๐ ๐๐ฃ ๐ฆ๐๐ฎ๐๐ฒ๐น๐ฒ๐๐ ๐ง๐ฟ๐ฎ๐ป๐๐ฝ๐ผ๐ฟ๐
Old MCP used sticky routing. One server held your session. If a server crashed, you lost your data. Load balancers pinned you to one spot. This made scaling hard.
The 2026-07-28 update changes this. Requests are now stateless. Every request carries all needed info. It includes auth and protocol versions.
Think of a post office.
- Old way: One clerk has a private notebook. You must wait for her.
- New way: Your info is on the envelope. Any clerk helps you.
This brings big wins:
- Easy horizontal scaling.
- Fast blue/green deploys.
- Quick crash recovery.
Some data still needs to stay. Use a shared store like Redis. The transport stays stateless. The store keeps the memory.
The Tasks extension (SEP-2663) also improves. You poll a task ID. Any server in the fleet handles the request.
Stop reading state from the connection. Put everything in the request. This makes your fleet robust.
Source: https://dev.to/pueding/mcp-2026-07-28-rc-stateless-transport-4ma9 Optional learning community: https://t.me/GyaanSetuAi