ONE ORCHESTRATOR.
MANY VIRTUOSOS.
ZERO CHAOS.
A production-ready Claude Code plugin. Five specialised AI agents. Model routing by task. Quality gates at every stage. Loop detection. Destructive command blocking. No Node, no Python.
WHY KLEIBER. THE OPINIONATED LAYER CLAUDE CODE LEFT OUT.
Kleiber makes them shippable.
Preconfigured Roles
Five battle-tested prompts. Architect, Frontend Engineer, Backend Engineer, Validator, Scribe. Stop rewriting role setup every run.
Model Routing
Opus for architecture. Sonnet for implementation. Haiku for validation. Cheapest reliable tier for every step. Token spend drops without quality loss.
Ralph Loop Guard
When an agent makes the same mistake three times in a row, it gets pulled. No infinite token-burning loops at 3 AM.
Quality Hooks
Tests, type checks, lint, and brand-drift checks at every stage. A teammate cannot mark a task complete if validation fails.
Destructive Block
rm -rf, git push --force, DROP TABLE. Pre-execution. Not after the bloodbath.
Delegate Mode
Lead agent coordinates only. Never implements. Scope boundaries stay crisp; no role drift; no "let me just fix this real quick".
FIVE ROLES. FIXED RESPONSIBILITIES. CLEAR HANDOFFS.
| Role | Model | Writes Code | Purpose |
|---|---|---|---|
| Architect | OPUS | NEVER | Design, review, approve. Coordinates the team in delegate mode. |
| Engineer-Frontend | SONNET | YES | UI components, state, accessibility, tests. |
| Engineer-Backend | SONNET | YES | Routes, queries, business logic, integration tests. |
| Validator | HAIKU | READ-ONLY | Runs tests, type checks, linting. |
| Scribe | HAIKU | DOCS ONLY | Docs, changelog, decision records. |
| Brand-Analyst | SONNET | DOCS ONLY | Probes ChatGPT / Gemini / Perplexity. Scores brand visibility across AI models. |
| Brand-Architect | OPUS | DOCS ONLY | Generates optimisation blueprints. Read-only on source; proposals only — humans approve. |
SIX HOOKS. STDIN JSON. EXIT 2 BLOCKS.
RALPH LOOP GUARD
Agents repeat themselves. When the same error fires three times in a row, the loop guard pulls the teammate, posts a structured report, and routes the work to a different role. Token burn stops at try #4 instead of try #400.
AI VISIBILITY AUDITING
Two extra agents — brand-analyst and brand-architect — probe ChatGPT, Gemini and Perplexity to score how your product shows up in AI answers, then write an optimisation blueprint to docs/.
Source code is read-only to these agents. Every proposal is a draft humans approve. The brand-drift-check pre-hook blocks any edit that strips the Brand Identity block from CLAUDE.md.
EVERYTHING PRELOADED. NOTHING TO STITCH.
Model routing rules, task decomposition, quality gates, coordination patterns, and anti-pattern detection — everything the Architect uses to keep the team aligned.
The methodology behind /kleiber-brand — probe templates, scoring rubric, optimisation playbook. BrandMind-style auditing, built in.
PREREQUISITE: ENABLE EXPERIMENTAL AGENT TEAMS.
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}
START CLAUDE CODE. RUN THE COMMAND.
Implement a payment processing module with Stripe integration.
Review PR #142 with security, performance, and test-coverage teammates.
Investigate a production bug with competing hypotheses and cross-challenge.
CHEAPEST RELIABLE TIER. EVERY TIME.
Architecture, security, compliance.
Feature implementation, default tier.
Validation, docs, boilerplate.
WHAT KLEIBER WILL NOT DO YET.
- →No session resumption: teammates are not restored by /resume or /rewind.
- →One team per session: teammates cannot spawn additional teams.
- →Task status may lag and require lead-agent nudges.
- →Split-pane requires tmux or iTerm2. In-process mode works everywhere.