To MCP or not to MCP. There's a debate right now about whether MCP is the future of AI integration or whether code gen makes it unnecessary. I think the debate is framing it wrong.
When Anthropic dropped the MCP spec, the ecosystem moved fast. Within weeks, everyone had an MCP server. Slack. Notion. Salesforce. GitHub. Linear. The demo videos were genuinely impressive: AI agents browsing your tools, pulling context, taking action. It felt like the integration problem had been solved.
Then you tried to use them in production.
Most MCPs turned out to be shallow wrappers. A get_issues call here. A create_task call there. Enough to check the "MCP supported" box, not enough to handle anything real. The moment you pushed into genuinely complex workflows, cross-tool, multi-step, data-dependent, you hit a wall. MCP gave us the plumbing. But most implementations forgot to build the house.
Meanwhile, code gen got good. Really good.
An LLM can now read your API spec, understand every endpoint, every parameter, every response shape, and generate the integration code at agent build time. Not at runtime. At build time. The resulting code is deterministic, testable, and version-controlled, just like any other code you'd ship.
Your product already has an API spec, which is a universal standard. An LLM reads it, generates the connector, and from that point forward it runs like any other code you'd validate and ship. No server to build. No hand-crafted tool list to maintain. No lag between your product shipping a new capability and an agent being able to use it.
The spec is the connector.
So does code gen kill MCP? I don't think that's the right question.
MCP and code gen from API specs aren't competing. They're solving different dimensions of the same problem: how should AI systems orchestrate across other systems at scale? Both are legitimate answers and will coexist. The right tool for the right scenario.
What actually dies is the old way of connecting systems: hand-built, opinionated connectors.
That's the real story underneath the MCP debate. The connector layer is being commoditized. The real differentiation is moving to orchestration and workflow intelligence, and that's what Agent Studios are for.
We're building exactly this at Typeface with BYOAPI (Bring Your Own API) for marketing workflows. Point it at any API spec and your AI content workflows can talk to any internal or external system, immediately. A customer recently pointed us at their internal API spec and had a working content workflow running in hours, not weeks. No connectors. No waiting. (We support MCP too, for the record.)
I'm a fan of MCP. The intent is right, and the protocol will keep evolving. Code gen capabilities will keep improving. The two paths are likely to converge. That's this lap. The next one could look completely different.