Mastra

Mastra is an open-source TypeScript framework for building AI agents and workflows, made for developers, with an optional hosted platform billed by usage.

Mastra handles the plumbing behind an agent that has to run for days: calling models, invoking tools, remembering earlier conversations and pausing for a person to approve a step. Kepler Software Inc., a Delaware corporation, sells it under the Mastra name, and the project started in 2024. The code lives in a public GitHub repository under the Apache 2.0 license, so a team can read, fork and run every part of it. Mastra Platform adds hosting, tracing and team controls for anyone who would rather not operate those servers. An Agent in Mastra is defined in code with instructions, a model, tools built with createTool, and memory. Each agent can use a different model provider. Workflows chain steps, then suspend at any point and resume later from a saved snapshot. Supervisor agents hand tasks to subagents, and the RAG module pulls context from a team's own documents. Input and output processors act as guardrails, while structured output forces replies into a schema. Scorers grade agent answers and can run in a CI pipeline. Tracing records how each request moves through agents, workflows, tools and model calls, and the Editor versions stored agents and prompt blocks through draft and publish steps. The people who get the most from it are TypeScript and JavaScript engineers who review and test agent logic like the rest of their application. A startup can build its first agent on the framework without paying anything. Larger product teams move to Mastra Platform when they need longer retention, single sign-on and several teams under one account. Hosting is bought as a flat monthly plan plus metered usage beyond each allowance, and Enterprise terms are quoted. Companies that keep every run inside their own network can instead license the self-hosted edition for a flat annual fee. Most deployments start from a concrete job. A support agent answers customers in Slack or Discord through Mastra's channels, and a research agent searches internal documents before it replies. An approval workflow waits for a manager's sign-off, then picks up exactly where it stopped. Webhook signals wake an agent when an outside system posts an event. The Mastra server publishes agents as HTTP endpoints for the application to call, or as an MCP server that other AI clients can use. Chat front ends connect through the AI SDK UI hooks. Mastra fits teams for whom an agent is a feature of the product they ship, written and deployed beside the rest of their TypeScript code.