Cloudflare AI Capabilities: The Complete 2026 Guide

Most people know Cloudflare as a CDN and DDoS shield. By 2026 it is also one of the most complete places to build an AI application or agent end to end, and you can do a surprising amount of it inside a free tier. Cloudflare’s AI capabilities are a full stack, not a single product: you can run models, cache and govern LLM calls, do retrieval, build stateful agents, give those agents web and voice tools, and secure the whole thing, all on the same serverless network that already sits in front of a large chunk of the internet.

This guide walks the entire lineup, grouped by what each piece actually does, with free tiers and the developer problem each one solves. If you are deciding whether to build your next chatbot or agent on Cloudflare, this is the map.

What is Cloudflare’s AI platform?

It is a set of serverless building blocks that run on Cloudflare Workers and its global edge network. Instead of renting GPUs, standing up a vector database, and bolting on an observability layer, you compose managed primitives that already talk to each other. The pieces fall into six groups:

• Inference - run the models (Workers AI)
• LLM operations - control and observe model calls (AI Gateway)
• Retrieval - RAG and vector search (Vectorize, AI Search)
• Agents - stateful, durable agent runtime (Agents SDK, Workflows)
• Agent tools - what agents can do (Browser Rendering, Realtime, MCP)
• Security and control - protect apps and govern AI crawlers (AI Security for Apps, AI Crawl Control)

Cloudflare AI products at a glance

How do you run AI models on Cloudflare?

With Workers AI, Cloudflare’s serverless inference service. You call a model through a Worker binding or a REST endpoint, and Cloudflare runs it on GPUs across its network with no infrastructure to manage. The catalog runs to more than 50 open models across four families: text and LLMs (Llama, Mistral, DeepSeek, Gemma, Qwen), embeddings (BGE), audio (Whisper for speech-to-text), and image generation (Flux and friends). You can bring your own fine-tune through LoRA adapters on supported base models, and there is an asynchronous batch inference API for large jobs.

Pricing is built on Neurons, Cloudflare’s unit of AI compute. You get 10,000 Neurons per day free, and beyond that it is roughly $0.011 per 1,000 Neurons on a Workers Paid plan. Each model burns Neurons at its own rate per token or per audio-minute, and the dashboard now shows per-model unit pricing on top. For a low-traffic chatbot or a prototype, you can often stay inside the free daily allocation entirely. Reach for Workers AI when you want low-latency, pay-per-use inference sitting next to your app and data, with open models rather than a single proprietary API. If you are optimizing spend, our guide on how to cut LLM costs for production chatbots and agents covers where edge inference helps and where it does not.

What is Cloudflare AI Gateway?

AI Gateway is a control plane that sits between your app and any LLM provider. You point your existing OpenAI, Anthropic, or other client at a single Gateway endpoint and immediately get analytics and cost tracking, response caching, rate limiting, and automatic retries or fallback between models. Its Dynamic Routing feature lets you define flows that evaluate conditions and pick a model with fallbacks, instead of hard-coding one. On the safety side, Guardrails scan prompts and responses for prompt injection, PII, and unsafe topics.

The core features - analytics, caching, rate limiting - are free. Caching alone can meaningfully cut spend on repetitive traffic, and guardrails are billed as Workers AI token inference. If you are weighing it against other options, we put it head to head in LLM gateways compared: LiteLLM vs Portkey vs Helicone vs Cloudflare AI Gateway. Reach for AI Gateway the moment you are calling a paid model in production and need visibility, caching, and failover without touching client code.

How does Cloudflare do RAG and vector search?

Two ways, depending on how much control you want.

Vectorize is the vector database. It stores embeddings close to your users for low-latency semantic search, supports up to around 5 million vectors per index and 1,536 dimensions per vector, and integrates tightly with Workers AI embeddings, R2, and D1. The free tier covers 5M stored and 30M queried dimensions a month, then about $0.01 per million queried dimensions. Use it when you want to own the RAG pipeline - your own chunking, your own ranking.

AI Search is the managed RAG primitive, renamed from AutoRAG in late 2025. You point it at an R2 bucket or upload files directly, and it handles chunking, embedding, continuous re-indexing, retrieval, and answer generation for you. It runs on R2 plus Vectorize under the hood, ships a built-in MCP server on every instance, and can now use OpenAI and Anthropic models in addition to Workers AI. It is free during its open beta, with the underlying Workers AI and AI Gateway usage billed separately. Reach for AI Search when you want a “chat with your docs” experience without wiring the pipeline yourself. This is exactly the pattern behind a modern cost-efficient lead-gen chatbot.

How do you build AI agents on Cloudflare?

This is where the 2026 platform is pointed. The Agents SDK builds stateful agents on Workers, where each agent is a Durable Object. That gives every agent persistent state, its own SQLite storage, WebSockets for real-time UIs, and alarms for scheduling, so an agent can be long-lived and pick up where it left off rather than being a stateless request. Durable Objects gained a free tier in 2025, which lowered the barrier to this considerably.

