Tzurot v3 Council MCP Collaboration

Use this skill when: Planning major refactors, debugging complex issues, getting code reviews, brainstorming solutions, validating architectural decisions, or needing a second opinion.

Overview

The Council MCP (v4.0.0) provides access to multiple AI models via OpenRouter. Default model is

google/gemini-3-pro-preview

Available Tools

Core Tools

mcp__council__ask

mcp__council__brainstorm

mcp__council__code_review

mcp__council__test_cases

mcp__council__explain

mcp__council__synthesize_perspectives

New in v4.0.0

mcp__council__debug

mcp__council__refactor

mcp__council__recommend_model

mcp__council__list_models

mcp__council__set_model

Multi-Turn Conversations

mcp__council__start_conversation

mcp__council__continue_conversation

mcp__council__list_conversations

mcp__council__get_conversation_history

mcp__council__end_conversation

When to Consult Council

Always Use For

Major Refactorings (>500 lines)

mcp__council__brainstorm({
  topic: 'Risks in refactoring PersonalityService',
  constraints: 'Must maintain exact functionality',
});

Structured Debugging

mcp__council__debug({
  error_message: 'Memory leak in BullMQ workers',
  code_context: 'Workers OOM after 2 hours',
  previous_attempts: ['Checked event listeners', 'Reviewed Redis connections'],
});

Safe Refactoring Plans

mcp__council__refactor({
  code: myCode,
  goal: 'reduce_complexity', // or: extract_method, simplify_logic, improve_naming, etc.
  language: 'typescript',
});

Before Completing Major PRs

mcp__council__code_review({
  code: changes,
  focus: 'behavior preservation, edge cases',
  language: 'typescript',
});

When Thinking "This seems unnecessary" STOP! Consult Council before removing code.

Don't Use For

• Questions answered by existing docs/skills
• Obvious code issues (typos, syntax errors)
• Small style preferences

Model Selection

Quick Guide

Model Classes

• FLASH: Fast & cheap (Haiku, GPT-4o-mini, Gemini Flash)
• PRO: Balanced quality/cost (Sonnet, GPT-4o, Gemini Pro)
• DEEP: Maximum quality (Opus, o1, DeepSeek R1)

Free Tier Options

• meta-llama/llama-3.3-70b-instruct:free

• deepseek/deepseek-chat:free

• qwen/qwen-2.5-72b-instruct:free

Using Model Selection

// Get recommendation for task
mcp__council__recommend_model({ task: 'code_review' });

// Change model for session
mcp__council__set_model({ model: 'anthropic/claude-3.5-sonnet' });

// Or specify per-call
mcp__council__code_review({
  code: myCode,
  model: 'anthropic/claude-3.5-sonnet',
});

Multi-Turn Conversations

For complex discussions that need context across multiple exchanges:

// Start a session
const { session_id } = await mcp__council__start_conversation({
  model: 'deepseek/deepseek-r1',
  system_prompt: 'You are a TypeScript architecture expert',
  initial_message: 'Review this service design...',
});

// Continue the conversation
await mcp__council__continue_conversation({
  session_id,
  message: 'What about the error handling?',
});

// End and get summary
await mcp__council__end_conversation({
  session_id,
  summarize: true,
});

Prompt Structuring

// BAD - No context
mcp__council__ask({ question: 'How do I fix this?' });

// GOOD - Full context
mcp__council__ask({
  question: 'How do I fix race condition in webhook reply tracking?',
  context: 'Using Redis to map message IDs. Bot-client and api-gateway both access Redis.',
});

// BAD - Generic review
mcp__council__code_review({ code: myCode });

// GOOD - Focused review
mcp__council__code_review({
  code: myCode,
  focus: 'resource leaks, error handling, Redis connection management',
  language: 'typescript',
});

The Safety Stack

Thinking -> MCP -> Action

1. Use thinking keywords to analyze ("Ultrathink about...")
2. Consult Council for second opinion
3. Follow project guidelines

Council Limitations

Council doesn't have access to:

• Your local filesystem
• Project-specific documentation (unless provided)
• Git history

Always validate against:

• Project guidelines (CLAUDE.md, skills)
• Existing codebase patterns
• Architecture decisions

When Council and Claude Disagree

Resolution hierarchy:

1. Project guidelines (CLAUDE.md, skills)
2. Existing codebase patterns
3. Technical correctness
4. User preference

Related Skills

• tzurot-architecture - Major design decisions
• tzurot-docs - Document recommendations
• tzurot-security - Security pattern validation
• tzurot-testing - Test case suggestions

References

• MCP tools:

  mcp__council__*

  functions
• Thinking keywords:

  ~/.claude/CLAUDE.md#mandatory-thinking-requirements

• Project guidelines:

  CLAUDE.md