🧠 Think MCP
Structured Sequential Thinking Server for LLMs
Break down complex problems. Branch out ideas. Remember insights.
Designed for LLMs to think deeper, smarter, and more efficiently.
Quick Start • Tools • Features • Changelog
📖 Overview
Think MCP transforms how LLMs approach problem-solving. It's not just a tool; it's a cognitive framework that enables:
- Sequential Reasoning: Step-by-step problem decomposition.
- Branching & Revision: Ability to backtrack, fork thoughts, and correct mistakes.
- Deep Analysis: Built-in methodology generator for rigorous code and logic auditing.
- Long-term Memory: Cross-session recall of insights and dead ends.
✨ Features
| Feature | Description |
|---|---|
| 🚀 Efficient Thinking | Imperative Prompts (v5.1) reduce token usage by ~55% using IF/THEN logic. |
| ⚡️ Burst Mode | Submit up to 30 thoughts in a single API call with think_batch. |
| 🧠 Methodology Generator | On-demand deep analysis frameworks (Chain Mapping, Crack Hunting, etc.) via think_logic. |
| 💾 Smart Memory | Cross-session learning via think_recall and auto-save with 24h retention. |
| 🔔 Nudge System | Proactive micro-prompts to detect low confidence, tunnel vision, or missed steps. |
| 🌳 Branching | Explore alternative paths without losing context. |
🚀 Quick Start
Installation
npx @gofman3/think-mcp
MCP Configuration
Add this to your MCP settings file:
{
"mcpServers": {
"think": {
"command": "npx",
"args": ["-y", "@gofman3/think-mcp"]
}
}
}
🛠️ Tools Reference
1. think
The core unit of reasoning. Adds a single thought to the chain.
{
thought: string, // The reasoning content
thoughtNumber: number, // Current step index
totalThoughts: number, // Estimated total steps
nextThoughtNeeded: boolean, // Is the chain complete?
confidence?: number, // Score 1-10
goal?: string, // Required for the first thought
subSteps?: string[], // Breakdown of current step (max 5)
alternatives?: string[], // Other approaches considered (max 5)
quickExtension?: { // Immediate micro-actions
type: 'critique' | 'elaboration' | 'correction' | 'innovation' | 'polish',
content: string,
impact?: 'low' | 'medium' | 'high' | 'blocker'
},
isRevision?: boolean, // Is this correcting a previous step?
revisesThought?: number, // Which step is being fixed?
branchId?: string // For branching paths
}
2. think_batch
High-velocity reasoning. Submit a complete chain (1-30 thoughts) at once.
⚠️ Constraints
IF similarity > 60%→ Reject ("Stagnation")IF thought < 50 chars→ Reject ("Too short")IF avg_confidence < 4→ Warning issued
{
goal: string, // Min 10 chars
thoughts: Array<{ // List of thought objects
thoughtNumber: number,
thought: string, // 50-1000 chars
confidence?: number,
// ... other standard fields
}>,
consolidation?: {
winningPath: number[],
summary: string,
verdict: 'ready' | 'needs_more_work'
}
}
3. think_logic
The Architect. Generates a strict thinking methodology for complex analysis.
Output Phases: CHAIN MAPPING → CRACK HUNTING → STANDARD BENCHMARK → ACTION PLANNING
{
target: string, // The subject of analysis (Min 10 chars)
depth?: 'quick' | 'standard' | 'deep',
focus?: ('security' | 'performance' | 'reliability' | 'ux' | 'data-flow')[],
stack?: ('nestjs' | 'react' | 'redis' | 'nextjs' | /* etc */)[]
}
4. think_recall
The Memory Bank. Search current session or past insights.
Best Practices:
BEFORE complex_task→ Checkscope: 'insights'IF repeating_logic→ Check for dead endsIF unsure→ Verify context
{
query: string,
scope?: 'session' | 'insights',
searchIn?: 'thoughts' | 'extensions' | 'alternatives' | 'all',
limit?: number
}
5. think_done & think_reset
think_done: Finalize session. Validates gaps, blockers, and confidence levels.think_reset: Wipe the slate clean. (Use only if task context changes completely).
💡 Intelligent Systems
The Nudge System
The server watches your back.
| Trigger Pattern | System Nudge |
|---|---|
confidence < 5 | "Low confidence. Validate?" |
3+ thoughts w/o alternatives | "No alternatives. Tunnel vision?" |
| Complex goal w/o subSteps | "Complex goal, no breakdown. Decompose?" |
| Unresolved blocker | "Blocker unresolved. Fix first." |
Complexity Budget
Recommended tool usage based on task size.
| Task Difficulty | Thoughts | Recommended Tool |
|---|---|---|
| Simple | 0-2 | Skip (Direct Answer) |
| Medium | 3-7 | think (Step-by-step) |
| Complex | 8-30 | think_batch (Burst mode) |
🔄 Changelog
v5.1.0 (Current)
- Imperative Prompts: Switched to IF/THEN style instructions.
- Performance: ~55% Token Reduction per request.
- Optimization: Faster parsing, less LLM overhead.
v5.0.0
- New Tool:
think_logicfor generating methodologies. - Framework: Added 4-phase analysis (Mapping, Cracking, Benchmarking, Planning).
v4.x.x History
- v4.6.0: Added NudgeService for proactive prompts.
- v4.5.0: Renamed to
think, added Cross-session insights. - v4.1.0: Introduced Burst Thinking (
think_batch).
MIT License • Created by @gofman3