What is tinysearch?

TinySearch is a local web research tool designed to give AI language models the ability to look things up on the internet without requiring a cloud service or account. It is built around MCP, the Model Context Protocol, a standard that lets AI agents call external tools, so AI coding assistants or chat clients can simply ask TinySearch to research a topic. When you give TinySearch a question, it runs a DuckDuckGo web search, scores and filters the best looking results, then crawls those web pages and extracts the most relevant chunks of text using a combination of semantic search, meaning based matching, and keyword search called BM25. It does not directly answer your question. Instead it assembles a structured prompt containing the relevant excerpts with source URLs attached, and hands that back to your language model to produce a cited answer. This keeps responses grounded in actual web content rather than only the model's training data. There is no hosted dashboard, no login system, and no remote analytics. Everything runs on your own machine. You can run it as a Docker container and connect to it over HTTP, or run it locally and have your AI client launch it directly as a subprocess using stdio. An optional FastAPI HTTP server is also included for calling it from tools that do not speak MCP. The embedding models used for semantic search can run locally as ONNX files, a portable format for running AI models efficiently, with several presets ranging from fast and lightweight to higher quality, or you can point it at an OpenAI compatible embedding API instead. Configuration such as search weights, chunk sizes, and crawl concurrency is controlled through a JSON config file. The project is written in Python and ships as a ready to run Docker image.