Kitsuvo / Performance field notes / 02 Nov 2026

Browser performance,
with the receipts.

One Mac. The same pages.
Every result in the open.

A measured look at Kitsuvo, Chrome, Edge and Firefox: what felt quick in a controlled test, what took longer, and what we still need to measure.

Runtime results use a Kitsuvo 0.2.38 candidate, not the public download. Mac only. Local fixtures. No universal speed claim.

Same work. Four browsers.

Read a table pagePage work · milliseconds · lower is less time
Read a table page: median and p95Kitsuvo candidate: median 1 ms, p95 6 ms; Chrome: median 1.5 ms, p95 1.9 ms; Edge: median 1.2 ms, p95 1.6 ms; Firefox: median 4 ms, p95 8 ms. Thirty observations each. Zero is below the timer bucket.Kitsuvo candidate1 / 6Chrome1.5 / 1.9Edge1.2 / 1.6Firefox4 / 8Bar: medianMark: p95

Kitsuvo is an unreleased 0.2.38 candidate. Timer resolution differs across engines; a sub-millisecond score is not zero work.

01 / Measured work

Fast is a workload.
Not a trophy.

1,680 successful timed observations across four browsers. The same helper produced identical full-snapshot text in every browser. Here is the complete picture, including the slower tails.

Kitsuvo’s table read: 1 ms median, 6 ms p95. Chrome and Edge had tighter tails on that task. Firefox had the lowest median for table navigation and verification. There is no across-the-board winner.

These are short, warm, synthetic tasks. Real websites, installed extensions and model reasoning can change the experience completely.

Page helper execution · median / p95 milliseconds · 30 observations per cell
WorkloadKitsuvo candidate0.2.38 · unreleasedChrome156.0.8078.76Edge155.0.4283.45Firefox157.0.1
small · fullShared helper, in-page work<1 / 10.1 / 10.1 / 1.1<1 / 1
small · boundedShared helper, in-page work<1 / 1<1 / 0.1<1 / 0.2<1 / 1
small · viewportShared helper, in-page work<1 / <10.1 / 0.70.1 / 0.7<1 / 1
small · form-edit-and-readShared helper, in-page work<1 / 10.1 / 0.30.1 / 0.3<1 / 1
article · fullShared helper, in-page work<1 / 10.4 / 0.50.4 / 11 / 1
article · boundedShared helper, in-page work<1 / 1<1 / 0.1<1 / 0.1<1 / 1
article · viewportShared helper, in-page work2 / 33.1 / 4.13.3 / 4.57 / 8
table · fullShared helper, in-page work1 / 61.5 / 1.91.2 / 1.64 / 8
table · boundedShared helper, in-page work<1 / 10.1 / 0.20.1 / 0.2<1 / 1
table · viewportShared helper, in-page work9 / 11.013.6 / 17.113.4 / 16.430.0 / 35.0

Full snapshots are bounded to 30,000 characters, bounded snapshots to 2,000. “Viewport” reads only the visible region. Viewport output differs across engines because text and controls lay out differently; those rows do not represent identical returned text. Full and bounded snapshot text matches exactly. <1 means the timer returned zero, not zero execution cost. Sub-millisecond ratios would be misleading.

Loading the page is a different question.

Navigation below includes the request, each adapter’s readiness check and title verification. Kitsuvo uses its local control API, Chrome and Edge use CDP, and Firefox uses WebDriver. These round trips are not pure engine-speed scores or MCP product comparisons.

Local navigation + verification · median / p95 milliseconds
WorkloadKitsuvo candidate0.2.38 · unreleasedChrome156.0.8078.76Edge155.0.4283.45Firefox157.0.1
small page6.5 / 95.7 / 7.66.6 / 8.76.1 / 7.7
article page15.2 / 112.812.5 / 16.415.8 / 23.58.3 / 10.1
table page43.5 / 84.228.1 / 35.032.6 / 35.615.8 / 17.8

02 / Weight, memory, CPU, network

Count the cost.
Keep the context.

A small app directory is useful. It does not mean a browser uses the same amount of memory, CPU or battery.

App directory on this Mac

Allocated MiB, measured locally. Public Kitsuvo release for this table; the minimal candidate bundle is excluded.