For reliable multi-step work, Workflows provides durable execution: pipelines that retry, sleep, pause for human approval, and resume after failures without losing progress. It is billed on active CPU time, so a workflow that is waiting costs nothing, and the free tier covers roughly 3 million requests a month. Together, Agents SDK and Workflows let you build an agent that calls a model, invokes a tool, waits for a signal, and continues, all without a server you manage. If you are comparing agent frameworks more broadly, see our AI agent framework comparison and Claude Agent SDK vs OpenAI Agents SDK.

What tools can Cloudflare agents use?

An agent is only as useful as what it can do. Cloudflare gives agents three big capabilities:

• Browser Rendering - a serverless headless browser with both Puppeteer and Playwright support, so an agent can navigate, scrape JavaScript-rendered pages, screenshot, and generate PDFs. It powers open-source natural-language browsing (Stagehand) and Playwright MCP. Pricing is about $0.09 per browser-hour, with a free daily allowance.
• Realtime (formerly Cloudflare Calls) - WebRTC infrastructure for voice and video, including a Realtime Agents runtime that orchestrates a speech-to-text, LLM, text-to-speech voice pipeline with interruption handling. It is the backbone for building a voice agent on Cloudflare, and it hooks into providers like ElevenLabs. If voice is your channel, compare it against dedicated platforms in our AI voice agent platforms comparison.
• Remote MCP hosting - deploy an MCP server to Workers from an official template and any MCP client can call your tools over the network. The free tier covers 100,000 requests a day with sub-5ms cold starts.

What are the supporting data services?

AI apps need somewhere to keep data, and the same primitives that power Cloudflare’s other products do double duty here:

The zero egress fees on R2 matter more than they sound: RAG and training pipelines move a lot of data, and egress charges are where other clouds quietly bill you.

How does Cloudflare secure AI apps and control AI crawlers?

Two distinct problems, two sets of tools.

For protecting your own AI app, AI Security for Apps (formerly Firewall for AI) sits in front of your LLM endpoints as part of the WAF. It scores each prompt for injection likelihood, detects PII exposure, flags unsafe topics, and even discovers AI endpoints across your property, all mitigated through the familiar WAF rule builder. Combined with AI Gateway’s guardrails, you get defense at both the network edge and the model call.

For controlling who trains on your content, AI Crawl Control (formerly AI Audit) governs which AI crawlers can access your site, with a shift toward opt-in. Pay Per Crawl lets publishers charge AI companies per fetch, and AI Labyrinth traps misbehaving crawlers in a maze of decoy pages. If your business depends on original content, these turn “AI is scraping me for free” into a policy you set.

What can you build with it?

The point of a full stack is that the pieces combine. Five common builds:

1. Managed RAG chatbot - AI Search over R2 docs, plus Workers AI for generation and AI Gateway for caching and observability. Minimal glue code.
2. Custom-control RAG - Vectorize as the vector store, Workers AI embeddings, and Hyperdrive to pull structured data from your Postgres, when you need to own the pipeline.
3. Autonomous web-research agent - Agents SDK for stateful memory, Browser Rendering to navigate and scrape, Workflows for durable multi-step runs, and AI Gateway for model fallback.
4. Real-time voice assistant - Realtime for WebRTC, a speech-to-text, LLM, text-to-speech pipeline via Workers AI or ElevenLabs, backed by a Durable Object for session state.
5. Secure public LLM app - a model behind AI Gateway guardrails and AI Security for Apps, with AI Crawl Control governing who trains on the output.

Should you build your AI stack on Cloudflare?

Cloudflare is a strong default when you want low-latency, pay-per-use AI co-located with your app, a genuinely cheap path to prototype (most pieces have real free tiers), and a single network for inference, retrieval, agents, and security. It is especially compelling for stateful agents, because the Durable Objects model fits agent memory naturally, and for anyone already serving traffic through Cloudflare.

It is less of a fit if you need the largest frontier proprietary models as first-class citizens (you would route those through AI Gateway rather than run them on Workers AI), or if your workload needs long-running heavy GPU training rather than inference. And because this space moves fast, with major launches at Cloudflare’s Birthday Week and Agents Week each year, some of the newest agent features are still early. The stable core, though - Workers AI, AI Gateway, Vectorize, AI Search, the Agents SDK, Workflows, and Browser Rendering - is production-ready today.

The honest takeaway: for building AI chatbots and agents in 2026, Cloudflare gives you more of the stack, cheaper to start, in one place, than almost anywhere else. The trade-off is a fast-moving, open-model-first platform rather than a single-vendor frontier-model shop. If you want help choosing and wiring the right pieces, our AI Agent Development and Enterprise AI Integration teams do exactly this.