Gemini LLM Council Extension

Consult multiple top-tier LLMs simultaneously with automated peer review and synthesis. Leverage the "Wisdom of the Crowd" to get high-confidence answers for complex architectural and debugging tasks.

Inspired by Andrej Karpathy's llm-council.

✨ Advanced Features

• Autonomous Investigator: New /council:investigate command that uses a specialized subagent to autonomously explore your codebase and gather evidence before deliberating.
• Hierarchical Config: Save your council settings Globally (for all projects) or Project-specifically (checked into your repo).
• Specialized Personas: Run targeted reviews using built-in audit personas (e.g., security, performance).
• Automatic IQ: The council automatically detects if your query requires a specific persona and applies it to guide the review phase.
• Customizable: Define your own personas in ~/.gemini/extensions/gemini-llm-council/personas.json.
• Ambient Grounding: System hooks automatically inject core project metadata (README, package.json) into council consultations.
• Deep Audit Trail: Offload massive raw deliberations to MCP Resources. Accessible via council:// URIs provided in the summary report.

Prerequisites

• Gemini CLI installed.
• An OpenRouter API Key.

Setup

1. Link the extension:

   gemini extensions link .

2. Configure API Key: Use the Gemini CLI to set your OpenRouter API key securely.

   gemini extensions config gemini-llm-council "OpenRouter API Key"

3. Build the extension:

   npm install
   npm run build

4. Configure Council Members:

   /council:setup

   Choose between Global (All Projects) or Project (Current Folder) scope.

Commands

| Command | Description | | :--- | :--- | | /council:setup | Select models, reasoning depth, and configuration scope. | | /council:ask <query> | One-shot consultation with automated project grounding and persona detection. | | /council:investigate <issue> | Autonomous: Subagent handles the file-reading and deliberation loop. | | /council:persona <name> <query> | Consult using a specific persona (e.g., security, performance, or your custom ones). | | /council:status | Show active members, reasoning effort, and active configuration scope. |

Usage Examples

Autonomous Debugging

/council:investigate "The database connection keeps timing out in production environments."

The council investigator will autonomously find your config files, logs, and connection logic to provide a verified fix.

Security Audit (Automatic or Explicit)

/council:ask "Audit the new user registration flow for potential injection flaws."

The Chairman will automatically detect the "Security" domain and load the appropriate persona.

Global vs Project Config

• Global: Stored in ~/.gemini/extensions/gemini-llm-council/config.json.
• Project: Stored in .gemini/llm-council.json. (Project config overrides global).

Architecture

• Autonomous Orchestration: Uses Gemini CLI Subagents to isolate the heavy lifting of multi-file investigations.

• Intelligent Grounding: Uses BeforeTool Hooks to automatically provide context about your tech stack.
• Clean UI: Moves raw multi-model critiques to MCP Resources, keeping your main chat readable.

Inspiration

This project was inspired by Andrej Karpathy's LLM council project, as shared in his Twitter (X) post.

License

MIT