MCP Scratchpad Server
An MCP (Model Context Protocol) server that provides a "think" tool for structured reasoning, based on the Anthropic blog post.
Overview
The Scratchpad MCP server allows Claude and other LLMs to add dedicated thinking steps during complex tool use scenarios. This improves problem-solving by creating space for structured reasoning and helps the model carefully process tool call outputs.
Features
Think Tool: Add thoughts to a timestamped log for structured reasoning
Thoughts Log Resource: View all logged thoughts in JSON format
Clear Tool: Reset the thinking log when needed
Structured Thinking Prompt: Template for systematic problem-solving
Installation
Quick Start (Recommended)
Run directly with uvx (no installation needed):
PyPI Installation
Install globally with pip or uv:
Development Installation
For local development:
Usage
With Claude Desktop
Add to your Claude Desktop configuration:
On macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
On Windows: %APPDATA%/Claude/claude_desktop_config.json
Using uvx:
Using pip installation:
Testing with MCP Inspector
For development and testing:
Tools
think
Add a thinking step to the log.
Parameters:
thought
(string): The thought to log
Example:
clear_thoughts
Clear all thoughts from the log.
No parameters required
Resources
log://thoughts
Returns all logged thoughts as a JSON array with timestamps.
Example output:
Prompts
Structured Thinking
A template for systematic problem-solving using the think tool.
Parameters:
topic
(string): The topic or problem to think through
Use Cases
The scratchpad tool is particularly useful for:
Long chains of tool calls: Keep track of intermediate results and reasoning
Complex problem-solving: Break down problems systematically
Policy-heavy environments: Verify compliance with guidelines
Sequential decision making: Document the reasoning process
Edge case handling: Think through unusual scenarios
Example Workflow
Start with a complex problem
Use the "Structured Thinking" prompt to set up the approach
Use the
think
tool multiple times to log reasoning stepsCheck progress with the
log://thoughts
resourceClear the log when starting a new problem
Best Practices
Use descriptive thoughts that explain the "why" not just the "what"
Break complex problems into smaller, manageable parts
Validate reasoning at each step
Summarize conclusions at the end
Clear the log between unrelated tasks
Development
To modify the server:
Fork the repository
Make your changes to
src/mcp_server_scratchpad/server.py
Test with
uv run mcp dev src/mcp_server_scratchpad/server.py
Submit a pull request
Future Enhancements
Potential improvements:
Persist thoughts to a file
Add search/filter capabilities
Support for thought categories or tags
Export thoughts in different formats
Integration with other reasoning tools
License
MIT License - see LICENSE file for details
Credits
Based on the thinking tool concept from Anthropic's blog post
This server cannot be installed
Provides a 'think' tool that allows Claude and other LLMs to add dedicated thinking steps during complex tool use scenarios, creating space for structured reasoning and improving problem-solving capabilities.
Related MCP Servers
- -securityAlicense-qualityAn MCP server that implements the 'think' tool, providing Claude with a dedicated space for structured thinking during complex problem-solving tasks to improve reasoning capabilities.Last updated -130MIT License
- AsecurityFlicenseAqualityOfficial implementation of Anthropic's 'think' tool that provides Claude with a dedicated space for structured reasoning, improving performance by up to 54% on complex tasks requiring multi-step problem solving.Last updated -1320119
- AsecurityAlicenseAqualityImplements Anthropic's 'think' tool for Claude, providing a dedicated space for structured reasoning during complex problem-solving tasks that improves performance in reasoning chains and policy adherence.Last updated -47534MIT License
- -securityAlicense-qualityEnables Claude's explicit thinking mode in Cursor, allowing users to see Claude's step-by-step reasoning process when prefacing queries with 'think'.