claude — ~/projects/web-app — commander v7.4.2 ⌘K

CC Commander — Master Claude Code instantly. Core free forever.

 ██████╗ ██████╗ ██████╗
██╔════╝██╔════╝██╔════╝
██║     ██║     ██║     
██║     ██║     ██║     
╚██████╗╚██████╗╚██████╗
 ╚═════╝ ╚═════╝ ╚═════╝C O M M A N D E R

Master Claude Code instantly. 82 plugin skills. 467 ecosystem skills. 549 total skills. 22 agents. 23 lifecycle hooks x 44 handlers. 19 vendors. One install, zero config, core free forever.

CC Commander demo — /ccc menu, /ccc-plan, /ccc-build, /ccc-fleet running live in Claude Code Desktop

$ install commander read the docs 𝕏 Share

~/projects/web-app — try it

› try a real command — claude responds in this box

v7.4.2 · MIT licensed· last shipped 2026-07-22· built by Kevin Z

runs in

Same plugin, every editor.

12 surfaces · MCP under the hood

Claude Code Desktopnative

Claude Code CLInative

Claude Coworkbeta

Cursorvia MCP

Windsurfvia MCP

OpenAI Codexvia MCP

VS Codevia MCP

JetBrainsvia MCP

Zedvia MCP

Continuevia MCP

Clinevia MCP

+ MCP clients7 more

tailored to you

Who's shipping?

Ship features without rebuilding the workflow.

You have product velocity to protect. Commander gives your team a shared methodology, severity-rated reviews, and parallel worktrees so two engineers don't redo the same plan.

top skills/ccc-saas · /ccc-plan · /ccc-ship

core agentsbuilder · reviewer · saas-specialist · security-auditor

typical install5 min · zero config

Run 4 client projects in parallel without dropping the rope.

The fleet-worker dispatches on git worktrees so each project stays clean. Commander remembers each client's CLAUDE.md, conventions, and review rubric — context-switching becomes a slash command.

top skills/ccc-fleet · /ccc-design · /ccc-makeover

core agentsfleet-worker · designer · technical-writer · architect

typical installper-repo · 30 sec each

One person. Full team output.

You don't need a project manager. You need 22 of them, on tap. Commander ships the discipline that solo devs end up reinventing on year three of a project.

top skills/ccc · /ccc-plan · /ccc-doctor

core agentsarchitect · builder · reviewer · debugger

typical install1 command · forever free

The audit trail your security team has been asking for.

Hooks log every tool use. PreToolUse blocks unknown commands. The security-auditor flags SEV-1 before deploy. MIT licensed, self-hosted, opt-out anonymous telemetry — pass the SOC2 review on the first try.

top skills/ccc-security · /ccc-audit · /ccc-doctor

core agentssecurity-auditor · architect · qa-engineer · reviewer

typical installself-hosted · telemetry opt-out

commander · v7.4.2 · status all systems online

plugin skills

specialist agents

23 × 43

hooks · handlers

467

ecosystem skills

vendors

why CC commander

Claude Code is powerful. CCC makes it instantly usable.

Stock Claude Code is a blank terminal with amnesia. Every other plugin solves one slice. Commander is the first guided AI PM — click-first UX, an intelligence layer that thinks before you do, and the whole ecosystem pre-wired.

[skills]plugin/commander81

Skills

81 click-first /ccc-* slash commands. Every menu is a native chip picker — no typing, no YAML, no syntax to memorize.

13 specialist workflows 11 domain routers

[agents]plugin/commander22

Agents

22 specialist sub-agents with persona voices — architect, reviewer, security-auditor, debugger, designer. Workflow-first orchestration keeps the lead session as the control plane while the Workflow tool and fleet do substantive work.

22 personas brain/hands SDK

[hooks]plugin/commander23 × 43

Hooks

23 lifecycle hooks x 44 handlers. SessionStart through SubagentStop fire automatically — the right agent runs at the right moment, every time.

