Crash

CRASH (Cascaded Reasoning with Adaptive Step Handling) is an MCP server that provides structured, iterative reasoning for complex problem-solving.

It facilitates systematic analysis through flexible validation, confidence tracking, revision mechanisms, and branching support while maintaining full backward compatibility with existing sequential thinking approaches.

Features

• 🎯 Flexible validation system with configurable strict mode.
• 📊 Confidence tracking on a 0-1 scale with uncertainty notes.
• 🔄 Revision mechanism for correcting and improving previous steps.
• 🌿 Branching support to explore multiple solution paths.
• 💾 Session management for concurrent reasoning chains.
• 📋 Multiple output formats including console, JSON, and markdown.
• 🛠️ Structured tool integration with parameter support.
• 🎨 Natural language flow without forced prefixes.

Use Cases

• Debugging complex issues where multiple hypotheses need testing and validation.
• Architecture planning that involves evaluating different design approaches.
• Code optimization requiring step-by-step analysis of performance bottlenecks.
• Research tasks where confidence levels and uncertainty need explicit tracking.
• Multi-step decision making with branching possibilities and revision needs.

Implementation Guide

1. Install the MCP server with NPM

npm install crash-mcp

OR

npx crash-mcp

2. Configure your MCP Client Configuration (Cursor example):

{
  "mcpServers": {
    "crash": {
      "command": "npx",
      "args": ["-y", "crash-mcp"]
    }
  }
}

3. Environment Variables:

• CRASH_STRICT_MODE: Enables legacy validation (default: false)
• MAX_HISTORY_SIZE: Controls step retention (default: 100)
• CRASH_OUTPUT_FORMAT: Sets display format (console, json, markdown)
• CRASH_NO_COLOR: Disables colored output

4. Basic Usage Structure:

{
  "step_number": 1,
  "estimated_total": 3,
  "purpose": "analysis",
  "context": "Current problem state",
  "thought": "Reasoning process",
  "outcome": "Expected result",
  "next_action": "Next step",
  "rationale": "Why this approach"
}

5. Advanced Parameters:

• Confidence tracking: confidence (0-1) and uncertainty_notes
• Revision support: revises_step and revision_reason
• Branching: branch_from, branch_id, branch_name
• Tool integration: tools_used, external_context, dependencies
• Session management: session_id for grouping chains

6. Standard purposes include analysis, action, reflection, decision, summary, validation, exploration, hypothesis, correction, and planning. Flexible mode allows any custom purpose string.

FAQs

Q: When should I use CRASH instead of internal planning?
A: CRASH provides most value for complex problems requiring systematic analysis of multiple solution paths. For simpler tasks or straightforward implementation, internal planning is usually faster and sufficient.

Q: How does CRASH compare to the original sequential thinking server?
A: CRASH maintains backward compatibility but adds token efficiency, confidence tracking, branching, revision mechanisms, and flexible validation. It’s more streamlined and doesn’t include code in thoughts.

Q: What’s the performance overhead of using CRASH?
A: Minimal – approximately 1-2ms per step with configurable history size to prevent memory issues. The token-optimized design actually reduces overall token usage compared to some alternatives.

Q: Can I use CRASH with multiple concurrent projects?
A: Yes, the session management system allows multiple reasoning chains with unique session IDs, making it suitable for handling several problems simultaneously.

Q: How do I handle module resolution issues?
A: Try using bunx instead of npx, or add the --experimental-vm-modules flag for ESM resolution problems. The package includes troubleshooting guidance for various runtime environments.