@myconsilium/cli

Command-line interface for Consilium -- a multi-model debate platform that lets you pit AI models against each other before writing code.

Installation

npm install -g @myconsilium/cli

Or run without installing:

npx @myconsilium/cli debate "your question"

Requirements

• Node.js >= 20.0.0
• A running Consilium backend (or access to a hosted instance)

Quick Start

# Interactive REPL session
consilium chat

# Start a debate
consilium debate "How should I implement auth?"

# Debate with options
consilium debate "Design API" --mode council -o output.md

# Resume a previous session
consilium sessions resume <id>

Commands

| Command | Alias | Description | |---------|-------|-------------| | consilium debate <topic> | ask | Start a multi-model debate | | consilium chat | | Interactive REPL with session persistence | | consilium config set\|get\|list | | Configuration management | | consilium login | | Web-based authentication (opens browser) | | consilium debug <debateId> | | Full debate trace | | consilium logs <debateId> | | Query debate logs | | consilium stats | | Model performance dashboard | | consilium sessions list\|resume\|rename\|delete | | Manage saved sessions |

Debate Options

| Flag | Description | |------|-------------| | -m, --models <models...> | Select models for the debate | | --output <format> | Output format: markdown, cursorrules, claude-md, json | | --mode <mode> | Set debate mode (see below) | | --file <paths...> | Attach files as context (e.g., --file src/auth.ts diagram.png) | | --git-diff | Include current git diff as context | | --ticket <id> | Include a Linear ticket as context (e.g., MYC-123) | | --apply | Apply structured edits from synthesis directly to files |

Debate Modes

| Mode | Rounds | Cost | Description | |------|--------|------|-------------| | quick | 1 | ~$0.01 | Single round, fastest results | | council | 3 | ~$0.04 | Multi-round deliberation (default) | | deep | 3 | ~$0.08 | Multi-round with sub-agent research | | blind | 3 | ~$0.04 | Anonymous — models don't see each other's names | | redteam | 4 | ~$0.10 | Adversarial testing, finds attack surfaces | | jury | 3 | ~$0.05 | Panel with mandatory dissent tracking | | market | 5 | ~$0.09 | Prediction-market style with confidence voting | | auto | 3 | ~$0.04 | Auto-selects the best mode for your topic |

consilium debate "Microservices vs monolith" --mode deep
consilium debate "Is this API secure?" --mode redteam
consilium debate "Which approach?" --mode auto

Output Formats

| Format | Use Case | |--------|----------| | markdown | General documentation | | cursorrules | Cursor IDE rules file | | claude-md | CLAUDE.md instructions | | json | Programmatic consumption | | text | Plain text |

consilium debate "Error handling strategy" --output cursorrules

REPL Mode

Running consilium chat drops you into an interactive session with persistent history.

REPL commands:

• /ask <topic> -- Start a debate within the session
• /help -- List available commands
• /exit -- Save session and quit
• Up/Down arrows -- Navigate input history

Codebase-Aware Debates

Consilium scans your project via ProjectContext and feeds relevant context into the debate. Three specialized agents -- architecture, structure, and config -- analyze your codebase so models understand your tech stack, directory layout, and existing patterns before responding.

Context Support

Attach files or images to provide additional context:

consilium debate "Review this architecture" --file diagram.png
consilium debate "Refactor this module" --file src/auth.ts
consilium debate "Compare these implementations" --file old.ts new.ts

Benchmarks

Run multi-model deliberation benchmarks against MMLU, TruthfulQA, or HumanEval:

# Run remotely via the Consilium API
consilium benchmark --benchmark mmlu -n 20

# Run locally via Python (requires apps/agents)
consilium benchmark --benchmark truthfulqa --local -n 10 --output results.json

| Flag | Description | |------|-------------| | --benchmark <name> | Required: mmlu, truthfulqa, or humaneval | | -m, --models <models...> | Models to use as debaters | | --mode <mode> | Deliberation mode (default: council) | | -n <count> | Number of questions to run | | --output <path> | Save JSON results to file | | --local | Run via local Python agent instead of API |

Eval

Run a blind evaluation of multiple responses to the same question:

# Evaluate inline (models generate and judge their own responses)
consilium eval "Which sorting algorithm is best for nearly-sorted data?"

# Evaluate a pre-generated set of responses from a file
consilium eval "Which sorting algorithm?" --responses responses.json

The --responses file should be a JSON array: [{"model": "gpt-5.4", "text": "..."}, ...]

Providers & Models

Consilium supports 7 LLM providers as of April 2026. Bring your own key for any provider — or run without keys and Consilium will fall back to a platform-hosted free-tier pool (Groq + OpenRouter) so you can keep working at zero cost.

| Provider | Current production models | |---|---| | OpenAI | gpt-5.5-pro, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano | | Anthropic | claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001 | | Google | gemini-3.1-pro-preview, gemini-3-flash-preview, gemini-3.1-flash-lite-preview | | Groq | llama-3.1-8b-instant, llama-3.3-70b-versatile, openai/gpt-oss-120b, openai/gpt-oss-20b, groq/compound, groq/compound-mini | | xAI | grok-4-20, grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-code-fast-1 | | Moonshot | kimi-k2.6, kimi-k2.5, kimi-k2-thinking, kimi-k2-thinking-turbo, kimi-k2-turbo-preview | | OpenRouter | google/gemma-4-26b-a4b-it:free, google/gemma-4-31b-it:free, qwen/qwen3-coder:free, nvidia/nemotron-3-super-120b-a12b:free, inclusionai/ling-2.6-1t:free |

Run consilium models for the live catalog with pricing and tier badges. Legacy IDs (e.g. gpt-4o, claude-3-5-sonnet-latest, gemini-2.0-flash) are forwarded to current replacements via aliases — but you should migrate your scripts.

Configuration

Manage API keys and settings with BYOK (Bring Your Own Keys). Supported providers: OpenAI, Anthropic, Google, Groq, xAI, Moonshot, OpenRouter.

consilium config set openai_key sk-...
consilium config set anthropic_key sk-ant-...
consilium config set moonshot_key sk-...
consilium config set openrouter_key sk-or-...
consilium config list

Configuration is stored in ~/.consilium/config.json. BYOK always wins over the free-tier pool. When a debate runs without a key for the requested provider, the CLI prints a pre-flight notice and the engine emits a routing:fallback SSE event so you always know when fallback is active.

Defaults target production (https://api.myconsilium.xyz, https://myconsilium.xyz). For a local Nest API, set:

export CONSILIUM_API_URL="http://localhost:4000"

MCP (Cursor, Claude Code, etc.)

Run consilium mcp for a copy-paste stdio config. The Python module consilium.mcp calls the same Nest API as the CLI using CONSILIUM_API_KEY (your consilium_ token) and CONSILIUM_API_URL (API origin, no /api/v1 suffix). Install: pip install -e packages/python-sdk and optional pip install 'consilium[mcp]' for the official MCP stdio transport.

Features

• Real-time streaming -- SSE streaming with progress bars and agent cards
• Cost estimation -- See estimated cost before a debate runs
• Health check -- Validates backend connectivity before operations
• Decision tracking -- Tracks decisions across conversations (decided/tentative/open/superseded)
• Session persistence -- Saved to ~/.consilium/sessions/, resume with consilium sessions resume <id>

Dependencies

commander ^12.1.0, chalk ^5, ora ^8, eventsource ^2, zod ^3, dotenv, open

Links

• Website
• Documentation