Skip to main content
Glama

Retrieval-Augmented Thinking MCP Server

by stat-guy
# RAT (Retrieval Augmented Thinking) MCP Tool ## Tool Name `rat` ## Description A context-aware reasoning system that orchestrates structured thought processes through dynamic trajectories. Implements iterative hypothesis generation and validation cycles while preserving context coherence across non-linear reasoning paths. ## Installation ### Simple 3-Step Process ```bash git clone https://github.com/stat-guy/retrieval-augmented-thinking.git cd retrieval-augmented-thinking npm install -g . ``` ### Verify Installation Test before configuring Claude Desktop: ```bash npx mcp-server-rat-node --help ``` **Success indicator:** If you see `RAT MCP Server (Node.js) running on stdio`, installation is complete! ## Claude Desktop Configuration Add to `claude_desktop_config.json`: ```json { "mcpServers": { "rat": { "command": "npx", "args": ["mcp-server-rat-node"] } } } ``` Restart Claude Desktop after adding the configuration. ## Parameters ### Required Parameters - **`thought`** (string): Structured reasoning step content - Primary analysis chains - Hypothesis formulation/validation - Branch exploration paths - Revision proposals - Context preservation markers - Verification checkpoints - **`nextThoughtNeeded`** (boolean): Signal for continuation of reasoning chain - `true`: Continue to next thought - `false`: Complete current reasoning trajectory - **`thoughtNumber`** (integer, min: 1): Position in current reasoning trajectory - Sequential numbering starting from 1 - Must be ≤ `totalThoughts` unless `needsMoreThoughts` is true - **`totalThoughts`** (integer, min: 1): Dynamic scope indicator (adjustable) - Initial estimation of reasoning steps needed - Can be expanded via `needsMoreThoughts` ### Optional Parameters - **`isRevision`** (boolean, default: false): Marks recursive refinement steps - **`revisesThought`** (integer): References target thought number for refinement - **`branchFromThought`** (integer): Indicates parallel exploration paths source - **`branchId`** (string): Context identifier for parallel reasoning chains - **`needsMoreThoughts`** (boolean, default: false): Signals scope expansion requirement ## Response Structure ```typescript { thought_number: number, total_thoughts: number, metrics: { complexity: number, // 0.0-1.0: Structural complexity score depth: number, // 0.0-1.0: Content depth analysis quality: number, // 0.0-1.0: Overall thought quality impact: number, // 0.0-1.0: Reasoning impact score confidence: number // 0.0-1.0: Certainty assessment }, analytics: { total_thoughts: number, average_quality: number, chain_effectiveness: number }, next_thought_needed: boolean, visual_output: string // Formatted display with metrics } ``` ## Usage Patterns ### Sequential Analysis ```json { "thought": "First, I need to establish the core problem parameters.", "nextThoughtNeeded": true, "thoughtNumber": 1, "totalThoughts": 5 } ``` ### Hypothesis Validation ```json { "thought": "Testing hypothesis: The root cause is data inconsistency.", "nextThoughtNeeded": true, "thoughtNumber": 3, "totalThoughts": 5 } ``` ### Revision with Refinement ```json { "thought": "Upon further analysis, the initial assumption was incomplete...", "nextThoughtNeeded": true, "thoughtNumber": 4, "totalThoughts": 5, "isRevision": true, "revisesThought": 2 } ``` ### Branch Exploration ```json { "thought": "Alternative approach: Consider the performance implications...", "nextThoughtNeeded": true, "thoughtNumber": 3, "totalThoughts": 4, "branchFromThought": 2, "branchId": "performance_analysis" } ``` ### Scope Extension ```json { "thought": "This requires deeper investigation beyond initial scope.", "nextThoughtNeeded": true, "thoughtNumber": 5, "totalThoughts": 5, "needsMoreThoughts": true } ``` ## Execution Protocol 1. **Initialize** with scope estimation (`totalThoughts`) 2. **Generate** structured reasoning steps sequentially 3. **Validate** hypotheses through verification cycles 4. **Maintain** context coherence across branches 5. **Implement** revisions through recursive refinement 6. **Signal** completion when validation succeeds ## Best Practices for AI Assistants - Start with realistic `totalThoughts` estimate (3-7 for most problems) - Use `isRevision` when refining earlier thoughts - Create branches for parallel analysis paths - Extend scope with `needsMoreThoughts` when complexity emerges - Always provide substantive `thought` content - Set `nextThoughtNeeded: false` only when reasoning is complete ## Error Prevention - Ensure `thought` content is non-empty and meaningful - Keep `thoughtNumber` ≤ `totalThoughts` unless extending scope - When using `isRevision: true`, always specify `revisesThought` - When branching, provide both `branchFromThought` and `branchId` - Use descriptive, analytical language in thought content ## Installation Notes The installation process includes automatic permission fixes via postinstall scripts. Most users will only need the simple 3-step process. Manual chmod commands are rarely needed as the system handles permissions automatically.

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/stat-guy/retrieval-augmented-thinking'

If you have feedback or need assistance with the MCP directory API, please join our Discord server