Mastra's $22M buys a runtime, not a framework

Mastra, the TypeScript agent framework, has raised a $22M Series A led by Spark Capital, taking total capital raised to $35M ^[1]. That implies roughly $13M before this round ^[3], but the number to read is what shipped with it: Mastra also launched a platform, with Studio for evals, logs, traces and datasets, Server to deploy agents and workflows, and a Memory Gateway for agent memory ^[2].

That is the trade. The framework itself is open source, built on Vercel's AI SDK by much of the former Gatsby team ^[4], and licensed Elastic v2 specifically to stop a hyperscaler repackaging it ^[5]. Frameworks do not have a price. Sessions, sandboxes and stored traces do.

The convergence is now hard to miss. Vercel's eve is an open-source agent framework that ships durable execution, sandboxed compute, human-in-the-loop approvals, subagents and evals in the box ^[6]; deployed, its sandbox runs on Vercel Sandbox, with local adapters for Docker, microsandbox or just-bash ^[8]. Cloudflare's pitch is to build and host agents on a durable runtime, deploy once, and scale to tens of millions of instances with no infrastructure to manage ^[9], with a starter that includes human-in-the-loop approval and task scheduling and defaults to Workers AI without API keys ^[10]. All three now offer the same three things: durable sessions, sandboxed execution, and an approval gate ^[11].

Mastra's dependency is the awkward part. Its framework sits on Vercel's AI SDK ^[4], and Vercel now ships a competing framework that it runs its own agents on, v0 among them ^[7]. Mastra's hedges are visible in the product: Studio is offered self-hosted as well as cloud, and Memory Gateway is pitched as working whether or not you use Mastra ^[2].

What would move this: the harness thesis. Mastra argues that over 18 months of model improvement harnesses got bigger rather than smaller ^[12], and it has been adding primitives accordingly, including parallel tool calls, subagents, sandboxes, filesystems and skills ^[13]. If that inverts, the hosting attach rate goes with it. The other tell is where large deployments land. Mastra cites a Marsh McLennan enterprise search used by more than 100,000 people daily ^[15]; whether workloads that size sit on vendor runtime or self-hosted infrastructure decides what the $22M is actually funding.