Benchmarks

cargo bench -p carmy-benches

The fixture tool does no work, so every number is Carmy’s own overhead on top of your code.

benchmark	measures
dispatch/direct_tool_call	baseline: calling the typed tool directly
dispatch/runtime_execute	the full runtime pipeline
json/result_dto	serializing an HTTP result
discovery/schema_generation	generating metadata and schemas
discovery/tool_catalog_json	serializing the catalog
http/execute_roundtrip	an in-process Axum request, runtime call and response
concurrent/64_read_only_executions	64 concurrent executions

One local run (Apple M3 Pro, rustc 1.93.0):

benchmark	median
dispatch/direct_tool_call	90 ns
dispatch/runtime_execute	2.5 µs
json/result_dto	0.31 µs
discovery/schema_generation	2.1 µs
discovery/tool_catalog_json	2.1 µs
http/execute_roundtrip	4.7 µs
concurrent/64_read_only_executions	98 µs

These numbers describe one machine; rerun the benchmarks before relying on them. Carmy makes no performance claims that these benchmarks cannot reproduce.