Kitsuvo public 0.2.37
69.5 MiB
Chrome 156.0.8078.76
1454.6 MiB
Edge 155.0.4283.45
924.1 MiB
Firefox 157.0.1
510.0 MiB
Brave 155.1.97.56
572.5 MiB

Excludes profiles, local models, downloads and system/shared frameworks. Kitsuvo uses WebKit supplied by macOS; Chromium and Firefox package their engines. This is not total installation footprint. Safari’s system-managed app was excluded from this disk measurement.

What crossed the fixture server

Each run recorded its loopback response bodies: Kitsuvo 6.15 MiB, Chrome and Edge 6.40 MiB, Firefox 6.15 MiB.

Adapter priming and favicon probes differ. The synthetic server returns HTML for every path. These totals include warm-ups and exclude headers, protocol traffic and external/background connections.

No data-saving claim follows from these figures. Real-network loading, cache effectiveness, total egress and ad-blocking savings need separate tests.

Inspect the request counts ↗

Memory and CPU: useful diagnostics, incomplete totals.

Only the owned main process was sampled. Renderer, WebKit XPC, GPU and network helpers are excluded. Process architectures differ, so this table cannot rank total-browser efficiency. Values show round 1 / round 2; CPU is cumulative time used during that workload, not an idle percentage.

Main process only · no total-memory or CPU ranking
BrowserEnd RSSCPU timeObserved elapsed
Kitsuvo candidate185.7 / 182.8 MiB1.12 / 0.88 s3.2 / 2.3 s
Chrome230.1 / 232.3 MiB0.53 / 0.59 s1.3 / 1.4 s
Edge274.4 / 272.0 MiB0.76 / 0.76 s1.6 / 1.5 s
Firefox719.1 / 715.1 MiB0.58 / 0.56 s1.7 / 1.7 s

Not measured here: total process-tree memory, energy, battery life, GPU efficiency, scrolling frame times, cold startup and long-running real-world stability. Missing measurements are unknown, not zero.

03 / Reproducible by design

The test is part
of the result.

Same setup. Reversed order.

  1. One shared host. Apple M5 Max, 64 GiB RAM, arm64, Darwin 25.6.0. Other host activity was not fully controlled.
  2. Matched display state. Headless and hidden, 800 × 560 CSS pixels, DPR 2. Disposable isolated profiles, local pages, no account sessions or model inference.
  3. Same work and output. Small form, 1,800-paragraph article, 1,500-row table. The identical 42,743-byte helper replaced preinstalled helpers; snapshot hashes and edited values were checked.
  4. Repeated, not selected. Two sequential rounds in reversed order, 15 timed observations per workflow per round. Warm-ups excluded. All valid samples pooled; nearest-rank p50/p95.

Four measured browsers, three honest gaps.

  • Public Kitsuvo 0.2.37: the extracted app’s control endpoint did not become ready within 30 seconds. No runtime score is substituted.
  • Brave 155.1.97.56: the first attempt reached the harness deadline. A bounded retry failed with “CDP WebSocket open timed out.”
  • Safari 26.6.2: the existing adapter is headed only. We did not mix it into a hidden/headless comparison.

Windows, Linux and mobile results are pending. This audit does not measure model quality, safety, detector accuracy or a user’s complete browsing session.

Check our work.

Application source 35d5a4ca3a7e. Binary and helper hashes, raw samples, browser versions and resource scopes travel with the data.

Summary JSON ↗All timing rows (CSV) ↗Method & reproduction ↗Earlier MCP comparison ↗

04 / The useful next steps

Less waiting.
More finding.

We are building Kitsuvo for the person tired of AI-built search results. Browser control is useful, but the point is getting you to pages worth reading.

  1. Measure the whole browser. Attribute engine helpers before making memory or CPU comparisons.
  2. Work on the slow tail. Repeat table and viewport tasks under load; preserve correctness before chasing smaller medians.
  3. Take it outside the lab. Add real-network, cold-start, scrolling, battery and multi-platform runs before broadening any claim.
Kitsuvo

Less AI. Less slop.

Filter likely AI-built websites. See why. Open any page anyway.

Get Kitsuvo