AI SDK Patterns
that compose into apps

Stream text responses from AI models in real-time using streamText and the useChat hook.

Generate typed, validated JSON data using Zod schemas and the generateObject function.

Let the AI model call functions and use external tools with the multi-step tool loop.

Render dynamic React components from AI tool calls using streamText and useChat.

Build an agent that orchestrates multiple tool calls in sequence, using the AI SDK's maxSteps and automatic tool result forwarding.

Integrate third-party search APIs (Tavily, Exa) with the AI SDK to build a grounded search agent with source citations.

Pause AI execution to get human approval before executing sensitive tool calls. Uses tool confirmation flow with the AI SDK.

Retrieval-Augmented Generation: embed documents, store in a vector database, and retrieve context for grounded AI responses.

Show the AI's chain-of-thought reasoning in a collapsible thinking section, then display the final answer — like Claude's extended thinking.

Generate images from text prompts using generateImage. Enter a description, click generate, and see the AI-created image.

AI-powered code generation with syntax highlighting and copy support. Describe what you need, and the AI generates clean, runnable code.

A chat interface that includes inline source citations with expandable source cards. Responses include numbered references that link to source details.

Stateful conversation agent with Workflow DevKit that persists messages across pauses/resumes, calls tools, and streams responses using DurableAgent.

Connect to any Model Context Protocol server, discover tools dynamically, and let the AI use them in a chat interface. Includes a demo MCP server with weather and calculator tools.

Natural language to SQL: ask questions about data in plain English, the AI generates and executes SQL queries, and displays results in a table. Uses an in-memory SQLite database with sample data.

Chat with images, files, and text in a single conversation. Drag-and-drop or paste images for vision analysis, attach files for context, and get AI responses that understand all modalities.

Create a Mastra agent with tools, register it in a Mastra instance, and call it with generate() or stream().

Create typed tools with Zod schemas that agents can call. Includes input validation, output schemas, and error handling.

Build multi-step workflows with createStep and createWorkflow. Chain steps with .then(), run in parallel, or branch conditionally.

Add conversation memory to agents with message history, working memory for user preferences, and semantic recall.

Build a retrieval-augmented generation pipeline with document chunking, vector embeddings, and semantic search.

Compose multiple agents into a network where a router agent delegates tasks to specialized sub-agents.

Suspend a workflow mid-execution to collect human approval, then resume with the decision. Essential for agent guardrails.

Use background agents to maintain a dense observation log that replaces raw message history. Keeps context small while preserving long-term memory.

Force agents to return typed, validated JSON using Zod schemas. Perfect for extraction, classification, and data transformation tasks.

Add input/output guardrails to agents for content filtering, PII detection, and response validation before they reach users.