Claude Code Skill v2.21.0 MIT License

AI Consultants

A harness for every question. A panel of 15 frontier models that writes its own playbook per question — fan out, debate to convergence, cross-examine under adversarial review, or run a tournament — and checks its work before it reaches you.

View on GitHub npm
Supported agents: Claude Code, Codex CLI, Gemini CLI, Cursor, Copilot, Windsurf, Aider, Amp, Kimi, Standalone Bash

Not a fixed pipeline. A dynamic deliberation engine.

AI Consultants convenes a panel of up to 15 frontier models — each with a distinct persona — and, crucially, decides how they should deliberate for the question in front of them. It classifies the question, picks an orchestration shape, and iterates until the answers converge: the way a workflow builds a harness for the task at hand.

Simple questions get a quick read. Hard ones loop through debate until the panel converges. Security questions are stress-tested by an adversarial gate that tries to refute the leading answer. “Which approach?” runs a tournament and crowns a winner. The process fits the problem — it isn’t hardcoded.

The result: one recommendation that survived scrutiny, with the points of agreement, the dissent, and the path it took to get there all on the table.

Everything you need

Built for developers who want more than a single opinion.

01

15 AI Consultants

Gemini, Codex, Mistral, Kilo, Cursor, Aider, Amp, Kimi, Claude, Qwen3, GLM, Grok, DeepSeek, MiniMax, and Ollama. Each with a unique persona.

02

Automatic Synthesis

All responses are combined into a single weighted recommendation with confidence scoring and consensus analysis.

03

Multi-Agent Debate

Consultants critique each other across multiple rounds. Positions evolve. Anonymous peer review identifies the strongest arguments.

04

Smart Routing

Questions are classified by category and routed to the most relevant consultants. Security questions go to security experts first.

05

Quality Tiers

Premium, standard, and economy model tiers. Choose max_quality for critical decisions or fast for quick checks.

06

Cost Control

Budget enforcement, semantic caching, cost-aware routing, and response limits. Stay within budget without sacrificing quality.

07

Token-Efficient Context

Share large files without burning the consultants' context window. AST-based skeleton extraction, category-aware project tree, and PRIMARY/CONTEXT relevance tags let you point the panel at what matters.

08

Dynamic Orchestration

A planner picks how the panel deliberates per question — a convergence loop that iterates until answers agree, an adversarial refutation gate for security, a tournament to pick one winner, or an exhaustive sweep. Not a fixed script.

09

Measured Roster

The panel calibrates itself. An uncorrelated-value audit flags consultants that only echo the others, while blind peer-review and pairwise-judge Elo derive each model's intelligence, taste, and cost from real runs — not hand-tuned guesses.

From question to recommendation

Classify Plan shape Fan out Deliberate & converge Synthesize
01

Classify & plan

Your question is categorized — architecture, security, performance, code review — and scored for complexity and intent. From that, the planner picks an orchestration shape: how the panel should deliberate this specific question.

02

Fan out

The relevant consultants are queried in parallel. Each returns a structured response — summary, analysis, pros/cons, and a 1–10 confidence score — from its own persona and focus area.

03

Deliberate to convergence

The chosen shape takes over. Converge loops debate rounds until the panel agrees (not a fixed count). Adversarial sends the leading answer through a refutation gate. Tournament pits approaches head-to-head. Exhaustive keeps probing until no new angle appears.

04

Synthesize

Consensus is scored, confidence intervals computed, and one weighted recommendation is produced — with the dissent and the convergence path it took, so you see not just the answer but how much to trust it.

15 consultants, 15 perspectives

Each consultant has a distinct persona that shapes their analysis. The invoking agent is automatically excluded to prevent self-consultation.

