Mastra

TypeScript-native AI agent and workflow framework built by the Gatsby team. Provides a typed Agent class with tool calling, conversation memory, and configurable LLM backend. Multi-step Workflow engine with branching, parallel steps, and retry logic. Built-in evals for LLM output quality. First-class Next.js integration. Self-hosted or Mastra Cloud. Confidence is docs-derived.

TypeScript API: new Agent({ name, instructions, model, tools, memory }) for agent definition; agent.generate(messages) for single-turn, agent.stream(messages) for streaming; Workflow class with .step() chaining, .branch(), .parallel(); Tool typed via createTool({ inputSchema: z.object({}), execute }) with Zod validation; MastraClient for remote agent invocation; REST API auto-generated from agent definitions.

No built-in auth layer in the framework — deployer controls access; Mastra Cloud uses account-level API key for deployment; LLM provider credentials (OpenAI, Anthropic, Groq, etc.) injected via environment variables and never stored in agent code; memory backends (PostgreSQL, Upstash) use standard credential patterns; tool credentials injected via env at startup.

Workflow steps are individually retriable with configurable backoff; agent.generate() propagates LLM provider errors as typed exceptions; streaming via agent.stream() uses async generators for backpressure-safe consumption; tool execution failures return structured error objects passed back to the agent for recovery reasoning; eval scores surface per-step quality signals.

Documentation covers Agent quickstart, Tool definition with Zod schemas, Workflow engine guide, Memory backends (PostgreSQL, Upstash), eval setup, and Mastra Cloud deployment. TypeScript-first with strong inference support in VS Code. Active Discord community. Confidence is docs-derived.