ZeroMCP v0.1.0 is here. v0.2.0 coming soon.
10 Languages HTTP Transport Open Source Harness

How fast is your
MCP server?

MCP over stdio is a demo. Production means HTTP. ZeroMCP gives you a handler function — plug it into whatever framework your team already runs.

See the numbers Methodology
HTTP Performance

ZeroMCP vs Official SDKs. Same tool. Same transport.

Both servers run the same hello tool in Docker containers over HTTP. 5 minutes sustained load. Same benchmark client, same methodology.

Rust 2.1x
ZeroMCP (Actix) 5,111 req/s 0.17% CPU · 4 MB
vs.
Official SDK 2,452 req/s 0.03% CPU · 2,420 MB*
* rmcp leaks memory: 126 MB → 2.4 GB over 5 minutes.
Full comparison →
Node.js 1.7x
ZeroMCP (bare http) 4,539 req/s 0.02% CPU · 26 MB
vs.
Official SDK 2,610 req/s 0.02% CPU · 174 MB
Full comparison →
C# 1.8x
ZeroMCP (ASP.NET) 4,421 req/s 1.47% CPU · 314 MB
vs.
Official SDK 2,517 req/s 0.05% CPU · 78 MB
Full comparison →
Go 4.5x
ZeroMCP (Chi) 4,024 req/s 0.00% CPU · 23 MB
vs.
Official SDK 893 req/s 0.03% CPU · 44 MB
Full comparison →
Java 1.4x
ZeroMCP (Javalin) 3,791 req/s 0.16% CPU · 207 MB
vs.
Official SDK 2,658 req/s 0.15% CPU · 217 MB
Full comparison →
Ruby 1.5x
ZeroMCP (Rack+Puma) 3,217 req/s 0.06% CPU · 26 MB
vs.
Official SDK 2,163 req/s 0.04% CPU · 56 MB
Full comparison →
Kotlin 5.2x
ZeroMCP (Ktor) 2,848 req/s 0.11% CPU · 194 MB
vs.
Official SDK 548 req/s 0.19% CPU · 204 MB
Full comparison →
Python 4.1x
ZeroMCP (Starlette) 2,623 req/s 0.28% CPU · 27 MB
vs.
Official SDK 635 req/s 0.04% CPU · 87 MB
Full comparison →
Swift 23x
ZeroMCP (Vapor) 1,730 req/s 0.21% CPU · 96 MB
vs.
Official SDK 74 req/s 1.76% CPU · 39 MB
Full comparison →
PHP 92x
ZeroMCP (Slim) 1,561 req/s 0.00% CPU · 33 MB
vs.
Official SDK 17 req/s 0.36% CPU · 64 MB
Full comparison →
Resilience

What breaks under pressure

"Every AI coding session spawns a tree of child processes. When the session ends, these processes are supposed to terminate. They don't." We sent 21 attack scenarios at 3 official MCP SDKs. Malformed JSON, binary garbage, memory leaks, hanging tools, rapid-fire floods.

C#
20/21
1 degraded (tool hang timeout)
PHP
19/21
2 corrupted (giant string, slow tool)
Ruby
19/21
1 crashed (binary garbage), 1 corrupted
Full attack matrix with 21 scenarios →
SDK Comparisons

Official SDK vs ZeroMCP

Side-by-side for each language. Code examples, architecture, what you gain and what you lose.

Node.js Python Go Rust Java Kotlin Swift C# Ruby PHP
Details

Methodology

Don't trust our numbers. Run them yourself. Docker containers, pinned versions, open-source harness.

Transport

JSON-RPC 2.0 over HTTP POST. The same protocol MCP clients use in production.

Environment

Docker containers (Ubuntu 24.04), isolated CPU/memory allocation.

Duration

300 seconds sustained load. Snapshots every 10 seconds. 50 warmup requests discarded.

Metrics

Throughput (req/s), latency (p50/p95/p99), memory (VmRSS), CPU (utime + stime).

Chaos

21 attacks per language. Malformed input, binary garbage, memory leaks, hanging tools, rapid-fire.

Frameworks

ZeroMCP embedded as a handler function. No proxy. No sidecar. Same process as your server.

Run it yourself on GitHub →