Performance Benchmarks

MCP servers are often the latency bottleneck in agentic AI loops. Slow parameter parsing, heavy reflection, unoptimized middleware, and un-cached repetitive tool calls add hundreds of milliseconds to model latency.

mcponce is engineered with extreme attention to performance. Below are the official benchmark results measured with Vitest and Tinybench on Apple Silicon (M-series).

Benchmark Results Summary #

Benchmark Suite	Operations / sec	Average Latency	Description
FIFO Concurrency Queue	~1,850,000 ops/sec	0.0005 ms	Enqueue and dequeue throughput across synchronized workers.
Smart Input Coercion	~810,000 ops/sec	0.0012 ms	Pre-compiled schema validation and type casting per invocation.
LRU Response Cache Hits	~5,500 ops/sec	0.18 ms	Full end-to-end tool execution cycle on cached responses.
Tool Registry Lookup	~2,400,000 ops/sec	0.0004 ms	Tool resolution, metadata lookup, and validation initialization.

Running Benchmarks Locally #

You can run the full benchmark suite on your own hardware at any time:

BASH

npm run bench

Sample output:

✓ bench/mcponce.bench.ts (4) 2854ms
  ✓ FIFO Concurrency Queue > 1,847,210 ops/sec ±0.42% (98 samples)
  ✓ Smart Input Coercion > 812,450 ops/sec ±0.31% (95 samples)
  ✓ In-Memory Cache Hits > 5,520 ops/sec ±0.85% (88 samples)
  ✓ Tool Registry Map Lookup > 2,415,800 ops/sec ±0.20% (99 samples)

BENCHMARK COMPLETE: All performance thresholds passed.

Architectural Optimizations #

How does mcponce achieve these speeds?

Pre-Compiled Schemas: Shorthand schemas are compiled into optimized Zod validators once at server initialization, avoiding runtime AST recreation.
Deterministic String Hashing: Response cache keys use a fast key-sorting algorithm that incurs zero garbage collector pressure.
Microtask Queue Scheduling: The concurrency queue utilizes native microtasks (queueMicrotask / resolved promises) rather than heavy timer primitives.
Zero-Allocation Middleware: Middleware chains avoid array slice allocations on each invocation by using linked index stepping.