UltraThink

UltraThink

A Python-based MCP server that facilitates structured problem-solving through sequential thinking, branching, and confidence scoring. It allows users to track assumptions and manage multiple concurrent reasoning sessions to break down complex tasks.

Category
Visit Server

README

UltraThink MCP Server

<div align="center">

A Python MCP server for sequential thinking and problem-solving

Python 3.12+ FastMCP MIT License Claude Code Plugin

</div>


Enhanced Python port of the Sequential Thinking MCP Server by Anthropic. Maintains full compatibility while adding confidence scoring, auto-assigned thought numbers, and multi-session support.

[!TIP] Using Claude Code? Install the UltraThink Plugin for seamless integration - no MCP server setup required!

# Via terminal
claude plugin marketplace add husniadil/ekstend
claude plugin install ultrathink@ekstend

# Or interactively in Claude Code
/plugin marketplace add husniadil/ekstend
/plugin install ultrathink@ekstend

[!NOTE] Meta: This MCP server was built iteratively using UltraThink itself - a practical example of the tool's capability to break down complex problems, manage architectural decisions, and maintain context across development sessions.


Features

  • UltraThink: Break down complex problems into manageable steps
  • Dynamic Adjustments: Revise and refine thoughts as understanding deepens
  • Branching: Explore alternative paths of reasoning
  • Confidence Scoring: Explicit uncertainty tracking (0.0-1.0 scale)
  • Auto-adjustment: Automatically adjusts total thoughts if needed
  • Multi-Session Support: Manage multiple concurrent thinking sessions with session IDs
  • Formatted Logging: Colored terminal output with rich formatting (can be disabled)
  • 100% Test Coverage: Comprehensive test suite with full code coverage
  • Type Safety: Full mypy strict mode type checking for production code
  • Simple Layered Architecture: Clean separation with models, services, and interface layers

Installation

Quick Install (Recommended)

Run directly with uvx from GitHub (no installation needed):

uvx --from git+https://github.com/husniadil/ultrathink ultrathink

Development Setup

For local development:

# Clone the repository
git clone https://github.com/husniadil/ultrathink.git
cd ultrathink

# Install all dependencies (including dev dependencies)
uv sync

Usage

Task Commands (npm-like)

# List all available tasks
uv run task --list

# Run the server
uv run task run

# Run tests with coverage
uv run task test

# Run tests without coverage (quick)
uv run task test-quick

# Run the test client
uv run task client

# Format code (ruff + prettier)
uv run task format

# Lint code
uv run task lint

# Type check with mypy
uv run task typecheck

# Clean cache files
uv run task clean

Direct Commands (Alternative)

For direct execution without task runner:

# Run the server directly
uv run ultrathink

# Run the test client directly
uv run python examples/client.py

Note: For testing, linting, and formatting, prefer using uv run task commands shown above.

Tool: ultrathink

The server provides a single tool for dynamic and reflective problem-solving through structured thinking.

Parameters

Required:

  • thought (str): Your current thinking step
  • total_thoughts (int): Estimated total thoughts needed (>=1)

Optional:

  • thought_number (int): Current thought number - auto-assigned sequentially if omitted (1, 2, 3...), or provide explicit number for branching/semantic control
  • next_thought_needed (bool): Whether another thought step is needed. Auto-assigned as thought_number < total_thoughts if omitted. Set explicitly to override default behavior
  • session_id (str): Session identifier for managing multiple thinking sessions (None = create new, provide ID to continue session)
  • is_revision (bool): Whether this revises previous thinking
  • revises_thought (int): Which thought number is being reconsidered
  • branch_from_thought (int): Branching point thought number
  • branch_id (str): Branch identifier
  • needs_more_thoughts (bool): If more thoughts are needed
  • confidence (float): Confidence level (0.0-1.0, e.g., 0.7 for 70% confident)
  • uncertainty_notes (str): Optional explanation for doubts or concerns about this thought
  • outcome (str): What was achieved or expected as result of this thought
  • assumptions (list[Assumption]): Assumptions made in this thought (id, text, confidence, critical, verifiable)
  • depends_on_assumptions (list[str]): Assumption IDs this thought depends on (e.g., ["A1", "A2"])
  • invalidates_assumptions (list[str]): Assumption IDs proven false (e.g., ["A3"])