CLI-Based — 11 consultants
Google Gemini
The Architect
Design, scalability, enterprise
OpenAI Codex
The Pragmatist
Simplicity, proven solutions
Mistral Vibe
The Devil's Advocate
Edge cases, risks, security
Kilo Code
The Innovator
Creative, unconventional
Cursor
The Integrator
Full-stack perspective
Aider
The Pair Programmer
Collaborative coding
Amp
The Systems Thinker
System design, interactions
Kimi
The Eastern Sage
Holistic, balanced views
Claude
The Synthesizer
Big picture, synthesis
Qwen3
The Analyst
Data-driven, metrics
MiniMax
The Pragmatic Optimizer
Performance, efficiency
API-Based — 3 consultants
GLM
The Methodologist
Structured approaches
Grok
The Provocateur
Challenge conventions
DeepSeek
The Code Specialist
Algorithms, code gen
Local — 1 consultant
Ollama
The Local Expert
Privacy-first, zero API cost

Up and running in 30 seconds

Option A: npx (recommended)

# Run directly — no install needed
npx ai-consultants "How should I structure my auth system?"

# With a preset
npx ai-consultants --preset balanced "Redis or Memcached?"

# Let the doctor recommend a preset for you (v2.13)
npx ai-consultants doctor --suggest-preset --question "How can I prevent SQL injection?"

# Scaffold persistent config at ~/.config/ai-consultants/ (v2.12)
npx ai-consultants init

# Run diagnostics
npx ai-consultants doctor --fix

# Install slash commands for Claude Code
npx ai-consultants install

# Keep the consultant CLIs up to date (v2.21)
npx ai-consultants update-clis

Option B: Claude Code skill (curl | bash)

# Install the skill
curl -fsSL https://raw.githubusercontent.com/matteoscurati/ai-consultants/main/scripts/install.sh | bash

# Ask your first question
/ai-consultants:consult "How should I structure my auth system?"

Option C: Standalone bash

# Clone and set up
git clone https://github.com/matteoscurati/ai-consultants.git
cd ai-consultants
./scripts/doctor.sh --fix

# Run a consultation
./scripts/consult_all.sh "How should I structure my auth system?"
Works with Claude Code, Codex CLI, Gemini CLI, Cursor, Copilot, Windsurf (via SkillPort), Aider (via AGENTS.md), Amp, Kimi, and standalone bash.

Presets and strategies

Presets
Preset Consultants Use Case
minimal2 (Gemini + Codex)Fast, cheap
fast2 + economy models, no debateQuick checks
balanced4 (Gemini, Codex, Mistral, Kilo)Standard consultations [default]
medium4 + standard models + light debateGeneral questions
thorough5 (+ Cursor)Comprehensive analysis
cost-cappedBudget-conscious optionsMinimal API costs
securitySecurity-focused + debateSecurity reviews
high-stakesAll + debateMaximum rigor for critical decisions
max_qualityAll + premium models + debate + reflectionCritical decisions
localOllama onlyFull privacy, no API calls
Synthesis Strategies
Strategy Description
majorityMost common answer wins (default)
risk_averseWeight conservative responses higher
security_firstPrioritize security considerations
cost_cappedPrefer simpler, cheaper solutions
compare_onlyNo recommendation, just comparison
Key Environment Variables
# Core features
ENABLE_DEBATE=true           # Multi-agent debate
ENABLE_SYNTHESIS=true        # Automatic synthesis
ENABLE_SMART_ROUTING=true    # Intelligent consultant selection
ENABLE_PANIC_MODE=auto       # Automatic rigor for uncertainty
ENABLE_STANCE_CONSENSUS=true # Exact-match stance consensus (v2.21, opt-in)

# Defaults
DEFAULT_PRESET=balanced
DEFAULT_STRATEGY=majority

# Ollama (local models)
ENABLE_OLLAMA=true
OLLAMA_MODEL=hf.co/prithivMLmods/VibeThinker-3B-GGUF

# Cost management
MAX_SESSION_COST=1.00        # Budget limit in USD
WARN_AT_COST=0.50
Copied to clipboard