Consult

Multi-agent consensus system. N LLM agents with domain-specific prompts analyze your problem in parallel, critique each other's outputs, iterate until approval threshold is met, then a synthesis agent produces one unified answer.

When to Use This

Single LLM calls optimize for one dimension. Use this when:

• The problem spans multiple domains (security + performance + data modeling)
• You want structured disagreement before commitment
• The cost of a bad architecture decision exceeds the cost of slower, more expensive analysis
• You're making decisions that are expensive to reverse

Don't use this for simple questions. It's 3-10x the cost and latency of a single API call.

Installation

pip install getconsult

export ANTHROPIC_API_KEY=sk-ant-...

How It Works

┌──────────────────────────────────────────────────────────────────┐
│                         WORKFLOW                                 │
├──────────────────────────────────────────────────────────────────┤
│                                                                  │
│  1. PARALLEL ANALYSIS                                            │
│     N agents with domain-specific system prompts                 │
│     analyze the problem concurrently                             │
│                                                                  │
│  2. PEER REVIEW                                                  │
│     Each agent reviews each other agent's output                 │
│     (N agents = N*(N-1) pairwise reviews)                        │
│                                                                  │
│  3. META REVIEW                                                  │
│     Cross-cutting analysis: integration gaps,                    │
│     unstated assumptions, conflicting recommendations            │
│                                                                  │
│  4. REVISION                                                     │
│     Agents incorporate peer + meta feedback, update solutions    │
│                                                                  │
│  5. APPROVAL VOTING                                              │
│     Each agent votes on each other's revised solution            │
│     APPROVE (1.0) / CONCERNS (0.7) / OBJECT (0.0)                │
│                                                                  │
│  6. RESOLUTION                                                   │
│     ≥80% aggregate approval → consensus                          │
│     <80%, iterations remain → back to step 2                     │
│     <80%, max iterations reached → orchestrator decides          │
│                                                                  │
│  7. SYNTHESIS                                                    │
│     Presentation agent produces ONE unified answer:              │
│     - Merges convergent recommendations                          │
│     - Highlights resolved trade-offs                             │
│     - Notes remaining caveats from peer review                   │
│     - Provides actionable next steps                             │
│                                                                  │
└──────────────────────────────────────────────────────────────────┘

Output: Single synthesized recommendation, not N separate answers.

Usage

consult -p "problem statement" [options]

Examples

# Zero-downtime schema migration with constraints
consult -p "Migrating 200M row PostgreSQL table from UUID to bigint PKs. \
47 foreign key references, 12 indexes, 3k QPS. Cannot lock table. \
Current plan: add column, backfill, swap. Backfill estimated 9 hours. \
What's wrong with this plan and what's the alternative?"

# Production debugging with incomplete information
consult -p "Kubernetes pods OOMKilled. Container limit 4Gi, JVM heap 2Gi, \
G1GC on OpenJDK 17. RSS grows to 3.8Gi over 6 hours. Native memory \
tracking shows 800MB unaccounted. Happens in prod not staging (same image). \
What are we missing?"

# Multi-domain architecture decision
consult -p "Event sourcing vs state-based for order service. \
100k orders/day, need 7-year audit trail, eventual consistency acceptable \
for reads, strong consistency required for inventory decrements. \
Team has zero ES experience. What are the actual trade-offs we'll hit?"

With Expert Selection

# Default: backend, database, infrastructure
consult -p "..."

# Security-focused (security, backend, infrastructure)
consult -p "..." -e security_focused

# Custom combination
consult -p "..." -e "database_expert,performance_expert,security_expert"

With Multiple Iterations (Pro)

# 3 revision cycles
consult -p "..." -i 3

# 90% agreement threshold
consult -p "..." -t 0.9

With Context (Pro)

# Include diagram
consult -p "Review for single points of failure" -a architecture.png

# Continue previous session
consult -p "Now add caching layer" --memory-session project-x

CLI Reference

consult -p "problem" [options]

Required:
  -p, --problem TEXT          Problem statement

Expert Selection:
  -e, --experts TEXT          Set name or comma-separated types
                              Sets: essentials, architecture, security_focused,
                                    performance, full_stack, data_platform, ai_system
                              Types: database_expert, backend_expert, security_expert,
                                     performance_expert, infrastructure_expert,
                                     software_architect, frontend_expert, cloud_engineer,
                                     ml_expert, data_expert

Analysis:
  -m, --mode [single|team]    single = one provider, team = all three (Pro)
  --provider NAME             anthropic, openai, or google
  -i, --max-iterations N      Max revision cycles (default: 1, Pro: up to 5)
  -t, --consensus-threshold F Approval threshold (default: 0.8)

Output:
  --markdown                  Save to ~/.consult/outputs/
  --markdown-filename TEXT    Custom filename
  -c, --copy                  Copy to clipboard

Context:
  --memory-session NAME       Session identifier (Pro)
  -a, --attachments FILES     Images/PDFs to include (Pro)

Info:
  --status                    Tier, limits, usage
  --list-experts              Available experts and sets
  --dry-run                   Validate without API calls
  --version                   Version

Expert Types

Expert Sets

Consensus Mechanism

Each agent votes on each other agent's solution:

3 agents = 6 pairwise votes. Aggregate = mean of all votes.

After consensus (or orchestrator resolution), a synthesis agent produces one unified answer incorporating all expert perspectives. You get a single recommendation, not N separate outputs.

Tiers

BYOK: you provide API keys, pay providers directly.

License Key

export CONSULT_LICENSE_KEY="CSL2_pro_..."
# or
echo "CSL2_pro_..." > ~/.consult/license

Configuration

# ~/.consult/.env

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...      # optional
GOOGLE_API_KEY=...         # optional

# Model overrides (defaults are cost-optimized)
ANTHROPIC_MODEL=claude-sonnet-4-20250514
OPENAI_MODEL=gpt-4o

Performance

Latency dominated by sequential LLM calls per phase. Agents within each phase run in parallel.

Data Storage

~/.consult/
├── .env              # API keys (chmod 600)
├── license           # License key
├── sessions/         # Session state (Pro)
├── outputs/          # Exports (Pro)
├── cache/            # Rate limit tracking
└── logs/             # Debug logs (keys redacted)

Security

• API keys never logged or persisted outside .env
• Session files use hashed identifiers
• Logs redact sensitive patterns
• No telemetry or phone-home

Issues

Bug reports and feature requests: github.com/1x-eng/consult-issues

License

Proprietary. See LICENSE.

Personal and internal business use permitted. Commercial distribution or SaaS integration requires separate license.

Built on AutoGen.