MCP Benchmark records repeatable raw observations for explicitly named public MCP endpoints. It runs the same bounded initialize and tools/list method for each target, preserves input order, fixes concurrency at two, and exports both JSON and CSV without inventing a composite score.
Each row includes its own timestamp, transport state, HTTP status, latency fields, tool count, truncation state, and full compatibility evidence. Methodology and limitations travel with the data so benchmark editions can be compared honestly.
What you can do
- benchmark_server — run one to five observations per explicit endpoint
- Benchmark up to five explicit targets and five total runs with fixed concurrency two
- Export ordered raw JSON and spreadsheet-ready CSV
- Preserve methodology, timestamps, caps, failures, and incomplete evidence
Who it's for
MCP maintainers, ecosystem researchers, release engineers, and technical writers producing transparent interoperability evidence. It is not a load test and does not rank servers with a hidden formula.
How to use it
- Call benchmark_server with only the public endpoints the user explicitly selected
- Choose the same run count and limits for every target in an edition
- Keep the returned methodology and limitations with the JSON or CSV
- Compare raw fields rather than deriving unsupported quality claims
Getting started
Start with one run per endpoint, then repeat only when the methodology requires multiple point-in-time observations.