23 hooks 43 handlers

[v7.4.2]plugin/commandernew

The full-audit release — everything confirmed got fixed

An 11-agent, adversarially-verified audit swept every subsystem, and every confirmed finding shipped. The compaction error is gone: every /compact used to trigger “Hook JSON output validation failed” because the PostCompact hook emitted output Claude Code's validator rejects — the emit layer now guards the validator's real 20-event union (extracted from the CLI binary itself) and the contract test drives every emitter under its real registered events, so this bug class can't ship green again. Mission Control finally tells the truth: finished subagents used to stay “working” forever (78 of 3,337 telemetry rows ever joined) — the roster now joins on agent_id. Keyless everywhere: every shipped surface — MCP guide, Cursor/Windsurf adapters, config templates — now says the same thing: free hosted MCP, no account, no license key, 100-call/mo anti-abuse cap. Plus: the rtk vendor picked up six upstream security fixes, root npm audit is 0, /ccc-connect captures credentials via a hidden-input terminal command (never chat), /ccc-tuneup is reachable by chips, and all 11 domain routers open a picker on bare invocation.

81st skill: /ccc-update 8th Cockpit tab: Prompts 10 Codex mirror events

[v7.2.0]plugin/commanderprior

Commander decks — one switcher, four artifacts

Every Commander artifact now shares a Commander decks switcher, so you can jump between the Cockpit, Mission Control, Usage & Cost, and Safety without leaving the page. Two new decks join the family this release. Usage & Cost (/ccc-usage) is a self-contained artifact showing what your agent work actually costs and what delegating to cheaper models saved you — an honest savings estimate, saved-per-day and cost-per-day trends, and a cost-by-app breakdown. Safety (/ccc-safety) shows what Commander's hooks caught for you: dangerous actions blocked, auto-fixes applied, actions approved, and a redacted tool-failure hotspot list — no raw command output, no secrets, just outcomes.

4 decks, 1 switcher 2 new decks 0 secrets in payloads

[v7.1.0]plugin/commanderprior

Commander Cockpit becomes a product

The Cockpit — every Commander tool inside one self-contained artifact — got a real redesign on the commanderplugin.com coral brand. Browse is now a filterable skill list with a rows/tiles toggle and real-time search, instead of a wall of one-skill "domains" — a clean 14-domain taxonomy, no misleading "core". Enhance turns 11 prompt-improvement strategies into multi-select chips with a GO that appends exactly the ones you pick. And the telemetry is honest now: real per-agent token and cost numbers, recovered from each subagent's own transcript (deduped so streaming turns aren't double-counted, cache reads priced in) — or a plain "—" when the data was never captured, never a fabricated zero. New in this release: a Tools tab that turns the Cockpit into an ecosystem launcher — open Commander surfaces, install the best MIT-licensed companion apps (Nimbalyst, Usage Monitor, Sniffly, Claude Control), or copy a prompt that has Claude build a live dashboard.

14 real domains 11 prompt strategies 5 MIT companion apps

[v6.8.3]plugin/commanderprior

Mission Control: charts, history, complete rosters

Mission Control's Claude+Codex rosters now come with the receipts. A zero-dependency Charts strip — inline SVG, one shared module rendering both the live dashboard and the CSP-safe snapshot artifact — plots cost/day, agents dispatched/day, tasks completed/week, and tool failures/day, theme-safe with a clean zero-state instead of a broken axis. A new claude-mem History panel adds a read-only, titles-only timeline over your claude-mem DB (opt-in — hidden entirely when claude-mem isn't installed). And the Agent Roster is finally complete: Codex Desktop's hook surface drops several events CCC relies on for real start records, so Codex work showed up in the Live Feed but never the roster — now those rows are synthesized from delegation events, marked "inferred" with a dimmed badge and tooltip, and never claim token/cost data that isn't real. All backed by a metrics.jsonl daily rollup (ccusage-sourced, Claude + Codex) and a per-skill Claude/Codex breakdown, with a TTL cache so the dashboard scales with more concurrent viewers.

4 chart types Claude + Codex complete rosters

[v6.8.2]plugin/commanderprior

Commander Cockpit

Every Commander tool, clickable, inside one artifact. Published from /ccc-browse → "🎛️ Open the Commander Cockpit": a self-contained, CSP-safe page with five tabs — Browse (78 plugin + 467 ecosystem skills, 22 agents, instant search, job-first row, copy-to-run), Ideas (catalog-grounded next-move generator), Enhance a prompt (six prompt-quality patterns as a live checklist), Analytics (your own agent runs — tokens, cost, 14-day sparkline, top agents, Claude and Codex side by side via source_app), and Agents (roster manager — edit a profile in-page, export it). A new skill-runs logger (the 42nd hook handler) records which skills you actually launch, powering Cockpit's per-skill analytics.

