Unconventional Thinking Server (v0.3.0)

A context-efficient MCP server for bold, unconventional, and boundary-breaking problem-solving.

This is a TypeScript-based MCP server that implements an unconventional thinking system optimized for context space savings based on Anthropic's latest MCP architecture patterns. It generates and tracks creative solutions to problems while maintaining efficiency.

> MCP spec 2025-11-25 compliant — uses `@modelcontextprotocol/sdk` v1.27.1 with tool `title`, `annotations`, `outputSchema`, `structuredContent` responses, and `resource_link` content type.

Architecture: Context-Saving Design

This server demonstrates Anthropic's recommended patterns for reducing context overhead by 98.7%:

Key Context-Saving Features

1. Resources API for On-Demand Data Loading

2. Server-Side Filtering

3. Metadata-First Returns

4. Persistent File-Based Storage

Features

Tools (All Context-Efficient, MCP spec 2025-11-25)

Each tool now includes:

• `title` — human-readable display name shown in client UIs
• `annotations` — behaviour hints (`readOnlyHint`, `destructiveHint`, `idempotentHint`, `openWorldHint`)
• `outputSchema` — JSON Schema describing the structured result
• `structuredContent` in responses — machine-readable output conforming to the schema
• `resource_link` content items — explicit links clients can subscribe to or fetch

• `generate_unreasonable_thought` — Generate new unconventional thoughts
  • Returns `resource_link` + `structuredContent`, not raw text blobs
  • Can build upon or rebel against previous thoughts
  • Full thought content available via Resources API

• `branch_thought` — Create new branches of thinking
  • Supports directions: `more_extreme`, `opposite`, `tangential` (now enum-typed)
  • Returns `resource_link` + `structuredContent` for the new branch

• `search_thoughts` — Efficient metadata search
  • Filters by branchId, isRebellion, challengesAssumption
  • Returns `structuredContent` with typed count + thoughts array
  • Includes limit parameter to control result size

Resources (On-Demand Content Loading)

• Each thought available as a resource: `thought://[thoughtId]`
• Metadata includes: isRebellion, challengesAssumption, timestamp, branch info
• Full thought content loaded only when Claude explicitly requests it
• Dramatically reduces token usage when many thoughts exist

How This Implements Context Efficiency

1. Progressive Disclosure

Claude doesn't need the full content of 100 thoughts upfront. Instead:

• `search_thoughts` returns just IDs and metadata (100 bytes per thought)
• Claude selectively fetches full content via Resources API for relevant thoughts
• Similar to how filesystems work: list files, then open specific files

2. Server-Side Filtering

Traditional approach (❌ inefficient):

All 1000 thoughts → Claude → Claude filters → Uses only 10
(costs tokens for all 1000)

This server (✅ efficient):

search_thoughts filter params → Server filters locally → Returns only 10 results
(Claude never sees the unused 990)

3. Metadata-First Pattern

Tool responses contain:

• Thought ID
• Resource URI to access full content
• Brief metadata (2-3 KB each)
• NOT the full 500-character thought (saves ~5KB per thought)

Example savings: With 100 thoughts:

• Old way: 500KB context usage
• New way: ~30KB + fetch only what's needed

Development

Install dependencies:

npm install

Build the server:

npm run build

For development with auto-rebuild:

npm run watch

Installation

To use with Claude Desktop, add the server config:

On MacOS: `~/Library/Application Support/Claude/claude_desktop_config.json`

On Windows: `%APPDATA%/Claude/claude_desktop_config.json`

{
  "mcpServers": {
    "unconventional-thinking": {
      "command": "/path/to/unconventional-thinking/build/index.js"
    }
  }
}

Usage Example

Claude: Generate an unreasonable thought about scaling problems
→ Tool: generate_unreasonable_thought("scaling problems")
← Returns: resource_link (thought://...) + structuredContent { thoughtId, isRebellion, ... }

Claude: What are all the rebellious thoughts?
→ Tool: search_thoughts(isRebellion=true, limit=5)
← Returns: structuredContent { count, thoughts: [...metadata] }

Claude: I need to see the full content of thought_xyz
→ Resource: Read thought://thought_xyz
← Returns: Full thought content (loaded only when needed)

Debugging

Since MCP servers communicate over stdio, debugging can be challenging. We recommend using the MCP Inspector, which is available as a package script:

npm run inspector

The Inspector will provide a URL to access debugging tools in your browser.

References

This server implements patterns from:

• Anthropic: Code Execution with MCP
• Model Context Protocol Docs