GodMode Core for Claude Code

currently published as CC_GodMode

"What happens when an AI system is used to improve itself?"

You're looking at the answer.

Not open source. Free for private, non-commercial use. Forks & pull requests on GitHub are welcome — re-hosting on third-party platforms/marketplaces is prohibited and attribution is required. See LICENSE and NOTICE.

Choose your GodMode

Product	Best fit	What it provides
GodMode Core for Claude Code — this repository (currently CC_GodMode)	You operate your own Claude Code setup	Self-installed orchestrator rules, 14 agents, 14 skills, hooks and release tooling, under its own license terms
GodMode Core for Codex	You operate your own Codex setup	The separately maintained Codex workflow package
GodMode Pro by Nerdsmiths	You want an integrated application and assisted setup	A separate proprietary desktop application for project control, result review, maintained integrations, and scoped onboarding and support

Core has no package subscription; Claude access and usage are paid separately. Free availability for private, non-commercial use does not make it open source and does not change the license. Pro has its own availability, pricing, and service terms; see the Nerdsmiths landing page.

---

The System That Builds Itself

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.

---

Install in 30 Seconds

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

Windows (PowerShell):

git clone https://github.com/cubetribe/ClaudeCode_GodMode-On.git
cd ClaudeCode_GodMode-On
.\scripts\apply-global-claude-setup.ps1

One run installs everything into ~/.claude/: all 14 agents, all 14 skills, the scripts, the templates, LICENSE and NOTICE — and, since v8.7.0, the hooks themselves, merged into ~/.claude/settings.json (your own settings, model, permissions, custom hooks, survive the merge untouched). Without this step, nothing enforces the rules below; a Claude Code session with no hooks wired is a session where @api-guardian never fires and the deterministic checks never run.

Activate GodMode per project — on purpose, not globally. The installer does not write a global ~/.claude/CLAUDE.md. It places the orchestrator template at ~/.claude/templates/CLAUDE-ORCHESTRATOR.md, and you copy it into each project you want GodMode to run in:

cd your-project
cp ~/.claude/templates/CLAUDE-ORCHESTRATOR.md ./CLAUDE.md
claude

That's deliberate: GodMode opts in per project instead of taking over every Claude Code session on the machine. Done. The orchestrator is active in this project. → MCP servers and prompt-based fallback: Installation Guide.

---

Update

Already installed? Same command, every time:

git pull && ./scripts/apply-global-claude-setup.sh

There is no separate updater — the installer is idempotent, backs up ~/.claude/settings.json before touching it, and prints the before/after version so you can see what changed. If nothing changed, it says so and stays quiet.

---

Verify your install

node scripts/verify-install.js

Checks that all 14 agents and all 14 skills are present under ~/.claude/, that every hook from config/claude-settings.json is wired into ~/.claude/settings.json and points at a file that still exists (a hook pointing at nothing is the same as no hook), that ~/.claude/templates/CLAUDE-ORCHESTRATOR.md, LICENSE, and NOTICE are in place, and that your installed version matches the repo's VERSION. Exit code 0 means the installation is complete; exit code 1 lists exactly what's missing.

MCP servers are a separate, deliberate step. ./scripts/install-mcps.sh is not part of the install/update command and stays that way. playwright is the one that matters for @tester's browser-based UX gate (Core Rule 6); without it, the UX gate falls back to human review instead of failing the sprint. That's a graceful degradation, not a missing dependency.

---

Daily Usage — Two Steps

The orchestrator loads from CLAUDE.md automatically. But its headline power — parallel-first dynamic workflows — only fans out when Ultracode is on, and Ultracode is session-scoped. So each new session is two steps (verified on Claude Code 2.1.286; use 2.1.28x or newer):

Step 1 — Pick the model, turn on Ultracode. Once per session, by command:

/model opus         # Opus 5.5 — the recommended default, the model this system is tuned for
/effort xhigh       # optional: raise reasoning effort for hard tasks
/effort ultracode   # separate session-only switch: automatic dynamic workflows, effort unchanged

ultracode is not an effort level. It is a switch that turns on automatic dynamic workflows for the session and leaves your effort setting alone. It needs dynamic workflows enabled in /config, and availability depends on the model. For a one-off, put the keyword ultracode in a single prompt instead.

Optional upgrade: /model best resolves to Fable 5.1 where your org has access (about 2.5× Opus 5.5's list price), otherwise to Opus 5.5. Nothing depends on it.

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).

---

What Is This?

