Framework Update | Mastra @mastra/core 1.62.0

🌏 中文版

Version Info

Why This Release Matters

The biggest change in Mastra 1.62 is turning desktop control into a native workspace capability. Previously, a Mastra sandbox could only run code or shell commands. This release adds a SandboxComputer capability through Daytona and a new E2B Desktop provider. Agents now get 11 additional tools — screenshot, click, double-click, right-click, drag, type, press key, scroll, read screen info, and wait — so they can operate a virtual desktop directly. This fills a gap the agent framework community has been working on: a code sandbox alone isn't enough for tasks like testing a web UI or driving a legacy system with no API, where an agent still needs to look at the screen and move a mouse. The same release adds two new storage backends, @mastra/elasticsearch and @mastra/valkey, giving memory, workflow snapshots, and semantic recall more production storage options. But this is also one of the releases with the most breaking changes recently — background task storage drops Cloudflare KV and ClickHouse support outright, so projects on either backend need a migration plan before upgrading.

Notable Changes

• Computer-Use Sandboxes (SandboxComputer capability): workspace sandboxes gain an optional desktop-control capability. Through Daytona or the new @mastra/e2b-desktop provider, 11 mastra_workspace_computer_* tools are exposed (screenshot / click / double-click / right-click / move / drag / type / press key / scroll / get screen info / wait) → agents can now operate a virtual desktop directly instead of being limited to code/shell; screenshots are returned to the model in native media format
• New storage backends: @mastra/elasticsearch supports memory, workflow snapshots, scores, and semantic recall; @mastra/valkey/@mastra/valkey-streams provide GLIDE-backed storage with PubSub and lease support → another round of expanded production storage options
• Sandbox lifecycle management: new getEnv()/setEnv() manage runtime environment variables inside a sandbox; Mastra.shutdown() now suspends remote sandboxes instead of destroying them outright; Workspace.stop() supports suspend/resume without a full teardown-and-rebuild → lowers the risk of losing state from accidentally killing a sandbox during long-running agent runs
• Observability / scoring improvements: Scorers gain declarative eligibility filters and deterministic sampling tied to trace decisions; PostgresStoreVNext now shows in-progress traces in Studio; span.endTree() can force-close an entire span tree → easier debugging for long-running agents with high trace volume
• Streaming / session UX: @mastra/ai-sdk adds withSseHeartbeat() so SSE connections survive long reasoning without being cut off by an intermediate proxy's timeout; AgentController#generateThreadTitle() can generate a thread title without initializing the full session → a good fit for UIs that list titles first and only load the full session on demand

Breaking Changes

• Mastra Code LSP is now opt-in — you must explicitly set "lsp": true in config
  • Affects: projects using Mastra Code editor integration features
• MarkdownRenderer no longer streams text word-by-word; it now uses useRevealedText
  • Affects: projects embedding the @mastra/playground-ui markdown rendering component directly
• CSS class mastra-markdown-arriving renamed to mastra-arriving
  • Affects: projects with custom styles relying on that class name
• SankeySignals now requires controlled selectedFrameId and onFrameIdChange props
  • Affects: projects using the playground-ui Sankey visualization component
• Background task storage no longer supports Cloudflare KV or ClickHouse
  • Affects: projects running background tasks on either backend
• DaytonaSandbox command execution result format changed: the args field is removed, and command now contains the full string
  • Affects: code that parses the Daytona sandbox command output structure directly
• agent.stream() removes the persistPartialOnAbort option; aborted runs now retain history automatically
  • Affects: code that previously set this option manually to control abort behavior

Migration Guide

Upgrading from 1.61.x to 1.62.0

# Step 1: update the dependency
pnpm add @mastra/core@1.62.0

// Step 2a: if you use Mastra Code LSP, enable it explicitly
export default defineConfig({
  lsp: true, // opt-in as of 1.62.0, no longer enabled by default
})

// Step 2b: update how you parse DaytonaSandbox command results
// Old (1.61.x)
const { command, args } = result

// New (1.62.0) — args is removed, command is now the full string
const { command } = result

If your project stores background tasks in Cloudflare KV or ClickHouse, switch to another supported storage backend before upgrading. If you use the persistPartialOnAbort option on agent.stream(), you can just delete it — abort behavior now defaults to retaining history, so nothing gets worse. The remaining UI-related breaking changes (LSP opt-in, MarkdownRenderer, the CSS class rename, SankeySignals) only affect projects embedding @mastra/playground-ui components directly; code that only calls the @mastra/core API is unaffected.

Cross-Framework Observations

For computer use, Mastra layers a SandboxComputer capability on top of its existing sandbox provider abstraction (Daytona, E2B Desktop) — a different layer than Anthropic's own Computer Use API or OpenAI's Operator, where the model itself is trained to operate a screen directly. Mastra instead turns desktop control into a 12th workspace tool at the agent-framework layer. The benefit: the same agent definition can move between environments with and without desktop capability just by swapping the sandbox provider, while governance mechanisms — approval, tracing, screenshot-after-action — carry over from the existing workspace tool framework instead of needing a separate rule set for computer use. Combined with the new Elasticsearch/Valkey storage in this release, Mastra is closing gaps faster than same-language competitors in its position as a "production-deployable TypeScript agent framework" — but that also means every upgrade now demands a careful read of the breaking changes list, and seven in one release is not a small number.

Takeaway

I used to think of "computer use" as purely a model-layer capability — feed in a screenshot, get back coordinates, and the model learns to operate a screen on its own. Seeing Mastra implement it as a workspace-tool-layer capability instead made me realize the framework layer can reuse the same approval/tracing/screenshot-after-action machinery to govern computer use alongside other workspace tools like read_file or shell commands — security controls don't need to be rewritten for every new tool type. It's a good example of decoupling a new capability from an existing governance framework.

References

• Mastra @mastra/core@1.62.0 — GitHub Release
• mastra-ai/mastra — GitHub
• Mastra Sandbox Overview — Computer-use tools
• Mastra @mastra/core@1.61.0 — GitHub Release (previous version)