream are marked "CCC safety:" in each skill: no force-push of any kind, no rebase of pushed branches (the base is merged in), no hook or secret-scan bypass, the full local suite and lint on the exact head before every push, CI read as a signal rather than the gate, PRs open as drafts, and merges only through the full gate. The per-skill decisions are in docs/tstack/OVERLAP.md and docs/tstack/SAFETY.md.
Data & network
Commander runs locally and reads no credentials. Outbound calls:
| Call | Default | Control |
|---|---|---|
Anonymous usage events to mcp.commanderplugin.com |
Off | Opt in with CCC_TELEMETRY=1 |
Update check (GET of the public plugin.json on GitHub) at session start |
On | CCC_OFFLINE=1 |
CI status via your own gh login |
Off | Opt in with CCC_CI_TICKER=1 |
Bundled MCP servers: npx fetches each package from the npm registry when Claude Code first starts it |
On | Remove the server from .mcp.json |
Context7 queries context7.com |
Only when a skill uses it | Remove the server |
CCC_OFFLINE=1 blocks every call Commander makes itself; ccc doctor prints the current mode; "disableAllHooks": true in settings.json turns every hook off. The hosted server is a separate service covered by the privacy policy.
Mirrors
commander/cowork-plugin/is the source plugin.commander/cowork-plugin-codex/is generated bynode scripts/build-codex.js; never edit it by hand.- The Anthropic directory listing is built from the source by
node scripts/build-directory-mirror.jsascc-commander; see docs/DIRECTORY.md.
Docs
BIBLE.md (guide) · CHEATSHEET.md (daily reference) · SKILLS-INDEX.md (all skills) · plugin README · CHANGELOG.md · docs.commanderplugin.com
Support, security, contributing
Bugs and questions: GitHub issues. Vulnerabilities: SECURITY.md. Contributions: CONTRIBUTING.md.
License
MIT. The Intelligence Layer (4 files) adds a Commons Clause. The 20 vendor packages under vendor/ keep their upstream licenses. The PR workflow skills adapted from tstack carry its MIT notice; see NOTICE.md.
