You're looking at the answer.
Welcome to the machine shop. Except the machines are building themselves — and they're getting better at it.
CC_GodMode v7.0.0, v7.1.0, and v8.0.0 weren't built by hand. They were planned, architected, implemented, validated, documented, and shipped by the exact agents defined in this repo. v8.0.0 went further: it was produced by a parallel dynamic workflow — a dozen edit subagents fanned out at once, each adversarially verified by another. The orchestrator delegated, gated, and opened the PR itself.
That's not marketing copy. That's what happened — the full story is here.
CC_GodMode is a Claude Code plugin that turns your workflow into a self-orchestrating multi-agent system.
macOS / Linux:
git clone https://github.com/cubetribe/ClaudeCode_GodMode-On.git
cd ClaudeCode_GodMode-On
./scripts/apply-global-claude-setup.sh
./scripts/apply-global-claude-setup.sh --check # Verify installationWindows (PowerShell):
git clone https://github.com/cubetribe/ClaudeCode_GodMode-On.git
cd ClaudeCode_GodMode-On
.\scripts\apply-global-claude-setup.ps1
.\scripts\apply-global-claude-setup.ps1 -Check # Verify installationThen activate in any project:
cd your-project
cp ~/.claude/templates/CLAUDE-ORCHESTRATOR.md ./CLAUDE.md
claudeDone. The orchestrator is active. → Full setup, MCP servers, and prompt-based fallback: Installation Guide.
The orchestrator loads from CLAUDE.md automatically. But its headline power — parallel-first dynamic workflows — only runs at Ultracode effort, and Ultracode is session-scoped. So each new session is two steps:
Step 1 — Turn on Ultracode. Set it once per session, in the effort selector at the bottom of Claude Code, or by command:
/model best # resolves to Opus 4.8 — the model this system is optimized for
/effort ultracode # xhigh reasoning + automatic parallel dynamic workflows
Step 2 — Say what you want, prefixed with GodMode:
GodMode: New Feature: user authentication with JWT
GodMode: Bug Fix: cart total miscalculates with discount codes
GodMode: Research: best approach for real-time sync in React 18
You say what you want — the system figures out which agents to call, in what order, at what cost. Ultracode is the one thing that does not persist across sessions, so make Step 1 a habit; skip it and GodMode still orchestrates and gates correctly, it just won't fan out to its full parallel width. The trigger GodMode: is case-insensitive (GODMODE: works too).
CC_GodMode transforms Claude Code into a self-orchestrating multi-agent development team — driven by the best / Claude Opus 4.8 orchestrator at ultracode effort, fanning work out across parallel Claude Code subagents for implementation, validation, and documentation.
You say WHAT. The AI figures out HOW.
You: "I need user authentication with JWT"
Orchestrator:
→ Analyzes the request and determines the version bump
→ Creates the report folder
→ Delegates to @architect for design
→ Delegates to @api-guardian for API impact
→ Delegates to @builder for implementation
→ @validator checks code quality ┐ in parallel
→ @tester checks UX quality ┘
→ @scribe documents everything
→ @github-manager opens the PR
You: *drinks coffee*
| Without CC_GodMode | With CC_GodMode |
|---|---|
| You: "Design the feature" | You: "Build Feature X" |
| You: "Now implement it" | ☕ |
| You: "Check the types" | ☕ |
| You: "Update the consumers" | ☕ |
| You: "Write the docs" | ☕ |
| You: "Did I forget something?" | AI: "Done. Here's the report." |
v8.0.0's headline: parallelization is the default, not an afterthought.
- Orchestrator optimized for Opus 4.8 at ultracode — xhigh reasoning plus automatic dynamic workflows for substantive tasks. The
bestalias resolves to Opus 4.8; if your org has access to a higher tier it is picked up automatically — strictly optional, no feature depends on it. - Fan-out by default — independent units (multi-file edits, audits, migrations, multi-angle research) spawn parallel subagents in a single message; the orchestrator fans in and synthesizes their verdicts.
- Dynamic-workflows escalation — when a job outgrows ~10 concurrent subagents, it escalates to tens-to-hundreds of subagents with adversarial verification (agents try to refute each other's findings). See
skills/dynamic-workflows/. - Smart Routing stays the default — risk-based, minimal-agent paths; ~30–50% token reduction vs. always-Full-Gates. Parallel is faster, not cheaper, so max-parallel is a deliberate opt-in.
