Cloudflare AI Cloud

Q: Does Cloudflare train AI on customer data?

Cloudflare does not train large language models (LLMs) on customer data, ensuring that data remains private and is not used for model training.

Q: Is there a free tier for Cloudflare AI?

Workers AI is available on both Free and Paid plans, offering a serverless, pay-for-what-you-use pricing model.

Q: How do I build AI agents on Cloudflare?

The Agents SDK allows developers to build goal-driven agents that can call models, use tools via MCP, and maintain state in a built-in SQLite database.

Q: What is the purpose of AI Gateway?

AI Gateway provides observability and control features including request caching, rate limiting, and model fallback to manage costs and performance.

Product · API / developer platform · Engineering & development by Cloudflare

The full-stack platform for building and scaling AI agents globally.

A serverless AI platform for developers to build, deploy, and scale AI agents and applications globally.

Launched: 2024-04-21

Category: AI Infrastructure

GPU Locations: 190+ cities
Maximum Context Window: 1,000,000 tokens
Inference Latency: <100 ms worldwide
Agent Storage: Built-in SQLite per instance
Supported Modalities: Text, Image, Audio, Video, Voice
Max File Size (AI Search): 10 MiB

Serverless GPU inference in 190+ locations
Stateful AI agents with built-in SQLite memory
Zero-egress storage for training data via R2
Up to 1M token context window on flagship models

Overview

Cloudflare AI Cloud is a comprehensive developer platform for building and deploying AI applications and stateful agents on a global serverless network. It provides businesses and developers with serverless GPU inference in over 190 locations, ensuring low-latency responses under 100ms worldwide without the need for cluster management. Developers can build proactive agents using the Agents SDK, which includes built-in SQLite state management and Model Context Protocol (MCP) integration for tool use. The suite is integrated with Cloudflare's broader ecosystem, including Vectorize for RAG workflows and AI Gateway for observability and cost control.

Inference and Models

Serverless GPU Inference: Developers can run machine learning models on serverless GPUs across Cloudflare's global network in over 190 locations.
Multi-interface Access: Models can be invoked directly from Workers, Pages, or via the Cloudflare REST API from any environment.
Specialized AI Models: The platform supports specialized decision models like Clef and extraction models like Clairic for structured agentic workflows.

Agent Infrastructure

Agents SDK: The Agents SDK provides primitives for building proactive, stateful AI agents with built-in memory and scheduling.
Built-in State Management: Every agent instance includes its own SQLite database to persist conversation history and user preferences automatically.
MCP Integration: Agents can connect to any MCP server to access external tools, APIs, and data sources.
Real-time Communication: The platform supports real-time communication via WebSockets with automatic hibernation and state synchronization.

Data and Observability

Vectorize Database: Vectorize is a globally-replicated vector database that enables semantic search and RAG workflows in a few lines of code.
AI Gateway Control: AI Gateway provides observability and control with built-in caching, rate-limiting, and model fallback for inference calls.
AI Search and RAG: AI Search offers complete RAG workflows with automatic indexing, pixel-direct image search, and OCR for scanned PDFs.
Agent Memory provides persistent, scoped memory for agents to remember users and organizations across conversations.

Use cases

Developers build proactive AI assistants that send scheduled emails and monitor systems without user prompts.
Engineering teams deploy RAG-powered search applications using Vectorize and AI Search for instant data retrieval.
SaaS companies create multi-tenant AI platforms where each customer runs isolated, sandboxed AI code.

Specifications

APIs and SDKs: Cloudflare REST API, Workers AI Binding, Agents SDK, OpenAI compatible endpoints
Platforms: API, Workers, Pages, and any environment capable of making REST API calls
Limits: Supports context windows up to 1 million tokens on select models like Claude Opus 4.7 and DeepSeek V4 Pro
Deployment: Cloud-based serverless execution across 190+ global GPU locations
Integrations: GitHub, GitLab, Model Context Protocol (MCP), Vercel AI SDK, Hugging Face
Languages: Supports 100+ spoken languages for embedding and text generation tasks
GPU Locations: 190+ cities
Maximum Context Window: 1,000,000 tokens
Inference Latency: <100 ms worldwide
Agent Storage: Built-in SQLite per instance
Supported Modalities: Text, Image, Audio, Video, Voice
Max File Size (AI Search): 10 MiB
API Compatibility: OpenAI-compatible endpoints

How it works

Select a Model
Build with SDKs: Use the Agents SDK to define agent behavior, including state management, tool use via MCP, and proactive scheduling.
Deploy Globally: Deploy the AI application or agent to Cloudflare's global network using a single command or the Cloudflare API.
Observe and Scale: Monitor and manage the application using AI Gateway for caching, rate limiting, and performance observability.

Pricing

Workers AI Free Plan: $0 per month
Workers AI Paid Plan: Usage-based pricing

FAQ

Does Cloudflare train AI on customer data?