5 tabs, one artifact 42 hook handlers

[v6.8.0]plugin/commanderprior

Commander Mission Control

A visual harness inside Claude Code Desktop — see who's working on what, delegation and data flow, plain-English status of every agent, built for non-coders managing coding agents. /ccc-mission-control opens a live dashboard (agent roster, delegation flow graph, task board, live event feed, an amber awaiting-permission state so unattended fleets never stall silently) or a snapshot artifact you can check from your phone. v6.8.1 added multi-source tracking — agents and events tagged by source app (Claude vs Codex) with a Source filter and colored pill — plus a Suggestions panel for proactive ideas you can promote straight to a Linear ticket.

live dashboard + snapshot mode Claude + Codex source tracking

[v6.7.3]plugin/commanderprior

Relay, SMB ops, spawn hub + the auto prompt fixer

Five new skills across the v6.7.x line. /ccc-relay chains independent, gated spec → build → review sessions with a durable handoff baton (merge stays human-approved). ccc-smb-ops adds a small-business-operations domain (invoicing/AR, cash-flow, payroll, CRM, weekly owner brief). /ccc-broadcast pushes scrubbed build/deploy/CI/cost status to your team's chat or email. /ccc-spawn is a click-first spawn hub — a spawn_task chip in Cowork, /spawn in the CLI. And /ccc-prompt-fix — the auto prompt fixer — diagnoses your prompt against six quality patterns and hands back a tighter rewrite, backed by the integrated claude-code-library (52 Claude Code prompts + a 7-module best-practices course).

77 plugin skills 544 total skills

[v6.6.0]plugin/commanderprior

Engagement Engine

Hook delivery rebuilt on the documented output fields (systemMessage / hookSpecificOutput.additionalContext) so recommendations actually reach the model and the user, not just stdout. Always-on skill suggestions now surface as clickable AskUserQuestion chips once confidence is high, instead of printed text you have to notice. /ccc-claudemd — the new CLAUDE.md auto-auditor — is the 72nd plugin skill. Verifier-separation and mandatory git-worktree isolation now cover every write path (ccc-migrate, ccc-orchestrate, ccc-plan-exec, ccc-fleet).

72 plugin skills 23 hooks × 39 handlers

[v6.4.0]plugin/commandernew

The Fable Method

The operating doctrine distilled from Claude Fable 5, encoded as 12 enforceable gates — never trust a single pass · loops need verifiers, state, and stops · prove before you alarm · context is disposable, state is durable · orchestrate, don't type — so ANY model running CCC, Sonnet, GPT, anything, produces Fable-shaped results. /ccc-fable is the master switch: on arms the doctrine, status reports what's armed, audit self-checks which gates the session is violating, off disarms. Every one of the 71 skills carries a Fable-contract footer enforced by an audit test.

12 gates 71 skills carry the contract

[v6.3.0]plugin/commandernew

Safe delegation

/ccc-yolo-setup turns on vetted safe-YOLO while keeping Plan mode for risky changes. /ccc-ecc selectively loads one ECC skill, agent, or hook without duplicating the full harness. Claude Teams is a hard-on default.

safe-YOLO · Plan mode ECC selective load

[v6.0.0]plugin/commandercore

Smart model routing

Opus 4.8 everyday default. Claude Fable 5 selective deep-mode escalates 4 deep personas (architect · debugger · security-auditor · product-manager) for heavy reasoning — a once-per-day nudge tells you when it pays off. Complexity-scored auto-routing across Haiku/Sonnet/Opus/Fable for subagents keeps costs honest. ccc --savings shows an estimated spend vs an all-Opus baseline (±30%).

Fable 5 deep-mode savings counter · ccc --savings

see it live

Click `/ccc` and watch the orchestrator dispatch.

