Agent team orchestration for Claude Code. Structured roles, model routing, quality gates, and loop prevention.

Named after conductor Carlos Kleiber — one orchestrator, many virtuosos, flawless coordination.

Website: kleiber-site.vercel.app | By: DevGap | AI Plumber

Anthropic shipped native agent teams in Claude Code. Kleiber adds the opinionated layer that makes them production-ready:

• Pre-configured roles — 7 agent roles with battle-tested spawn prompts so you don't rewrite them every session
• Model routing — Opus for architecture, Sonnet for implementation, Haiku for validation. Cheapest model that can reliably do the job.
• Quality hooks — 6 hooks enforce tests, types, safety, and brand consistency at every stage of the pipeline
• Loop detection — Ralph Loop Guard stops agents after 3 repeated errors
• Destructive command blocker — Blocks rm -rf, git push --force, DROP TABLE before execution
• Delegate mode enforced — Lead agent coordinates only, never implements
• AI Visibility Auditing — BrandMind-style probes that score how ChatGPT, Gemini, and Perplexity perceive your product, with automated optimization blueprints

Agent teams are experimental and disabled by default. Enable them first:

// Add to your Claude Code settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Or set it in your shell:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

# Add the marketplace
/plugin marketplace add Devgapperk/kleiber

# Install the plugin
/plugin install kleiber@kleiber-marketplace

Slash commands:

• /kleiber — creates and orchestrates agent teams with a single command
• /kleiber-brand — runs a full AI brand visibility audit

7 agent roles with pre-configured spawn prompts and model routing:

6 hooks enforcing quality at every stage:

2 skills:

• Orchestration — model routing rules, task decomposition protocol, quality gates, and anti-pattern detection
• Brand Visibility — AI probe patterns, scoring rubric, gap detection, and optimization blueprint patterns

Start Claude Code and use the /kleiber command:

/kleiber Build user authentication with login, signup, and password reset

Or describe the team you want directly:

Create an agent team to build [feature]. Spawn architect (opus), frontend
engineer (sonnet), backend engineer (sonnet), validator (haiku), scribe (haiku).
Use delegate mode.

Keyboard shortcuts once your team is running:

Full feature build:

/kleiber Implement a payment processing module with Stripe integration

Parallel code review:

Create an agent team to review PR #142. Three reviewers: security, performance,
test coverage. Have them report independently.

Bug investigation:

Users report [symptom]. Create a team with 4 teammates investigating competing
hypotheses. Have them challenge each other's theories.

Kleiber includes BrandMind-style AI visibility auditing that measures how ChatGPT, Gemini, and Perplexity perceive and recommend your product.

What it does:

1. Brand Analyst probes each AI model with 5 structured prompts (identity, category, weakness, competitor, recommendation)
2. Scores responses across 4 dimensions: Mention (0-25), Accuracy (0-25), Recommendation Rank (0-25), Sentiment (0-25)
3. Detects gaps — displacement, accuracy, coverage, category, and authority
4. Source tracking — extracts citation domains from AI responses (especially Perplexity), categorizes them (UGC, Editorial, Corporate, Reference), and flags missing brand presence
5. Historical trends — appends scores to docs/brand-visibility-history.json after every audit so you can track improvement over time
6. Brand Architect generates an optimization blueprint with specific content patches
7. Brand drift hook warns in real-time when doc edits contradict your visibility baseline

Setup:

# Set one or more API keys (missing keys are skipped gracefully)
export OPENAI_API_KEY=sk-...
export GEMINI_API_KEY=...
export PERPLEXITY_API_KEY=pplx-...

Run an audit:

/kleiber-brand
/kleiber-brand --competitors "Cursor, Cline, Aider"
/kleiber-brand --models "chatgpt,perplexity"
/kleiber-brand --re-audit
/kleiber-brand --trend

Output:

• docs/brand-visibility-report.md — scores, gap table, source analysis, raw probe responses, and optimization blueprint
• docs/brand-visibility-history.json — append-only score history for trend tracking
• docs/brand-identity-template.md — template for adding a Brand Identity block to your CLAUDE.md

The ## Brand Identity section in CLAUDE.md serves as the ground truth that probe responses are scored against. See docs/brand-identity-template.md for the format.

• Claude Code CLI
• Agent teams enabled (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
• jq (used by hook scripts)
• Git

Agent teams are in research preview. Current constraints:

• No session resumption — /resume and /rewind won't restore teammates. Spawn new ones.
• One team per session — Teammates can't spawn their own teams.
• Task status can lag — Teammates may finish without marking tasks complete. Nudge the lead.
• Split-pane needs tmux/iTerm2 — Not supported in VS Code terminal, Windows Terminal, or Ghostty. In-process mode works everywhere.

Built by DevGap. Powered by Claude.

Part of the AI Plumber governance-first framework by Koen Van Lysebetten.