Cloudflare AI Search adds one-command indexing and public endpoints

Cloudflare has rolled out a series of developer experience improvements to AI Search that eliminate the need to manually assemble Workers AI, Vectorize, R2, and Browser Run components. The platform now handles the full stack automatically, according to details first reported by Cloudflare.

The company positions the service as giving agents their own search engine to retrieve data for better responses. Alongside the feature updates, Cloudflare shared preview pricing designed to scale predictably by making embedding and reranking free when using default models.

What's new in AI Search

Developers can now index structured and unstructured data from individual files to entire websites with a single command. Previously, websites required a sitemap, but a new "Discover" parsing option crawls sites by following links using Browser Run's /crawl functionality.

Public endpoints represent a significant workflow simplification. Enabling public URLs on a namespace generates /search and /mcp endpoints that query multiple instances or websites simultaneously without authentication. Developers can add custom domains to brand these endpoints—for example, search.example.com/mcp—and layer Cloudflare Access on top to create private search instances.

The platform also ships with an AI Search plugin for EmDash, Cloudflare's open-source CMS, adding semantic search capabilities to sites built on that framework.

How Cloudflare uses it internally

Cloudflare built its own Dev Stack MCP server using AI Search, which is now available in the company's AI Playground. The implementation indexes ten separate surfaces—including Docs, Blog, API Docs, Community, and framework documentation for Astro, Vite, Vitest, Hono, Replicate, and OpenNext.

Creating an instance requires one command. For sites without sitemaps, developers add a --parse-type discover flag. The system then handles crawling, ingestion, embedding, and retrieval automatically.

Querying across all instances happens through either a Worker binding or by enabling public endpoints. The Worker approach involves binding the namespace and making a single multi-instance call that fans out across all sources. Results return cited and tagged with their source instance. The public endpoint option requires no code—flipping on public URLs immediately generates shareable /search and /mcp endpoints.

Cloudflare Blog, Developer Docs, and Cloudflare.com all run on AI Search, using hybrid search that combines semantic and keyword matching in one query.

Preview pricing structure

AI Search remains free during beta, with billing not yet enabled. Preview pricing covers ingestion at $0.75 per million tokens (with 5 million tokens included), storage at $2.00 per GB-month (10 GB included), and queries starting at $0.10 per thousand for full-text and $0.75 per thousand for semantic searches (2,000 queries included).

Image processing adds $0.50 per million tokens. Embedding and reranking run free with select Workers AI models, removing a major variable cost from the equation.

A sample monthly bill for indexing 20,000 documents plus 1,000 images with 30,000 semantic queries would total approximately $35, with most of that cost coming from queries rather than the largely one-time indexing expense.

Why it matters

The shift from assembly-required primitives to automated indexing lowers the barrier for developers building agent-powered applications that need current, cited information. By making embedding and reranking free with default models, Cloudflare removes the token-counting uncertainty that complicates cost forecasting for semantic search implementations. The public endpoint option particularly matters for teams that want to ship search functionality without writing or maintaining query infrastructure.

The AI Search bot identifies itself as Cloudflare-AI-Search, follows robots.txt, and respects site bot controls. Developers can enable and configure AI Search through the Wrangler CLI, with full documentation available on Cloudflare's developer site.

These details were announced by Cloudflare on their blog.