Mastra released 1.10.0 today. This version is quite informative, but the most worth grasping is actually three points: the tool definition now supports 'inputExamples', the MCP client's custom fetch can receive 'RequestContext', and a number of stability fixes related to proxy execution, streaming output, and memory cleaning. Taken together, it is not a single feature update, but a patch for the most problematic layers of the proxy framework.
The 'inputExamples' change is very representative. It essentially tells the model "what kind of tool input is correct", which directly affects the hit rate of tool calls, especially in complex agent workflows, which is more stable than simply changing the prompt. On the other hand, MCP requests support continuing to carry cookies or bearer tokens in the context of the request, making the tool links that require authentication more like real enterprise usage rather than lab demos.
More importantly, Mastra didn't spend all its energy on new features. It also fixes provider stream error swallowing, vector residue after memory deletion, Cloudflare Workers build failure, file attachment routing exceptions, and sub-agent loop boundaries. For teams that have already integrated the agent framework into their production process, these fixes are more tangible than just "one more feature" on the surface.
FAQs
Q: What are the core additions in 1.10.0?
A: Tool definitions support 'inputExamples', which makes it easier for models to make the correct tool calls.
Q: Why is RequestContext forwarding important?
A: It allows MCP tool requests to be authenticated in context, which is suitable for real business system access.
Q: Why can't this update be seen as just a feature upgrade?
A: Because it also fixes a number of underlying issues that affect the stability of the agent.
Q: Which teams will be paying special attention to this update?
A: Do the development team for agent frameworks, tool calls, and enterprise AI workflows.
Q: What changes does this message reflect?
A: Mastra began to put "whether the tool can be calibrated and whether the agent can run steadily" to a position that is ahead of the heap function.