Three surfaces, one plugin. Pick a chip on the left — the brain routes to the right specialist, hooks fire on lifecycle events, hands ship the diff.

claude — ~/web-app — commander⌘K

you

add stripe checkout to the pricing page

commander

I'll route this through /ccc-saas. Pick a workflow:

$/ccc-saas21 skills · pick one

brain orchestrator picked stripe-checkout

hands dispatching → builder + saas-specialist

hook PreToolUse — verifying scope, deps, secrets

cowork — web-app · sprint 14⌘K

builder · #142 stripe checkout

Scaffolded /api/checkout, wired client SDK, added webhook handler. Ready for review.

reviewer · #143 auth refactor

3 SEV-2 findings on session rotation. Suggested patches attached. /ccc-review --apply

fleet-worker · 4 worktrees in flight

FOR/AGAINST analysis on the email provider choice. Resend vs Postmark vs SES. ETA 2m.

~/web-app — zsh — commander

$ /ccc ┌─ commander ─────────────────────────────────────────────┐ │ pick a workflow │ │ [1] /ccc-plan spec interview → plan.md │ [2] /ccc-build scaffold and ship │ [3] /ccc-review severity-rated diff review │ [4] /ccc-fleet parallel git-worktree dispatch │ [5] /ccc-doctor diagnose · fix · report └─────────────────────────────────────────────────────────┘ › choose [1-5]: 3 ✓ dispatching reviewer agent ✓ hook PreToolUse fired scanning diff: 14 files · 312 lines ⚠ SEV-2 · src/auth/session.ts:84 session rotation missing on privilege escalation ⚠ SEV-3 · api/checkout.ts:21 stripe webhook signature not verified › apply patches? y ✓ 2 patches written, hook PostToolUse green

install demo

One command. Whole ecosystem online.

From cold-start to /ccc in under 30 seconds. No API keys, no signup, no credit card. The plugin marketplace handles the rest.

claude-code · plugin install

$/plugin install commander

→Resolving from KevinZai/commander@v7.4.2

✓Installing commander v7.4.2

✓82 plugin skills activated (/ccc-*)

✓22 specialist agents wired (architect · reviewer · debugger · …)

✓23 lifecycle hooks active (44 handlers)

✓2 MCP servers connected (context7 · sequential-thinking)

▸Ready. Type /ccc to start.

┌────────────────────────────────────────────────────┐ │ /ccc → Build · Review · Ship · Design · Learn │ └────────────────────────────────────────────────────┘

claude code desktop claude code cli cursor windsurf cline · continue · codex same install, every editor.

200 lines vs one slash

The plan you'd write. The plan it writes.

Stock Claude Code is a blank prompt. You write the methodology, the spec template, the review rubric, the hook handlers. Or you type one slash command.

stock claude code~200 lines · CLAUDE.md

# CLAUDE.md
## Persona
You are a senior engineer working on a SaaS product.
You plan first, ship small, and operationalize fixes.

## Workflow for new features
1. Spec interview: ask 5-7 clarifying questions before code.
2. Write a plan.md with: scope, schema, files touched, tests.
3. Wait for approval before any implementation.
4. Implement in vertical slices, one PR at a time.
5. Run reviewer pass: severity-rate every diff.

## Review rubric
SEV-1 = blocks deploy (auth, payments, data loss)
SEV-2 = must fix before merge (race, perf, regression)
SEV-3 = should fix this sprint (polish, dx)

## File conventions
- src/* uses TypeScript strict mode
- API routes return Result<T, AppError>
- Webhooks verify signatures, no exceptions
- Migrations are reversible
- New deps require one-line justification

## Hooks
- Before any tool use, log to .claude/audit.log
- After session end, summarize diffs in NOTES.md
- On error, dispatch debugger sub-agent
- ... (180 more lines)

+ commander1 line · zero config

$ /plugin install commander

✓ 81 skills active
✓ 22 agents wired
✓ 23 hooks × 44 handlers
✓ BIBLE.md methodology loaded
✓ review rubric: SEV-1/2/3

# now type:
$ /ccc-plan add stripe checkout

→ commander interviews you (5 questions)
→ writes plan.md
→ waits for approval
→ dispatches builder + saas-specialist
→ reviewer passes severity check
→ ships

# the methodology is already written.
# the agents are already wired.
# the hooks are already firing.
# you just type slash commands.

stock vs commander

The same Claude Code. A different machine.

