When architecting AI agents that execute multi-step planning loops, tool invocation latency is frequently dismissed as a rounding error compared to model token generation.
However, in autonomous engineering agents (like Cursor Agent or Claude Desktop executing 10 to 15 sequential queries to triage a codebase or inspect an infrastructure cluster), transport and serialization overhead compound rapidly.
We benchmarked 10,000 tool executions across the two primary Model Context Protocol (MCP) transport models: Stdio and Server-Sent Events (SSE).
1. The Numbers: Invocation Latency
| Metric | Stdio (UNIX Pipe / IPC) | Remote SSE (HTTP/1.1 + TLS) |
|---|---|---|
| Mean Latency | 2.1 ms | 19.4 ms |
| p95 Latency | 3.8 ms | 32.1 ms |
| p99 Latency | 6.2 ms | 48.7 ms |
| Connection Setup | 0 ms (Persistent Pipe) | 45 ms (TCP Handshake + TLS) |
2. Memory Footprint
- Node.js Stdio Worker: ~32MB RSS per active process.
- Python FastMCP Worker: ~21MB RSS per process.
- Compiled Go/Rust Worker: <7MB RSS per process.
- Centralized SSE Daemon: ~42MB shared across all incoming client streams.
3. Architecture Verdict
For developer workstations and desktop agents (Claude Desktop, Cursor), stdio is strictly superior: sub-3ms invocation, zero network port binding, and OS-supervised sandboxing.
For multi-tenant cloud environments where agents share access to a centralized cluster or database, SSE behind an Envoy or Traefik reverse proxy provides the necessary mTLS authentication and rate-limiting controls.
Explore our full benchmark suite, architecture comparisons, and verified native server recipes at MCP Bridge.
Top comments (0)