Methodology
Measured numbers for the GMAN proxy and gateway, with constraints surfaced first. Reproducible scripts in scripts/; result artifacts in benchmarks/.
Local GMAN proxy overhead
Measured local GMAN proxy overhead in a deterministic mock-provider environment. Not a universal SDK tax — the published delta is this proxy's hot-path cost on the recorded environment, on the date in the artifact below. Real-world deployments add upstream network RTT to whatever provider the proxy points at; that is not measured here. See full methodology for what is and is not measured.
- run_at
- 2026-05-16T17:10:50.337Z
- node
- v24.3.0
- os
- darwin-arm64
- cpu
- Apple M4 Pro
- sample size
- 200 per path
- warmup
- 20 per path (discarded)
- payload
- 76 bytes (same both paths)
- keep-alive
- enabled
| path | p50 (ms) | p95 (ms) | p99 (ms) | mean (ms) |
|---|---|---|---|---|
| direct | 0.118 | 0.183 | 0.273 | 0.13 |
| proxied | 0.334 | 0.548 | 1.086 | 0.36 |
| delta · proxy tax | 0.216 | 0.365 | 0.813 | 0.23 |
- · Measured against a deterministic local mock provider; real-world tax adds upstream network RTT to whichever provider you point the proxy at.
- · Telemetry POSTs to the mock gateway are fire-and-forget (proxy.ts:264) and do not block the caller's response. Their completion is not included in measured latency.
- · Keep-alive enabled on the shared http.Agent to mirror production connection reuse.
- · Same payload size both paths (see config.payload_bytes).
Gateway ingest envelope
Measured sustained ingestion on a single Railway replica with free-tier Supabase and a hard 100 req/min/IP gateway ceiling. The envelope is the headline; the numbers are valid only inside it. Rate-limit reality is preserved in the raw data — pauses and 429s are not smoothed. full methodology, including what this benchmark is NOT.
- run_at
- 2026-05-17T18:40:01.904Z
- gateway
- getmyagentnow-ui-production.up.railway.app
- replicas
- 1 Railway replica
- db tier
- free Supabase
- rate limit
- 100 req/min/IP
- queue layer
- none — synchronous insert
- region
- us-west
- bench config
- 5000 events · 5 in-flight
| metric | value |
|---|---|
| sustained accepted / min | 97 |
| events sent | 5000 |
| events accepted | 4643 |
| drop % | 7.14 |
| 429 rate-limit hits | 357 |
| 5xx server errors | 0 |
| duration (min) | 47.64 |
| latency p50 (ms, accepted) | 86.34 |
| latency p95 (ms, accepted) | 248.66 |
| latency p99 (ms, accepted) | 387.07 |
- · not a horizontally scaled deployment (single replica)
- · not multi-region (US-West only)
- · not durability-tested under node failure
- · not representative of enterprise throughput
- · not an end-to-end correctness test (synthetic payloads)