Repository navigation
Benchmark typed capability retrieval against VirtualMCPServer tool filtering #6742
Description
Activity
Thanks for the question, @JDeun, and for being upfront about your connection to SchemaRouter.
We have decided not to host a benchmark or example in this repo that centers a specific third-party project. We would rather not take on maintaining a comparison against an external tool, and we do not have a plan to add retrieval or preselection comparisons to the VirtualMCPServer docs.
If you would like to publish the comparison, your own repo is a good home for it, and you are welcome to link to it from there.
Thanks again for reaching out. Closing this one out.
ToolHive's VirtualMCPServer/registry layer is a strong host for a large-catalog retrieval question. I maintain SchemaRouter, which indexes MCP/OpenAPI/Python capabilities at provider/endpoint/field granularity and returns a bounded selection while the host keeps auth/policy/execution authority. Would an optional frozen benchmark/example be in scope comparing existing VirtualMCPServer filtering/exposure with typed field-aware preselection on the same catalog? Metrics: required-tool/field recall, unsupported rejection, exposed schema bytes, shortlist size, latency, plus behavior under tool-definition changes. Negative results would be reported unchanged. https://github.com/JDeun/SchemaRouter