→ Deep dive: Architecture · cost model: Agent Model Selection.
15 specialists, each with one job, a model assignment, and effort tuning:
- 8 core —
@researcher@architect@api-guardian@builder@validator@tester@scribe@github-manager(always available). - 1 security gate —
@security, for secrets, injection, auth/authz, crypto, and dependency review. - 6 department —
@ci-security-guardian@docs-dx@quality-operations@runtime-platform@workflow-design@workspace-governance(activate when their domain is in scope).
After @builder, the dual quality gates — @validator (code) and @tester (UX) — run in parallel and both must pass before @scribe documents and ships.
→ Full roster, quality gates, and workflows: The Agents.
- Plan-First — Non-trivial work starts with a plan (
plans/vX.Y.Z/) split into sprints with explicit write-scope ownership; the version is decided at release, never at work start (ADR-004) - Smart Routing default — Risk-based routing; Full-Gates for high-risk signals
- Architecture gate (split) — Inline brief for small/medium; @architect (Opus) for new modules / breaking changes
- @api-guardian is MANDATORY — For any API/schema/type change (enforced by hook)
- Dual Quality Gates — Both @validator AND @tester must pass (parallel execution)
- @tester MUST create screenshots — Every page tested at 3 viewports
- No Skipping within the selected path — Smart Routing picks the minimal set; that set executes fully
- Sprint-scoped reports & single-writer hot files — parallel agents get disjoint write scopes;
VERSION/CHANGELOG.mdhave exactly one writer (the release tooling / @scribe) - NEVER push without permission — Applies to ALL agents
- Release invariant, machine-checked —
VERSION == CHANGELOG == tag == GitHub release, enforced locally and in CI
Guides (start here):
- Installation Guide — script install, MCP servers, prompt-based fallback, recovery
- Architecture — parallel-first orchestration, file structure, dual-location model, the hook
- The Agents — the 15-agent roster, quality gates, workflows, and modes
- The Story & Design Philosophy — how (and why) the system builds itself
- ROADMAP.md — living roadmap · plans/ — active plans & sprint files
Reference:
- CHANGELOG.md — full version history
- AGENT_MODEL_SELECTION.md — model aliases, pricing, and cost optimization
- AGENT_ARCHITECTURE.md — dual-location install/update/verify procedures
- orchestrator/AGENTS.md — agent registry & handoff matrix
- orchestrator/VERSIONING.md — the release law (single source of truth, invariant, procedures)
- orchestrator/WORKFLOWS.md · MODES.md · QUALITY-GATES.md · VERSIONING.md · META-DECISIONS.md
Policies:
Q: Why 15 agents?
A: 8 core cover the standard workflow, 1 optional @security gate activates for security-sensitive changes, and 6 optional department agents activate only when their domain is in scope. Separation of concerns — each agent has ONE job.
Q: What's the difference between @validator and @tester?
A: @validator = code quality (TypeScript, tests, security). @tester = UX quality (E2E, visual, a11y, perf). They run in parallel.
Q: Can agents push without my permission? A: No. "NEVER git push without permission" is enforced across all agents.
→ More, plus the origin story: The Story & Design Philosophy.
CC_GodMode v8.6.0
What's in the box:
- 15 agents (8 core + 1 security gate + 6 department) with effort-field budget tuning
- 13 skills for workflows, quality gates, release, research, API changes, modes, teams, bootstrap, and dynamic workflows
- Parallel-first orchestration — fan-out by default, dynamic-workflows escalation with adversarial verification
- Dual quality gates (parallel execution for speed)
- Smart Routing by default (~30–50% token savings vs. old always-Full-Gates)
- Version-first workflow with automated checks
See the CHANGELOG for the full history.
Dennis Westermann (www.dennis-westermann.de) Years of suffering, distilled into this repo. Now the repo improves itself. Was it worth it?
Proprietary License — Private use permitted. Commercial use requires permission.
Copyright (c) 2025 Dennis Westermann
Made with mass sleep deprivation
The experiment continues.
⭐ Star if you're not too unsettled ⭐