The hosted runtime thesis, in three questions

I've been asked variations of the same question every week for a month. It usually starts with: "isn't running an MCP server just a Docker container?" The short answer is yes. The longer answer is why we started Cognoverge.

There are three questions that decide whether the hosted-runtime thesis holds. If any of them flips, the whole bet weakens. I want to walk through where we've landed on each — and how you can argue with us if you think we're wrong.

1. Is MCP actually the standard, or a phase?

Every major AI client shipped MCP support in the last eighteen months — Claude Desktop first, then Cursor, then Windsurf, Cline, Continue, Zed. That's a strong signal. But the concerning version of the same story is that model vendors have historically used protocols as leverage: OpenAI shipped Functions, then Assistants, then GPT Store. Anthropic shipped MCP. The next contender ships something similar-but-incompatible, and the protocol wars begin.

Our bet: it doesn't matter. We built the runtime layer to be protocol-agnostic. If OpenAI ships "Apps" and it wins, we support Apps. If MCP wins, we're already there. The auth + observability + billing layer we sit above is reusable.

2. Do developers care enough about the ceremony to pay?

The pitch for a hosted runtime is that you skip two weeks of glue — HTTPS, OAuth, rate limits, logs, versioning, billing. But "skip two weeks of work" isn't a business by itself. You need those two weeks to have real leverage: many developers never finish them, so the servers stay on localhost and die in half-finished Dockerfiles.

That's what we see empirically. Most MCP servers are weekend projects. They never reach a second user. The ones that do reach a second user hit the OAuth wall — Claude and Cursor have subtly different token expectations — and often stop there. The pain is real.

3. Will Anthropic or Cloudflare just ship this themselves?

Cloudflare has already shipped Workers-based MCP hosting. Vercel could ship it any weekend they wanted. This is the real threat.

Our answer isn't "they won't." They might. Our answer is that the differentiation is above the compute layer — multi-client OAuth conformance, marketplace-style billing (Stripe Connect for tool authors), and per-server dashboards with the level of detail dev buyers actually pay for. Cloudflare wins compute. We compete on the product.

That's the thesis. Two years from now, if the runtime layer turned out to be commoditized and the pain wasn't there, we'll write the honest post-mortem. Until then, we're shipping.