MCP just went stateless and nobody's talking about what this means for indie devs
I've been building on MCP since it was a weird spec with maybe 200 stars on GitHub. Yesterday it crossed a threshold that's going to quietly reshape how a lot of us ship AI products, and almost nobody outside the maintainer circle has connected the dots yet.
Here's the short version: the 2026-07-28 spec release just made MCP stateless.
If you've ever tried to run an MCP server at any real scale, you know why that matters. The original protocol was bidirectional and stateful, which is great for a local dev loop and genuinely painful for anyone trying to serve it behind a load balancer. You needed sticky sessions, you needed to keep connection state alive, and scaling horizontally meant fighting the protocol instead of your actual product. That's the exact complaint that's been all over MCP server threads for the past year.
Now it's request/response. Stateless. The kind of architecture you can put behind any standard infra without inventing workarounds.
A few other things buried in this release that indie hackers building on MCP should care about:
The protocol now has a real extensions framework. MCP Apps, Tasks, Enterprise Managed Auth ship as domain-specific extensions instead of getting jammed into the core spec. That means there's now an official lane for experimental additions, which is exactly the kind of surface area a small team can build a niche tool around before the big players notice.
Multi round-trip requests are formalized. Header-based routing is in. List results are cacheable now, which sounds boring until you realize it's a direct cost and latency win for anyone running an MCP server at volume.
Auth got hardened. If you've been nervous about shipping an MCP integration that touches real user data, this is the release where that got noticeably less scary.
TypeScript and Python SDKs both hit v2.0.0 for the new spec with better ergonomics, and a C# SDK is coming. If you've been holding off on MCP because the tooling felt early, this is the moment it stopped feeling early.
The part that actually surprised me: the growth numbers behind this. MCP is closing in on half a billion downloads a month, with both Tier 1 SDKs having crossed 1 billion downloads total in just a few months. That's not a niche protocol anymore. That's infrastructure.
Why this matters if you're indie hacking right now: the barrier to building and hosting an MCP server just dropped. Stateless means cheap, boring, standard hosting. A real extensions framework means there's room to build something opinionated without waiting on the core spec to catch up. And the download curve tells you the audience is already there, you're not evangelizing a new protocol to people, you're arriving early to one that's already going vertical.
If you've been sitting on an MCP-based idea waiting for the protocol to feel stable enough to build a business on, this release is probably the signal you were waiting for.
Anyone else digging into the 2026-07-28 spec yet? Curious what people are planning to build now that hosting got this much simpler.
Worth being precise about what "stateless" means here: the transport got stateless, not your product. Someone still has to remember which tool calls happened, in what order, for which user session, once you're doing anything beyond a single request/response. Now that the protocol stops holding that for you, it's your app's job, which is exactly the audit/session-state problem people already get wrong in stateful backends. Boring hosting is a real win, but it moves the hard part one layer up, it doesn't remove it.
One thing that often gets overlooked is that infrastructure becomes most interesting right around the point people stop talking about the infrastructure itself.
When the constraints become predictable, builders can spend more time creating differentiated products instead of engineering around the platform.
The stateless shift is massive for hosting complexity. Going from sticky sessions to standard request/response unlocks so much. The real hidden win is that you're no longer fighting infra to scale - you can use boring, commodity hosting (Lambda, Cloud Run, standard containers) instead of needing specialized stateful setups. Half a billion downloads/month is the kind of adoption curve that signals infrastructure, not trend. If you've been hesitant about MCP because it felt fragile at scale, this release removes that excuse.