Skip to content

Performance

Jump to bottom
github-actions[bot] edited this page Mar 3, 2026 · 1 revision

Performance

Baseline performance of ZeroMCP endpoints and how to reproduce numbers.

Running the benchmarks

From the repository root, run the BenchmarkDotNet project in Release:

dotnet run -c Release --project ZeroMCP.Benchmarks/ZeroMCP.Benchmarks.csproj

To run only the MCP endpoint benchmarks:

dotnet run -c Release --project ZeroMCP.Benchmarks/ZeroMCP.Benchmarks.csproj -- --filter "*McpEndpoint*"

The benchmarks use WebApplicationFactory with the ZeroMCP.Sample app (multiple controllers and minimal APIs, ~15 tools). Each benchmark measures the full HTTP round-trip.

What is benchmarked

Benchmark Description
GET /mcp/tools (inspector) Time and allocation for the tool inspector JSON response
POST /mcp tools/list JSON-RPC tools/list (discovery + schema serialization)
POST /mcp tools/call list_orders Full dispatch for list_orders (no arguments)
POST /mcp tools/call get_order Full dispatch for get_order with one route argument

Baseline numbers (reference)

Run the benchmarks on your machine to get local results. Typical orders of magnitude on a development machine (Release, .NET 10, Sample app with ~15 tools):

  • GET /mcp/tools — Usually in the low milliseconds (single-digit ms mean); allocation depends on tool count and schema size.
  • POST tools/list — Similar to inspector (same discovery path, different response shape).
  • POST tools/call — Depends on the action (e.g. in-memory list vs. no-op); often sub-ms to a few ms for simple actions.

Exact numbers depend on hardware, OS, and load. Use the benchmark project to establish baselines after changes (e.g. schema generation, dispatch, or serialization).

Reproducibility

  • Use Release configuration.
  • Close other heavy applications to reduce variance.
  • For comparison across commits, run the same filter and note the reported Mean and Error (and Allocated if using MemoryDiagnoser).

See also

Clone this wiki locally