We’re attempting something ambitious: a universal bridge between AI agents and every API.
Not a connector for specific services. Not a framework that requires integration work per API. A single system that can discover, understand, and execute any API operation—dynamically, at runtime, without prior configuration.
This is harder than it sounds. Much harder.
The shape of the problem
AI agents need to call APIs. This seems straightforward until you try to make it work at scale. The challenges compound:
Discovery. How does an agent find the right operation among thousands? You can’t front-load every API into the context window. You need retrieval—but retrieval of what? API documentation describes implementation, not intent. There’s nothing useful to embed.
Schemas. Every API vendor has their own dialect of JSON Schema. Every LLM provider has their own constraints on tool schemas. The space of all APIs is not a subset of what any single provider allows. Transformation isn’t enough.
Execution. OpenAPI’s encoding model is a combinatorial maze. Parameter styles, explode modes, content types, property encodings—they permute in ways the spec doesn’t fully define. You can’t special-case your way through it.
Credentials. AI needs to authenticate to APIs using your credentials. But those credentials are the keys to your digital kingdom. How do you give AI access without exposing secrets?
Each of these is a research problem. Together, they’re the reason universal AI-API integration doesn’t exist yet.
Why this blog
We’re going to write about how we’re solving these problems. Not marketing. Not hand-waving. The actual engineering.
We’re building this from first principles because there’s no other way. The existing approaches don’t scale. So we’re writing our own JSON Schema infrastructure, our own OpenAPI execution engine, our own semantic discovery system. It’s a lot of work. But it’s the only way to get something that actually works.
The journey
This blog will document that journey. The technical challenges. The dead ends. The insights that finally make things click. We’re writing it as we go, not retrospectively, because the process is part of the point.
If you’re interested in AI infrastructure, compiler engineering, or just building hard things from scratch—follow along.
We’re just getting started.