Sequential Thinking MCP

Sequential Thinking — Structured Reasoning

You are an expert at using the Sequential Thinking MCP server for structured, revisable multi-step reasoning. This tool helps decompose complex problems into numbered thoughts with support for revision and branching.

Critical: These Are Direct Tool Calls

MCP tools are direct tool calls — exactly like Read, Grep, or Bash. They are NOT CLI commands.

CORRECT — call the tool directly:

Tool: mcp__sequential-thinking__sequentialthinking
Parameters: {
  "thought": "First, let me identify the possible causes of the RLS policy failure...",
  "thoughtNumber": 1,
  "totalThoughts": 5,
  "nextThoughtNeeded": true
}

WRONG — do NOT shell out:

Bash: claude mcp call sequential-thinking sequentialthinking ...  # This does not work

The single tool uses the mcp__sequential-thinking__sequentialthinking prefix.

When to Use (and When NOT to)

Rule of thumb: If you can solve it in your head in one step, don't use this tool. If you need to "think out loud" with potential backtracking, use it.

Tool Parameters

Workflow 1: Problem Decomposition

Trigger: Complex bug with multiple possible causes, or a decision requiring systematic analysis.

Steps

1. Frame the problem (thought 1):

   sequentialthinking({
     thought: "The timesheet approval is failing silently. Three possible causes: (1) RLS policy blocking the update, (2) has_permission() returning false, (3) trigger rejecting the state transition.",
     thoughtNumber: 1,
     totalThoughts: 5,
     nextThoughtNeeded: true
   })
   

2. Analyze each hypothesis (thoughts 2-4):

   sequentialthinking({
     thought: "Testing hypothesis 1: The RLS policy for timesheets requires has_role_on_account(account_id) AND the user must be manager_or_above. Let me check if the user has the right role...",
     thoughtNumber: 2,
     totalThoughts: 5,
     nextThoughtNeeded: true
   })
   

3. Revise if earlier thinking was wrong (optional):

   sequentialthinking({
     thought: "Revising thought 2: Actually, the new permission system uses has_permission() not has_role_on_account(). The RLS policy might be using the old function.",
     thoughtNumber: 4,
     totalThoughts: 6,
     nextThoughtNeeded: true,
     isRevision: true,
     revisesThought: 2,
     needsMoreThoughts: true
   })
   

4. Conclude (final thought):

   sequentialthinking({
     thought: "Root cause identified: The RLS policy uses has_role_on_account() which checks system roles, but the user only has a custom role assigned. Need to update the policy to use has_permission() instead.",
     thoughtNumber: 6,
     totalThoughts: 6,
     nextThoughtNeeded: false
   })
   

Workflow 2: Architectural Decision

Trigger: User asks "should we use X or Y?", "what's the best approach for Z?", or there are multiple valid implementation strategies.

Steps

1. Define the decision criteria (thought 1)
2. Evaluate option A against criteria (thought 2)
3. Evaluate option B against criteria (thought 3)
4. Branch if a hybrid approach emerges (optional):

   sequentialthinking({
     thought: "What if we combine the service pattern from option A with the caching strategy from option B?",
     thoughtNumber: 4,
     totalThoughts: 6,
     nextThoughtNeeded: true,
     branchFromThought: 3,
     branchId: "hybrid-approach"
   })
   

5. Recommend with trade-off summary (final thought)

Workflow 3: Multi-Factor Analysis

Trigger: Complex decision involving performance, security, maintainability, and other competing concerns.

Pattern

• Thought 1: List all factors and their relative importance
• Thoughts 2-N: Analyze each factor independently
• Revision thoughts: Update earlier analysis as new interactions between factors emerge
• Final thought: Synthesize into a recommendation with explicit trade-offs

Key Patterns

• Start with an estimate of totalThoughts — 5 is a good default. Adjust with needsMoreThoughts: true if you need more.
• Use isRevision: true when you realize an earlier thought was wrong or incomplete — don't just silently change direction.
• Use branchFromThought to explore "what if?" alternatives without abandoning the main analysis.
• Set nextThoughtNeeded: false only when you have a clear conclusion — not just when you run out of thoughts.
• Each thought should be substantive — avoid placeholder thoughts like "Let me think about this more."

Troubleshooting

Reasoning Chain Gets Too Long

If you're past thought 10 and still going:

1. Set nextThoughtNeeded: false and summarize what you know so far
2. The problem might need breaking into sub-problems — solve each separately
3. Consider if you actually need this tool or if the problem is simpler than you thought

Losing Track of the Thread

Use isRevision and revisesThought to explicitly connect corrections to earlier thoughts. This helps maintain a coherent chain rather than just appending unconnected thoughts.

Tool Seems Unnecessary for This Problem

If you find yourself forcing sequential thinking on a straightforward problem, just stop and solve it directly. This tool adds value for genuinely complex reasoning — not for every problem.