CC_GodMode transforms Claude Code into a self-orchestrating multi-agent development team — driven by an Opus 5.5 orchestrator (/model opus) with ultracode switched on, 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:
  → Plans the work and splits it into sprints with write-scope ownership
  → Creates the report folder
  → Delegates to @architect for design
  → Delegates to @api-guardian for API impact
  → Delegates to @builder for implementation
  → Deterministic hook runs typecheck/lint/tests/build (always, ~0 context on success)
  → @tester checks UX quality, only if the sprint declared `ux_gate: auto`
  → @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."

---

Parallel-First & Ultracode

v8.0.0's headline: parallelization is the default, not an afterthought.

• Orchestrator tuned for Opus 5.5 (/model opus) with ultracode on — ultracode switches on automatic dynamic workflows for substantive tasks. The best alias resolves to Fable 5.1 where your org has access (about 2.5× Opus 5.5's list price) — an optional upgrade, 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.

---

The Agents

14 specialists, each with one job, a model assignment, and effort tuning:

• 7 core — @researcher @architect @api-guardian @builder @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, a deterministic hook (typecheck, lint, tests, build) always runs and costs no context on success. @tester (UX) runs only when the sprint declared ux_gate: auto; @security runs on security surfaces; a /code-review pass is pulled when risk or doubt warrants it. No standing agent re-reads the diff a compiler already checked.

→ Full roster, quality gates, and workflows: The Agents.

---

The Rules

1. 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)
2. Delegate when it pays — Delegate large, genuinely independent, parallelizable work, or work needing a specialist's tools or a separate write scope; do what fits in a handful of tool calls yourself; never delegate to double-check your own work (that is what the deterministic checks are for). Smart Routing is the default — risk-based, Full-Gates for high-risk signals
3. Architecture gate (split) — Inline brief for small/medium; @architect (Opus) for new modules / breaking changes
4. @api-guardian is MANDATORY — For any API/schema/type change (enforced by hook)
5. Verification matches the evidence — After @builder, a deterministic hook (typecheck/lint/tests/build) always runs; @tester or @security only run where they open evidence @builder didn't have; a /code-review pass is pulled on risk or doubt
6. The UX gate is declared, not assumed — Every sprint carries ux_gate: auto | human | skip, decided once at planning (default human); under auto, @tester screenshots every page at 3 viewports (375×667 / 768×1024 / 1920×1080)
7. No Skipping within the selected path — Smart Routing picks the minimal set; that set executes fully. Whatever it leaves out is logged in the Routing Log, not silently dropped
8. Sprint-scoped reports & single-writer hot files — parallel agents get disjoint write scopes; VERSION/CHANGELOG.md have exactly one writer (the release tooling / @scribe)
9. NEVER push without permission — Applies to ALL agents
10. Release invariant, machine-checked — VERSION == CHANGELOG == tag == GitHub release, enforced locally and in CI

---

Documentation

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 14-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:

• REPORT_TEMPLATES.md · CONTEXT_SCOPE_POLICY.md · SECURITY_TOOLING_POLICY.md

---

FAQ

Q: Why 14 agents? A: 7 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 happened to @validator? A: It was dissolved, not replaced. Its deterministic part (typecheck, lint, tests, build) moved into a hook that runs after every @builder pass and costs no context on success. Its judgment part (code review) is pulled on demand via /code-review instead of a standing agent re-reading the same diff.

Q: What does @tester do now? A: UX quality (E2E, visual, a11y, perf) — but only when the sprint declares ux_gate: auto. It's the one true writer-verifier left in the system because it opens evidence (browser, screenshots) @builder never had.

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.

---

Version

CC_GodMode v9.0.0

What's in the box:

• 14 agents (7 core + 1 security gate + 6 department) with effort-field budget tuning
• 14 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
• Evidence-matched verification — deterministic hook always, @tester opt-in via ux_gate, @security on security surfaces, /code-review on risk
• Smart Routing by default (~30–50% token savings vs. old always-Full-Gates)
• Version-at-release workflow — VERSION is written once by the release tooling; the release invariant is machine-checked

See the CHANGELOG for the full history.

---

Credits

Dennis Westermann (www.dennis-westermann.de) Years of suffering, distilled into this repo. Now the repo improves itself. Was it worth it?

---

License

Proprietary License — Copyright (c) 2025-2026 Dennis Westermann. Free for private, non-commercial use; commercial use requires written permission. Attribution is required for every permitted copy. Forks and pull requests on GitHub are expressly welcome; re-hosting, mirroring, or redistribution on third-party platforms is prohibited.

See LICENSE and NOTICE for the full terms.

---

Made with mass sleep deprivation

The experiment continues.

⭐ Star if you're not too unsettled ⭐