MCP-Think is an MCP server that implements a dedicated think tool for large language models to record, inspect, and retrieve their internal reasoning steps. The server connects LLM clients, such as Cursor and Claude Desktop, directly to an isolated reasoning scratchpad via standard input and output transport. Developers and prompt engineers use MCP-Think to enforce structured reasoning workflows, policy verifications, and intermediate chain-of-thought documentation before an agent executes tools or produces final outputs. By exposing tools to record thoughts, fetch prior reasoning steps, clear scratchpad memory, and view thought statistics, it prevents complex instructions from polluting standard conversation contexts. Written in Go, it operates locally with minimal overhead and supports standalone binaries across Linux, macOS, and Windows. This setup enables models like Claude 3.7 Sonnet to systematically review multi-step plans, validate user inputs against business logic, and verify tool execution results without requiring external database integrations or complex cloud configurations.
Category: AI & LLM Tooling
Tags: chain-of-thought, context, memory, reasoning
npx -y @smithery/cli install @iamwavecut/mcp-think --client claude, or install via Go using go install github.com/iamwavecut/MCP-Think@latest. 2. Alternatively, download the pre-built binary for your OS from GitHub Releases and ensure the binary think-tool is in your system PATH. 3. Configure your MCP client. For Cursor, add the server to ~/.cursor/mcp.json under mcpServers: json { "mcpServers": { "think-tool": { "command": "think-tool", "transport": "stdio" } } } 4. Add guidance in your client's rules instructing the model to record and review its thinking steps before executing external actions.Providing a dedicated scratchpad for LLMs to verify user inputs and constraints before triggering tool executions. Enforcing multi-step business logic validation and rule checks during automated airline or service booking workflows. Iterating over and sanity-checking raw tool outputs to confirm correctness before formulating a final response to the user. Tracking reasoning statistics and clearing interim thought histories across distinct subtasks within long coding agent sessions.
Part of MCP Servers
MCP-Think is an open Model Context Protocol server that provides a structured thinking tool for large language models. It acts as an isolated scratchpad where models can log reasoning steps, review earlier conclusions, and evaluate complex decisions before responding or running tools.
The server exposes four primary functions: Think Tool to record reasoning steps, Get Thoughts to retrieve previously logged steps, Clear Thoughts to reset the memory buffer, and Get Thought Stats to generate metrics about recorded reasoning data.
You can install it automatically for Claude Desktop using Smithery with npx. Alternatively, download pre-compiled binaries from the GitHub releases page, run the shell install script, or compile it locally using go install github.com/iamwavecut/MCP-Think@latest.
It functions with any MCP-compliant client that supports stdio transport. Tested clients include Cursor and Claude Desktop, where it can be integrated into system prompts and user rules to structure agent planning and reasoning.