think

This MCP server implements Anthropic's think tool as an attachable reasoning step for Claude-like assistants. When integrated into an MCP workflow, the think tool can be invoked to pause the generation and record an intermediate reasoning segment without altering external data sources. This can help with multi-step problem solving, policy-aligned decision making, and more deliberate tool usage. To use it, configure your MCP client to expose the think tool to the AI and ensure prompts instruct the model when to call think, what to include as the thought, and how to proceed after the thinking step. You can enable think in your prompts by referencing the tool name think and providing a thought field that captures the intermediate reasoning. The server will log the think content and continue with the final answer, keeping external state unchanged.

Prerequisites:

• Node.js and npm installed on your machine
• Access to run npm/npx commands

1. Install dependencies and build (development flow):

npm install
npm run build

2. Run the MCP server using npx (as configured in the server config):

npx -y think-mcp-server

3. Alternative installation via Smithery (if you want Claude Desktop integration):

npx -y @smithery/cli install @marcopesani/think-mcp-server --client claude

4. Docker (optional):

• Build: docker build -t think-mcp-server .
• Run: docker run -it think-mcp-server

Prerequisites for Docker include having Docker installed and access to the repository to mount sources if developing.

Notes and tips:

• The think tool is designed to be a non-intrusive reasoning step that appends to a log rather than changing state.
• In MCP client configurations, you may want to set autoApprove for think if your workflow consistently uses it; otherwise, prompts may require explicit consent.
• When debugging, MCP Inspector tooling (e.g., npm run inspector) can help reveal how think is invoked during a session.
• The server exposes the think tool with a schema that requires a thought string; ensure prompts align with this expectation to maximize utility.
• If integrating with desktop clients (Claude Desktop, Cline, Cursor), follow the example JSON blocks provided in the README to populate your client config.