What is docket-mcp?

Docket MCP turns meeting transcripts into ready-to-file task items, but always with a human checking the list before anything gets created. You paste in a transcript, and an AI agent using Google's Gemini 2.5 Flash model reads through it and pulls out action items along with who is responsible, when it is due, and how confident it is about that assignment. Before creating anything, it checks each proposed task against your existing Notion tracker to see if something similar already exists, so you are not creating duplicate work. The tracker connection is built using the Model Context Protocol, or MCP, an open standard that lets AI agents call outside tools through a clean, typed interface without needing to know how those tools work internally. In this project, that means the logic for talking to Notion lives in one server, and both this app's own agent and any other MCP compatible tool, like Claude Desktop or Claude Code, can use the same three functions: listing open tasks, searching for similar existing tasks, and creating a new task. Each extracted item is scored for similarity against open tasks: a high similarity score marks it as a likely duplicate that gets skipped by default, a middle range asks a human to decide, and a low score marks it as new and pre-checked for creation. Everything is shown in a review screen where you can check or uncheck items before clicking to create the approved ones in Notion, after which you get a receipt listing exactly what was filed and to whom. Setting it up needs Node.js and Python installed, cloning the repository, installing both Node and Python dependencies, and configuring environment variables for a Gemini API key and Notion credentials. The project also includes a demo mode that works with no API keys at all, using a simpler rule-based extractor and a fake Notion tracker so you can try the whole flow immediately.