wtf is search-mcp?

search-mcp is an open source toolkit built around Cloudflare's AI Search product, designed to let you turn a set of documents or code repositories into a searchable knowledge base that both people and AI agents can query. The toolkit is made of three main pieces. The first is an MCP Worker, a small server that exposes a search tool over the Model Context Protocol so AI coding agents can look things up, protected by a bearer token. The second is a Query Worker, which handles a streaming ask endpoint meant for a browser based widget, with rate limiting, cross origin protection, and an optional Turnstile human check. The third piece is a corpus sync tool, a set of scripts that pull files from git repositories and upload them into a Cloudflare R2 storage bucket, renaming file extensions like TypeScript or Dockerfiles so Cloudflare's AI Search will actually index them instead of skipping them as non text files. The corpus sync tool and the browser widget files are published as an npm package, while the two Workers themselves need to be deployed separately from a cloned copy of the repository, each with its own configuration file for account, storage bucket, and which repositories to index. Once running, the ask widget can be dropped into any documentation site with a small script tag, and the MCP server can be added to any agent tool that speaks the Model Context Protocol, using a bearer token for access control. The project also includes careful handling for keeping the search index up to date: before asking Cloudflare to reindex the content, the sync scripts wait for any already running reindex job to finish and for a short cooldown period afterward to clear, so a new reindex job reliably includes everything that was just uploaded. The code is released under the AGPL v3 license, a copyleft license that requires anyone who runs a modified version of the software as a network service to share their source code changes too.