Another week, another round of big changes for Mastra! This week, we’re excited to announce a major licensing update, new features for Playground and Memory, and a bunch of other updates.
Let’s get into it:
Mastra is now licensed under Apache-2.0
Big news — we've moved Mastra to the Apache-2.0 license.
This gives you complete freedom to use, modify, and distribute Mastra in any project. Build commercial applications, fork the repo, contribute back — it's all fair game.
We're excited to see what the community builds next. Read our full blog post on the transition here.
Playground: now multi-modal
We shipped an upgrade to playground: it’s now multi-modal. You can test agents with images, audio, and text.
For example: Upload an image and see how your agent interprets it. Send audio clips and watch the transcription happen in real-time.
Memory: new SaveQueueManager method
We've added SaveQueueManager to handle message persistence during streaming and generation. This solves a tricky problem: how do you save streaming data without losing messages or hitting race conditions?
The manager batches saves intelligently and debounces operations, dramatically reducing data loss risk. Agent conversations should now be more reliable, especially during high-throughput scenarios.
RAG: new relevance provider
We've added the ZeroEntropy relevance provider for better relevance scoring in your RAG pipelines. This means more accurate document retrieval and better context matching for your agents.
Other updates
- We added support for configuring CloudflareD1Store with a custom D1 client implementation via a new D1ClientConfig option. This enables integration with custom HTTP endpoints while maintaining backward compatibility.
- We now support passing a function that returns the sleep duration or date to the
sleep()andsleepUntil()workflow methods.