Response

Returns a JSON object with:

  • session_id: Session identifier for continuation
  • thought_number: Current thought number
  • total_thoughts: Total thoughts (auto-adjusted if needed)
  • next_thought_needed: Whether more thinking is needed
  • branches: List of branch IDs
  • thought_history_length: Number of thoughts processed in this session
  • confidence: Confidence level of this thought (0.0-1.0, optional)
  • uncertainty_notes: Explanation for doubts or concerns (optional)
  • outcome: What was achieved or expected (optional)
  • all_assumptions: All assumptions tracked in this session (keyed by ID)
  • risky_assumptions: IDs of risky assumptions (critical + low confidence + unverified)
  • falsified_assumptions: IDs of assumptions proven false

Example

Basic Usage

from fastmcp import Client
from ultrathink import mcp

async with Client(mcp) as client:
    # Simple sequential thinking with auto-assigned fields
    result = await client.call_tool("ultrathink", {
        "thought": "Let me analyze this problem step by step",
        "total_thoughts": 3
        # thought_number auto-assigned: 1
        # next_thought_needed auto-assigned: True (1 < 3)
    })

With Enhanced Features

async with Client(mcp) as client:
    # With confidence scoring and explicit session
    result = await client.call_tool("ultrathink", {
        "thought": "Initial hypothesis - this approach might work",
        "total_thoughts": 5,
        "confidence": 0.6,  # 60% confident
        # next_thought_needed auto-assigned: True
        "session_id": "problem-solving-session-1"
    })

    # Continue the same session with higher confidence
    result2 = await client.call_tool("ultrathink", {
        "thought": "After analysis, I'm more certain about this solution",
        "total_thoughts": 5,
        "confidence": 0.9,  # 90% confident
        # next_thought_needed auto-assigned: True
        "session_id": "problem-solving-session-1"  # Same session
    })

    # Branch from a previous thought
    result3 = await client.call_tool("ultrathink", {
        "thought": "Let me explore an alternative approach",
        "total_thoughts": 6,
        "confidence": 0.7,
        "branch_from_thought": 1,
        "branch_id": "alternative-path",
        # next_thought_needed auto-assigned: True
        "session_id": "problem-solving-session-1"
    })

With Uncertainty Notes and Outcome

async with Client(mcp) as client:
    # Track uncertainty and outcomes
    result = await client.call_tool("ultrathink", {
        "thought": "Testing the authentication fix",
        "total_thoughts": 5,
        "confidence": 0.8,
        "uncertainty_notes": "Haven't tested under high load yet",
        "outcome": "Login flow works for standard users"
    })

    # Response includes the new fields
    print(result["confidence"])          # 0.8
    print(result["uncertainty_notes"])   # "Haven't tested under high load yet"
    print(result["outcome"])             # "Login flow works for standard users"

With Assumption Tracking

async with Client(mcp) as client:
    # Thought 1: State assumptions explicitly
    result = await client.call_tool("ultrathink", {
        "thought": "Redis should meet our performance requirements",
        "total_thoughts": 4,
        "assumptions": [
            {
                "id": "A1",
                "text": "Network latency to Redis < 5ms",
                "confidence": 0.8,
                "critical": True,
                "verifiable": True,
                "evidence": "Based on preliminary network tests in staging environment"
            }
        ]
    })

    # Thought 2: Build on previous assumptions
    result2 = await client.call_tool("ultrathink", {
        "thought": "Based on low latency, Redis can handle 10K req/sec",
        "total_thoughts": 4,
        "depends_on_assumptions": ["A1"],
        "session_id": result["session_id"]
    })

    # Thought 3: Invalidate if proven false
    result3 = await client.call_tool("ultrathink", {
        "thought": "After testing, latency is 15ms, not 5ms!",
        "total_thoughts": 4,
        "invalidates_assumptions": ["A1"],
        "session_id": result["session_id"]
    })

