ClawQL
Equal-arm latency · p50 / p95 / p99
Apples-to-apples: both sides return the same tiny pets JSON via one MCP execute.
Headline compares ClawQL gateway overhead (execute − same-host mock fetch) to
Executor (live same-host).
n=100.
ClawQL — equal arm
EQUAL — single MCP execute → tiny same-host mock (2 pets), fields projection only. No search bundle, no where, no WORM, no audit in the timed execute sample.
Executor — equal arm
EQUAL — single MCP execute returning the same pets JSON in-process (Executor sandbox cannot fetch; logical result matched).
ClawQL execute (e2e) equal
MCP → tiny same-host mock + fields
p50 1.8
p95 2.7
p99 4.2
ms
p50
p95
p99
ClawQL gateway overhead equalized
execute − direct fetch (apples-to-apples)
p50 1.3
p95 2.1
p99 3.4
ms
p50
p95
p99
Executor execute equal
live · same pets JSON in-process
p50 3.0
p95 5.5
p99 5.9
ms
p50
p95
p99
Direct HTTP mock control
bare fetch of the same tiny body
p50 0.4
p95 0.5
p99 0.8
ms
p50
p95
p99
ClawQL audit control
local ring append (no upstream)
p50 0.6
p95 14.1
p99 31.2
ms
p50
p95
p99
Equalized ClawQL p50 1.3ms
Equalized ClawQL p95 2.1ms
Equalized ClawQL p99 3.4ms
Executor p50 3.0mslive same-host
Executor p99 5.9ms
Compare 2.26× — ClawQL overhead faster
Honesty — Equalized ClawQL gateway overhead p50 1.341ms beats live Executor p50 3.032ms on the same result shape.
- Both arms return the same tiny pets JSON via one MCP execute. ClawQL goes through same-host mock HTTP; that cost is subtracted for the equalized overhead metric.
- Executor is LIVE equal-arm MCP execute returning the same pets JSON.
- Raw ClawQL execute e2e p50 1.8ms still includes same-host mock HTTP; equalized row subtracts it.
- Do not use UsefulSoftwareCo/executor#1519 50–100ms as measured — wire EXECUTOR_BIN for live.
- p99 needs adequate n; default iters=100. Still sensitive to rare GC/scheduling spikes.
Measured 2026-10-11T00:31:37.146Z (linux/x64, Node v22.14.0). Equal arm: MCP execute returns the same pets JSON ClawQL gets from the mock. In-process (Executor has no fetch); ClawQL pays same-host mock HTTP (~subtracted in overhead). Distinct from token compare.
latency.json ·
token benchmark notes ·
GitHub ·
Reproduce: EXECUTOR_BIN=… npm run benchmark:executor-comparison:latency