Cloudflare Workers AI

Executes LLM inference (Llama 3, Gemma 3, Mistral) across Cloudflare's 190+ global edge locations, routing requests to the nearest datacenter for sub-100ms response times. Uses Workers compute runtime paired with optimized model serving infrastructure, eliminating centralized API bottlenecks. Supports streaming responses via WebSocket for real-time token delivery.

Enables LLMs to invoke external tools and APIs through a declarative schema registry, with automatic model-specific formatting (OpenAI function_calling, Anthropic tool_use, etc.). Supports synchronous tool execution, multi-step reasoning chains, and model fallback via AI Gateway when primary model fails. Built on Workers compute for stateless execution and Durable Objects for multi-turn state persistence.

Enables agents to generate images using built-in image generation models (specific models not documented). Agents can specify generation parameters (style, size, quality, etc.) and receive generated images as outputs. Images are stored in R2 for persistence and can be returned to users via HTTP or embedded in agent responses.

Provides built-in embedding generation that converts text into vector representations for semantic search and similarity matching. Embeddings are generated using a built-in model (specific model not documented) and can be stored in Vectorize for later retrieval. Supports batch embedding generation for processing multiple texts efficiently.

Deploys agents as serverless functions on Cloudflare Workers, automatically scaling to handle traffic spikes without manual provisioning. Agents are deployed to 190+ edge locations globally, ensuring low latency for users worldwide. Billing is based on actual usage (requests, compute time) with no minimum fees or reserved capacity. Deployment is triggered via Git push or API, with automatic rollback on errors.

Provides integrated object storage (R2) for persisting agent outputs, training data, checkpoints, and user uploads. R2 is replicated globally and offers zero egress costs (no charges for downloading data), making it cost-effective for storing large files. Agents can read and write to R2 directly, and files can be served via HTTP or embedded in agent responses.

Persists agent conversation state, memory, and execution context in a built-in SQL database per agent instance, with automatic client-side state synchronization via WebSocket. Uses Durable Objects as the state coordination layer, ensuring consistency across multiple Workers instances and preventing race conditions in multi-turn conversations. Supports both server-side state (agent reasoning, tool call history) and client-side state (UI context, user preferences).

Enables agents to receive and respond to user input via multiple channels—WebSocket for real-time chat, email for asynchronous communication, and voice for audio-based interaction. Each interface is abstracted through a unified agent API, allowing the same agent logic to serve multiple input modalities without channel-specific code. Voice input is processed via Whisper speech-to-text, and responses can be delivered as text-to-speech audio.

Provides a built-in RAG pipeline (AI Search) that automatically indexes documents and web content, enabling agents to retrieve relevant context without manual embedding or vector database setup. Supports fresh data by re-indexing on-demand, and integrates with Vectorize for vector storage and semantic search. Agents query the index via natural language, and retrieved documents are injected into the LLM context window automatically.

Provides a managed vector database (Vectorize) that stores and retrieves embeddings across Cloudflare's global network with automatic replication. Integrates natively with Workers AI for embedding generation and AI Search for RAG. Supports semantic search queries, filtering by metadata, and batch operations. Vectors are replicated globally for low-latency retrieval from any edge location.

Provides a proxy layer (AI Gateway) that sits between agents and LLM inference endpoints, implementing request caching, rate limiting, and model fallback. Caches identical prompts to avoid redundant inference calls, applies per-user or per-IP rate limits, and automatically routes requests to fallback models if the primary model is unavailable. Supports observability features (logging, metrics) for monitoring inference usage.

Enables agents to execute long-running tasks (hours or days) asynchronously without blocking the user request. Uses Durable Objects to coordinate workflow state, Workers to execute tasks, and R2 for storing intermediate results and checkpoints. Agents can pause, resume, and checkpoint progress, allowing recovery from failures without restarting from the beginning. Supports email or webhook notifications when workflows complete.

Enables agents to connect to remote MCP servers (e.g., GitHub, Slack, databases) using the Model Context Protocol standard. Agents authenticate via OAuth 2.1 with granular permission scoping, allowing users to authorize specific capabilities (read-only, write, delete) without exposing full credentials. Includes an MCP playground for testing server connections and a built-in OAuth provider implementation for custom MCP servers.

Provides built-in speech processing via OpenAI's Whisper model for converting audio to text, and text-to-speech (TTS) synthesis for converting text responses to audio. Both are integrated into the agent runtime, allowing agents to receive voice input and deliver audio responses without external speech services. Supports multiple audio formats (WAV, MP3, etc.) and languages for Whisper.