    # Track all assumptions and detect risky ones
    print(result3["all_assumptions"])      # {"A1": {...}}
    print(result3["falsified_assumptions"]) # ["A1"]

Configuration

Environment Variables

  • DISABLE_THOUGHT_LOGGING: Set to "true" to disable colored thought logging to stderr

Usage with Claude Desktop

Add to your claude_desktop_config.json:

Using uvx from GitHub (Recommended)

{
  "mcpServers": {
    "UltraThink": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/husniadil/ultrathink",
        "ultrathink"
      ]
    }
  }
}

Local Development

For local development from source:

{
  "mcpServers": {
    "UltraThink": {
      "command": "uv",
      "args": ["--directory", "/path/to/ultrathink", "run", "ultrathink"]
    }
  }
}

Local Configuration File

For local development and testing, you can create a .mcp.json file (see .mcp.json.example):

# Copy the example file
cp .mcp.json.example .mcp.json

# Edit to match your local path
# Change /path/to/ultrathink to your actual directory

Example configuration (.mcp.json.example):

{
  "mcpServers": {
    "UltraThink": {
      "command": "uv",
      "args": ["--directory", "/path/to/ultrathink", "run", "ultrathink"],
      "env": {
        "DISABLE_THOUGHT_LOGGING": "false"
      }
    }
  }
}

This configuration:

  • Enables thought logging by default (DISABLE_THOUGHT_LOGGING: "false")
  • Can be used with MCP clients that support .mcp.json configuration
  • Useful for testing the server locally with colored output enabled
  • Note: .mcp.json is gitignored - customize it for your local setup

Session Management

Session Lifecycle

Important: Sessions are stored in-memory only and will be lost when the server restarts or terminates. Each session is identified by a unique session ID and maintains:

  • Thought history for that session
  • Branch tracking
  • Sequential thought numbering

Implications:

  • Sessions do not persist across server restarts
  • All thinking context is lost when the server stops
  • For production use cases requiring persistent sessions, you would need to implement custom session persistence (e.g., to disk, database, or external state management)

Best Practices:

  • Use custom session IDs (instead of auto-generated UUIDs) for resilient recovery if you need to recreate session context
  • Keep session-critical information in your application layer if persistence is required
  • Consider sessions as ephemeral working memory for active problem-solving tasks

Architecture

Built with Simple Layered Architecture principles for clean separation of concerns and maintainable code.

Files

src/ultrathink/ (3-layer structure)

Models Layer (models/)

  • thought.py: Thought, ThoughtRequest, ThoughtResponse models
  • session.py: ThinkingSession model

Services Layer (services/)

  • thinking_service.py: UltraThinkService business logic

Interface Layer (interface/)

  • mcp_server.py: MCP server entry point with FastMCP tool registration

Package Entry Points

  • __init__.py: Package exports
  • __main__.py: CLI entry point (enables uv run ultrathink)

tests/ (100% coverage, mirroring source structure)

Models Tests (models/)

  • test_thought.py: Thought model tests (properties, formatting)
  • test_session.py: Session logging and formatting tests

Services Tests (services/)

  • test_thinking_service.py: Service tests (validation, functionality, branching, multi-session)

Interface Tests (interface/)

  • test_mcp_server.py: MCP tool function tests

Root Test Files

  • test_cli.py: CLI entry point tests

examples/

  • client.py: Test client demonstrating tool usage

Architecture Layers

1. Models Layer

Pydantic models for data representation and validation:

Thought: Core model representing a single thought with validation and behaviors ThoughtRequest: Input model from MCP clients with validation ThoughtResponse: Output model to MCP clients with structured data ThinkingSession: Session model managing thought history and branches

# Type-safe model usage
request = ThoughtRequest(
    thought="My thinking step",
    thought_number=1,
    total_thoughts=3,
    next_thought_needed=True
)
response = ThoughtResponse(
    thought_number=1,
    total_thoughts=3,
    next_thought_needed=True,
    branches=[],
    thought_history_length=1
)

2. Services Layer

Business logic and orchestration:

