Mastra

Recently shipped span-level trace queries, token and cost aggregation across observability stores, and self-embedding vector stores for semantic recall.

Observability queries went span-level and cost-aware. storage.querySpans() and client.querySpans() find individual completed spans with filters, cursors, previews, and model cost — "every failed tool_call in the last hour" without locating traces first. aggregateTraces() gained token and cost measures with coverage and currency handling in the ClickHouse, DuckDB, and Postgres stores. Servers also advertise observabilityStorageCapabilities so Studio calls only what a store supports.

Agents now recover from provider failures by default. Three error processors repair history before retrying, disableable via errorProcessorDefaults: false. Streaming emits a tool-call-resumed chunk through @mastra/ai-sdk and @mastra/react, so UIs clear pending approvals before the result lands, and tools can read the full conversation via context.agent.getMessages().

Semantic recall dropped the client embedder. Vector stores that generate their own embeddings, MongoDB among them, work directly. Recall pages before and after a matched observation group, and getObservationalMemoryHistory gained group filtering, sort order, and recordId lookup.

Classifiers, per-request model routing, and live channels remain core. Register classifiers on new Mastra({ classifiers }) and use them as typed workflow steps; ClassifierProcessor fails closed by default. ModelSelectionProcessor picks the best or cheapest model per run. Channel connections resolve at runtime, so platform changes appear without redeploy. @mastra/connect added Google Docs, Drive, and Sheets toolsets.

Also shipped: @mastra/mcp@2.0.0 on the stateless 2026-07-28 MCP revision, plus @mastra/azure-ai-search and @mastra/weaviate vector stores.