Stock Claude Code ships a powerful blank canvas. Commander adds the workflow layer most teams end up rebuilding from scratch — except it’s already built, opinionated, and free.

capability

stock claude code

+ commander

Slash-command workflows

none built-in

81 skills, click-first picker

Specialist sub-agents

roll your own

22 wired personas, day one

Lifecycle hooks

config required

23 hooks × 44 handlers, on by default

Plugin discovery

manual

/ccc-suggest recommends by name

Parallel work

single thread

git-worktree fleet, FOR/AGAINST runs

Persistent methodology

your CLAUDE.md

BIBLE.md + the Kevin Z method

Cost

free

also free, MIT, forever

22 specialist agents

Brain orchestrates. Hands ship.

Each agent embodies a distinct role, model and voice — wired through the Claude Agent SDK sub-agent architecture. Ask once; the right specialist runs every time.

architect

Designs system shape and constraints before a line is written.

reviewer

Severity-rated code review on the current diff or branch.

builder

Scaffolds projects, generates code, executes the plan.

debugger

Reproduces, isolates, patches — never guesses past a stack trace.

security-auditor

Path traversal, cwd validation, dependency vuln triage.

performance-eng

Profiles, benchmarks, trims the bottleneck path.

designer

UI/UX flows, component design, Figma-to-code.

product-manager

Spec interviews, plan files, scope discipline.

technical-writer

Docs, READMEs and changelogs that read like a person wrote them.

content-strategist

Marketing, launch posts, narrative shape for releases.

data-analyst

SQL, dashboards, ad-hoc data spelunking with citations.

devops-engineer

CI, deploys, infra-as-code, rollback playbooks.

qa-engineer

E2E suites, Playwright automation, visual regression.

researcher

Surveys the field; returns options with trade-offs, not opinions.

fleet-worker

Parallel git-worktree runs — fan-out, pipeline, FOR/AGAINST.

typescript-reviewer

Strict types, generics, narrowing, exhaustive checks.

python-reviewer

Type hints, ruff, async patterns, packaging hygiene.

all 81 skills

Browse the index.

Every plugin skill that ships with Commander. Filter by domain, click a card to peek at the YAML manifest. The agent doesn't pick by accident — it picks from this menu.

showing 81 of 81

11 CCC domains

Every workflow has a domain.

Domain routers narrow the 467-skill ecosystem down to the half-dozen moves that matter for what you’re doing right now. Type the slash command, click the chip.

$ls -lh ./domains total 11

drwxr-xr-x62 skillsmay 06/ccc-designUI · UXlanding · components drwxr-xr-x21 skillsmay 06/ccc-saasappsstripe · auth · email drwxr-xr-x54 skillsmay 05/ccc-devopsinfraci · deploy · rollback drwxr-xr-x48 skillsmay 05/ccc-marketinggrowthlaunch · copy · ads drwxr-xr-x42 skillsmay 04/ccc-dataSQL · BIdashboards · pipelines drwxr-xr-x38 skillsmay 04/ccc-testingQAe2e · unit · visual drwxr-xr-x33 skillsmay 03/ccc-securityauditvuln · authn · authz drwxr-xr-x31 skillsmay 03/ccc-mobileiOS · Androidexpo · react-native drwxr-xr-x29 skillsmay 02/ccc-seogrowthschema · sitemap drwxr-xr-x26 skillsmay 02/ccc-researchreconsurvey · trade-offs drwxr-xr-x23 skillsmay 01/ccc-makeoverrefactorUI polish · dx

548 total skills · 81 plugin + 467 ecosystem · click any domain to browse the catalog

23 hooks · 44 handlers

Lifecycle, automated.

Every Claude Code event has a handler attached. Hover the timeline to see what fires when, and which handlers ship by default.

architecture

Brain orchestrates. Hooks fire. Hands ship.

A two-layer SDK pattern with lifecycle automation between. Skills become inputs, agents become tools, hooks become guardrails. You write nothing — it’s already wired.

1. input

/ccc

/ccc-plan

/ccc-review

/ccc-yolo-setup

/ccc-ecc

+ 76 more

↓ 81 skills

2. brain

control plane

workflow-first router

plan-file writer

scope guard

↓ dispatch

3. hands

architect

builder

reviewer

debugger

security-auditor

+ 12 more

↓ lifecycle

4. hooks

SessionStart

PreToolUse

PostToolUse

UserPromptSubmit

SubagentStop

+ 4 more

the Kevin Z method

“Plan first. Ship small.
Operationalize fixes.”

01 — plan first

Spec before code.

Every feature starts with a spec interview. /ccc-plan writes a plan file you can read, review and reject before a token is burned on implementation.