UltraThinkService: Orchestrates the thinking process

Responsibilities:

  • Model Translation: ThoughtRequest → Thought model (input)
  • Business Logic: Delegate to ThinkingSession
  • Response Building: Session state → ThoughtResponse model (output)
  • Validation: Leverages Pydantic for automatic validation
  • Session Management: Create and manage multiple thinking sessions

Key Method:

  • process_thought(request: ThoughtRequest) → ThoughtResponse: Main orchestration
service = UltraThinkService()

# Full flow:
# 1. Receives ThoughtRequest from interface layer
# 2. Translates to Thought model
# 3. Calls session.add_thought() (business logic)
# 4. Builds ThoughtResponse from session state
# 5. Returns response
request = ThoughtRequest(thought="...", thought_number=1, ...)
response = service.process_thought(request)

3. Interface Layer

External interface using FastMCP:

mcp_server.py: MCP server tool registration

Responsibilities:

  • Define MCP tools using @mcp.tool decorator
  • Map tool parameters to model types
  • Call services layer for processing
  • Return responses to MCP clients
@mcp.tool
def ultrathink(thought: str, total_thoughts: int, ...) -> ThoughtResponse:
    request = ThoughtRequest(thought=thought, total_thoughts=total_thoughts, ...)
    return thinking_service.process_thought(request)

Type Safety Benefits:

  • Pydantic validation on all inputs/outputs
  • No arbitrary dicts - strict typing throughout
  • Automatic validation errors
  • Clear separation between interface and business logic

Architecture Benefits

  1. Clear Separation of Concerns:

    • Models layer = Data models with validation and behaviors
    • Services layer = Business logic and orchestration
    • Interface layer = External API (MCP tools)
  2. Simpler Structure: Flatter folder hierarchy (2 levels instead of 3)

  3. Easier Imports: Shorter relative import paths (..models vs ...domain.entities)

  4. Consolidated Models: Related models grouped together (Thought, ThoughtRequest, ThoughtResponse in one file)

  5. Testable: Easy to test each layer in isolation

  6. Maintainable:

    • Change interface? → Update interface layer only
    • Change business rules? → Update services layer only
    • Change validation? → Update models layer only
  7. Extensible: Easy to add new models, services, or tools

  8. Interface Independence: Services can be reused with different interfaces (REST API, gRPC, CLI, etc.)

  9. Type Safety: Pydantic models throughout ensure validation at all boundaries

Development

Running Tests

# Run all tests with coverage (recommended)
uv run task test

# Run tests without coverage (quick)
uv run task test-quick

# Coverage is 100%

Type Checking

# Run mypy type checker on all code
uv run task typecheck

# Mypy runs in strict mode on entire codebase

The project uses mypy in strict mode across the entire codebase (src/, tests/, examples/) to ensure complete type safety.

Test Organization

Tests are organized by layers, mirroring the source structure (100% coverage):

Models Layer Tests (tests/models/)

  • test_thought.py: Model properties, auto-adjustment, formatting, validation, and confidence scoring
  • test_session.py: Session logging and formatted output

Services Layer Tests (tests/services/)

  • test_thinking_service.py: Service validation, functionality, branching, edge cases, response format, reference validation, and multi-session support

Interface Layer Tests (tests/interface/)

  • test_mcp_server.py: MCP tool function invocation

Root Tests

  • test_cli.py: CLI entry point

Credits

This project is a Python port of the Sequential Thinking MCP Server by Anthropic, part of the Model Context Protocol servers collection. The original implementation provides the foundation for structured thinking and problem-solving.

New Features

While maintaining full compatibility with the original design, UltraThink adds several enhancements:

  1. Confidence Scoring - Explicit uncertainty tracking with 0.0-1.0 scale for each thought
  2. Auto-assigned Thought Numbers - Optional thought numbering (auto-increments if omitted)
  3. Multi-Session Support - Manage multiple concurrent thinking sessions with session IDs
  4. Assumption Tracking - Make reasoning transparent with explicit assumptions, dependencies, and invalidation tracking

License

MIT

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured