Full percentile distribution
See median and tail behavior instead of trusting an average that hides slow users.
Performance assurance
Test the latency tail—not just the happy path.
Load-test MCP tools, HTTP endpoints, and AI workflows against explicit SLOs. Measure p50 through p99, throughput, errors, retries, and the percentage of work completed inside budget.
p50 / p95 / p99Local profilesCI gates
search-tool · 50 virtual users
$ prooflane perf --profile release requests 12,840 throughput 106.7 /s errors 0.18% p50 214 ms p95 681 ms p99 1.48 s under budget 96.2% FAIL · p99 exceeds 1.20 s SLO
Measure what matters
Connect load shape to user-visible reliability.
Sustainable work rate
Measure completed operations under controlled concurrency, ramp, duration, and pacing.
Categorized failures
Separate timeouts, transport errors, rate limits, invalid output, and application failures.
Percent inside SLO
Express the user promise directly and fail when too many tasks miss their budget.
Realistic inputs
Drive load from versioned MCP contracts and representative scenario distributions.
Baseline comparison
Detect meaningful latency and reliability movement between releases.
Workflow
Performance evidence that reproduces.
Model load
Define concurrency, duration, ramp, pacing, cases, and environment metadata.
Warm up
Separate startup effects and record the conditions that shape the run.
Measure
Collect timings, throughput, outcomes, retries, and classified failures.
Gate
Compare SLOs and baselines, then publish machine-readable CI evidence.
prooflane perf --profile release --json perf.json prooflane gate --performance perf.json --github
Part of the spine
Performance is one dimension of release readiness.
Protect the tail
Ship the speed your users were promised.
Make latency, throughput, and reliability part of the same deployment decision.
30-day full-access trial. No credit card required.