Think MCP

Structured reasoning tools for MCP-compatible LLM clients

Reason step by step, branch when needed, block weak finals, and keep useful memory across sessions.

Install | Tools | think_cycle | Quality Gates | Changelog

Why this exists

Most LLM workflows fail in predictable ways:

• They answer too early.

• They stay linear when the task needs alternatives or critique.

• They lose earlier insights between steps.

• They finish with confidence that is not backed by verification.

Think MCP adds an external reasoning layer for those failures. It does not replace the model's intelligence. It constrains and structures the way that intelligence is used.

What is in 5.5.1

• Fixed mojibake and broken runtime coaching text.

• Refreshed the README and aligned it with the current toolset.

• Kept the think_cycle hard-gate workflow introduced in 5.5.0.

• Preserved release validation, eval coverage, and hard quality policy checks.

Core model

Think MCP combines three layers:

Install

Run directly

npx -y @gofman3/think-mcp

MCP config

{
  "mcpServers": {
    "think": {
      "command": "npx",
      "args": ["-y", "@gofman3/think-mcp"]
    }
  }
}

Local development

npm install
npm run build
npm test

Toolset

think_cycle

think_cycle is the main depth-control tool in the current release.

It runs a session as a state machine:

start -> step -> status -> finalize

If the reasoning quality is weak, finalize does not silently pass. It blocks completion and returns concrete next prompts plus a required minimum of additional thoughts.

Key behavior

• Adaptive required depth based on goal complexity and risk markers.

• Hard gate for phase coverage: decompose, alternative, critique, synthesis, verification.

• Quality score with penalties for repetition, weak verification, and unstable confidence.

• Fallback interop with the regular think backend when backendMode=auto.

• Loop budget control to avoid infinite cost and latency growth.

Input shape

{
  action: 'start' | 'step' | 'status' | 'finalize' | 'reset',
  sessionId?: string,
  goal?: string,
  context?: string,
  constraints?: string[],
  thought?: string,
  thoughtType?: 'decompose' | 'alternative' | 'critique' | 'synthesis' | 'verification' | 'revision',
  confidence?: number,
  finalAnswer?: string,
  backendMode?: 'auto' | 'independent' | 'think',
  maxLoops?: number,
  showTrace?: boolean
}

Output shape

{
  status: 'in_progress' | 'blocked' | 'ready' | 'completed' | 'error',
  sessionId: string,
  loop: { current: number, max: number, required: number, remaining: number },
  quality: {
    overall: number,
    coverage: number,
    critique: number,
    verification: number,
    diversity: number,
    confidenceStability: number
  },
  gate: { passed: boolean, reasonCodes: string[] },
  requiredMoreThoughts: number,
  nextPrompts: string[],
  shortTrace: string[],
  finalApprovedAnswer?: string
}

Example flow

// 1. Start
{
  action: 'start',
  goal: 'Design a safe migration from Redis session cache to Postgres-backed sessions',
  constraints: ['zero logout spike', 'rollback in under 5 minutes'],
  backendMode: 'auto'
}

// 2. Add steps
{
  action: 'step',
  sessionId: 'cycle_xxx',
  thought: 'Break the migration into dual-write, read-fallback, rollout metrics, and rollback paths.'
}

// 3. Try to finalize
{
  action: 'finalize',
  sessionId: 'cycle_xxx',
  finalAnswer: 'We should migrate in phases and monitor it carefully...'
}

Typical blocked response:

{
  status: 'blocked',
  gate: { passed: false, reasonCodes: ['MISSING_VERIFICATION', 'LOW_DIVERSITY'] },
  requiredMoreThoughts: 10,
  nextPrompts: [
    'List concrete rollback failure modes.',
    'Verify whether session consistency breaks during dual-write.',
    'Compare at least two rollout strategies.'
  ]
}

Other tools

think

Use when you want incremental reasoning with revisions, branches, substeps, and quick extensions.

{
  thought: 'The bug likely comes from stale branch state after retry.',
  thoughtNumber: 3,
  totalThoughts: 7,
  nextThoughtNeeded: true,
  confidence: 6,
  quickExtension: {
    type: 'critique',
    content: 'Verify whether retry state is recreated or reused.'
  }
}

think_batch

Use when you already know the rough chain and want to submit it in one call.

{
  goal: 'Audit deployment rollback flow',
  thoughts: [
    { thoughtNumber: 1, thought: 'Identify entry points for rollout state changes.' },
    { thoughtNumber: 2, thought: 'Trace rollback triggers and timeout behavior.' }
  ]
}

think_logic

Use for strict methodology generation before a deep audit.

{
  target: 'Review the payment retry pipeline for consistency and failure isolation',
  depth: 'deep',
  focus: ['reliability', 'performance', 'data-flow']
}

think_recall

Use before starting a familiar class of problem.

{
  query: 'rollback strategy cache migration',
  scope: 'insights',
  searchIn: 'all',
  limit: 5
}

Quality and release gates

Release verification is built into the repo:

npm run validate:release

Main checks:

• TypeScript typecheck

• Unit tests

• Local eval scenarios

• Repo structure validation

• Security audit

• Hard quality baseline in docs/quality/HARD_QUALITY_STANDARD.md

Runtime storage

• Default data directory: ~/.think-mcp

• Override with THINK_MCP_DATA_DIR

• think_cycle sessions persist in runtime storage with TTL cleanup

Package links

• npm: @gofman3/think-mcp

• repo: GofMan5/think-mcp

Changelog

v5.5.1

• Fixed broken mojibake output in runtime coaching strings.

• Rebuilt and refreshed README for the current toolset and quality model.

• Released documentation and packaging cleanup on top of 5.5.0.

v5.5.0

• Added think_cycle for adaptive external reasoning with hard quality gates.

• Added release-gated hard quality policy based on NEED_ADD.

• Added eval scenarios for cycle gating, fallback behavior, autonomy quality, safety gates, and bounded retries.

v5.1.0

• Switched prompt style toward imperative IF/THEN instructions.

• Reduced token overhead significantly for common reasoning flows.

v5.0.0

• Added think_logic methodology generation.