02 — ship small

One commit, one idea.

The reviewer agent enforces it. Diffs that try to do two things get rejected with a severity rating. Boring is the strategy.

03 — operationalize fixes

A bug becomes a hook.

Every regression that lands in production turns into a lifecycle hook. The session never has to remember — it’s already enforced.

14 months of production methodology, 200+ community sources, distilled in BIBLE.md. Ships with the plugin.

core free forever

No paywall. No feature gating.
No license checks.

Every skill, every agent, every hook, every domain router — installable today, in beta, in production, on day one thousand. MIT licensed. It will not get worse.

How it stays free: transparent affiliate links inside /ccc-connect (you see them, you choose) and Kevin Z’s consulting practice underneath. The plugin itself is, and stays, MIT.

quick start · 30 seconds

Three steps. You’re live.

Works in Claude Code Desktop, the CLI, and any MCP-capable IDE. Pick the path that matches where you live — same plugin underneath.

01 — add the marketplace

Point Claude Code at the repo.

Settings → Plugin Marketplace → Add from GitHub. Or run it from the CLI.

$/plugin marketplace add KevinZai/commander

02 — install commander

Pull the plugin.

One command activates 82 plugin skills, 467 ecosystem skills, 22 agents, 23 lifecycle hooks x 44 handlers, and 19 vendors.

$/plugin install commander

03 — start commanding

Type `/ccc`.

The chip picker pops. Build · Review · Ship · Design · Learn · More. Click and go.

$/ccc

faq

Questions, answered.

Is this really free? What’s the catch?

Yes — MIT licensed, core free forever. The plugin itself never gates features. /ccc-connect shows transparent affiliate links to optional MCP integrations (Supabase, Vercel, Notion, etc.) — you see them and choose. Kevin Z’s consulting practice covers the rest.

Do I need an API key or signup?

No. The two bundled MCP servers — context7 and sequential-thinking — are credential-free. Sixteen optional integrations are opt-in via /ccc-connect; only those need keys, and only if you want them.

Does it work outside Claude Code?

Yes. Native plugin install in Claude Code Desktop and the CLI. For Cursor, Windsurf, Cline, Continue, and Codex, point the MCP client at mcp.commanderplugin.com — same 467 ecosystem skills across 11 domains, one URL.

What’s the difference between a skill and an agent?

A skill is a slash command — a workflow you invoke (/ccc-plan, /ccc-review). An agent is a persona with a distinct role, model, and voice that the orchestrator dispatches to. Skills route to agents. The brain/hands pattern.

How does it interact with other plugins?

Commander is a meta-layer, not a walled garden. /ccc-suggest recommends specific third-party plugins by name when they’re the right tool — claude-mem for persistent memory, superpowers for structured thinking, caveman for token-heavy iteration, and a dozen more.

Can I uninstall cleanly?

Yes — /plugin uninstall commander removes everything. Or for the binary install, ./uninstall.sh at the repo root. No background daemons, no system-wide configs, nothing left behind.

What about privacy and telemetry?

The plugin collects anonymous usage telemetry (enabled by default, easy to opt out). It sends event names, OS, Node version, and plugin version — never prompts, file contents, paths, or secrets. Opt out with CCC_TELEMETRY=0 or ~/.commander/config.json → {"telemetry": false}. The hosted MCP endpoint logs tool name, latency, and call count for rate-limiting (100 calls/mo free tier) but never logs prompt content or Claude's responses. This site uses Plausible (cookieless) + PostHog (first-party cookie) for pageview and click analytics — no session replay, no third-party trackers. Full breakdown in our privacy policy. Self-host the MCP server to keep everything on your box.

How do I report a bug or request a skill?

GitHub issues at KevinZai/commander, or DM @commanderplugin on X. Or just run /ccc-doctor — if it’s a known regression, it’ll tell you the workaround. Otherwise it generates a bug report ready to paste.

support cc commander

Keep it free. Stay in the loop.

⭐

Star on GitHub

The fastest way to help. Stars signal traction and bring CC Commander to more developers.

❤️

Sponsor on GitHub

Help keep CC Commander free. Recurring or one-time — every dollar funds the hosted MCP infrastructure.

💼

Custom Claude Code Consulting

Engagements for teams adopting Claude Code at scale. Architecture, onboarding, and agent workflows.