Mastra remote sandboxes: Daytona and E2B tutorial

Alex Booker demonstrates remote sandbox support in Mastra, an open source TypeScript framework. The examples give an AI agent a separate environment for shell commands and generated files. These demos use remote sandbox providers rather than showing how to run models locally.

The first example adds a Daytona sandbox to the default weather agent. It converts temperatures with awk, saves weather.txt, and resumes a stopped sandbox to read the file. Booker configures a one-minute inactivity pause and explains the ephemeral option for deleting a sandbox after use.

An activity planner uses E2B with Linux and Python packages prepared through a template. A Mastra workspace mounts S3 storage for activity guides, skills, and output files. The agent searches Markdown guides with BM25, checks the weather, and uses Anthropic's PDF skill to create a London itinerary and add a watermark. The sample activities are fictional. Booker also describes support for vector and hybrid search.

The final coding assistant demo assigns a Daytona sandbox to each user ID. A starter snapshot supplies the application setup, while a background server provides a preview link with hot module reloading. Booker presents sandbox isolation as a security benefit; the demonstration does not establish a universal guarantee against prompt injection or data exposure.