SEP-2106 made the full 2020-12 vocabulary legal in an MCP tool's
inputSchema โ oneOf,
if/then, prefixItems,
$ref and more. Do models and providers actually
honor those keywords when they emit a tool call?
Measured across 24 models and three API arms, with a
Hugging Face-ready fidelity dataset โ and twin live, keyless demo
MCP servers you can point your own client (and model) at.
(An interactive playground is coming soon.)
Research on MCP, Agent Skills & adjacent topics
A place to gather research by Ola Hungerford and collaborators. Each topic is a self-contained write-up; the styled versions are published here, and the same content renders as Markdown on GitHub.