Artificial intelligence news.
Straight from the original announcements.

Mastra adds MCP 2.0 support, letting servers scale like normal HTTP services

October 21, 2026By the AI News Bites Editorial TeamSource: Mastra
Mastra adds MCP 2.0 support, letting servers scale like normal HTTP services
Image: Mastra

Mastra now supports the MCP 2026-07-28 specification through @mastra/mcp ^2.0.0. The new spec removes server sessions, so every request is self-contained and MCP servers built with Mastra can scale like any other HTTP service.

Key points
  • Mastra supports the MCP 2026-07-28 spec via @mastra/mcp ^2.0.0.
  • The spec removes the initialize handshake and protocol-level sessions, so any instance can answer any request.
  • Servers can run behind round-robin load balancers or scale to zero on serverless platforms.
  • @mastra/mcp 2.0 removed SSE transport, roots, sampling and Dynamic Client Registration, and accepts only 2026-07-28 clients.
  • Requires @mastra/[email protected] and @mastra/[email protected] or later.

Why sessions were a problem in MCP 1.0

Under the previous MCP spec, a client opened every connection with an initialize request. The server created a session, returned a session ID, and every later request had to carry it. According to Mastra, this tied the client to the specific server instance that issued the session.

Behind a load balancer, that could cause trouble: a request might land on an instance that did not recognize the session ID, producing a 400 session not found error. The workarounds were sticky sessions or a shared session store such as Redis. Mastra says both worked, but they added complexity, latency and potentially cost.

What changes in the 2026-07-28 spec

Two proposals drive the change. SEP-2575 removes the initialize handshake, and SEP-2567 removes the Mcp-Session-Id header and protocol-level sessions. Each request is now independent and behaves like a regular HTTP request.

Any instance can answer any request, so a traditional round-robin load balancer works without sticky sessions or a session store. If an instance crashes, the next request goes to another instance and the client carries on. With no session to keep alive, a server also does not need to stay up between requests, so on serverless platforms it can scale to zero when idle, which Mastra says saves money.

Other additions described in the post:

  • HTTP headers: every request carries an Mcp-Method header, plus an Mcp-Name header for tool calls, prompts and resources. Gateways, rate limiters and firewalls can decide on a request without parsing its body.
  • Cache hints: list results such as tools/list and resources/list include ttlMs, how long the result stays fresh, and cacheScope, which marks it as public or private to the caller. In Mastra these are set through cacheHints on an MCPServer.
  • server/discover: reports a server's supported spec versions, capabilities, name and description. In Mastra, getServerInfo() on an MCPClient returns the server's name, version and description.

Asking users questions and handling long tasks

The requestState mechanism lets a server ask the user a question mid-call without holding the connection open. In Mastra, a tool calls suspend() from its execute function to pose the question and reads the answer from resumeData. The response contains the question and a requestState, a serialized string of the paused call. Deserialized, it records the tool, a fingerprint of its arguments, the question and who asked.

Tasks, previously experimental, are now an official extension. A long-running tool call can return a task handle straight away and keep working in the background while the client polls with tasks/get. Mastra says it is still investigating how to support tasks in @mastra/mcp. Meanwhile, long-running tools can report progress with context.mcp.progress().

Deprecations and what developers need to do

The spec deprecates the HTTP+SSE transport, roots, sampling, logging and OAuth Dynamic Client Registration. According to Mastra, @mastra/mcp 2.0 has already removed the SSE transport, roots, sampling and Dynamic Client Registration, and it only accepts clients on the 2026-07-28 spec.

To upgrade, developers install the latest @mastra/core and @mastra/mcp. The post notes that @mastra/[email protected] and @mastra/[email protected] or later are required. Mastra points to a migration guide for moving from v1 to v2.

For teams running MCP servers in production, the practical effect is simpler infrastructure: no sticky sessions or session store, and the option of serverless hosting that scales to zero. The trade-off is that 2.0 only works with clients on the new spec.

MastraMCP 2.0SEP-2575SEP-2567AI agentscloudenterprise AI

Related news

AWS details four ways to speed up Amazon Bedrock AgentCore sessions
For developersNovember 13, 2026, 23:02

AWS details four ways to speed up Amazon Bedrock AgentCore sessions

AWS has outlined four architecture patterns for reducing session startup delays in Amazon Bedrock AgentCore runtime, including snapshot-based initialization, build hooks…

Dynatrace and Microsoft bring observability context to Azure AI workflows
Products and toolsNovember 13, 2026, 18:36

Dynatrace and Microsoft bring observability context to Azure AI workflows

Dynatrace will showcase integrations with Microsoft tools that bring observability data into AI-assisted development, Azure AI applications and cloud operations. The…

Mastra Adds Private Model Gateways for Self-Hosted AI Agents
Products and toolsNovember 12, 2026

Mastra Adds Private Model Gateways for Self-Hosted AI Agents

Mastra has introduced private model gateways that let agents route requests to self-hosted open-weight models through a custom gateway. The feature can connect Mastra…