Hound sends the diff, not the page

Every mainstream browser-automation MCP server returns the page's full accessibility tree on every action. Hound budgets each response and returns only what changed. Measured here, on three real pages, in an 8-step agent flow.

99.2%

fewer tokens

311,967

conventional total

2,625

hound total

pages measured

Measured 28 October 2026 · budget 1,200 tokens/call · 8-step flow · single run per page.

Results

Page	Conventional	Hound	Saved	Reduction
example.com	8,781	845	7,936	90.4%
news.ycombinator.com	93,485	806	92,679	99.1%
en.wikipedia.org/wiki/Web_scraping	209,701	974	208,727	99.5%
Total	311,967	2,625	309,342	99.2%

The gap widens with page size: 91% on a near-empty page, 99.5% on a long Wikipedia article. That is the point — cost becomes a function of change, not page size.

Where the saving comes from

Per-step detail for news.ycombinator.com, the same flow with both strategies:

Step	Conventional	Hound	Ratio
open	11,683	570	20.5×
snapshot	11,683	29	402.9×
act:click	11,683	29	402.9×
snapshot	11,683	29	402.9×
snapshot	11,683	29	402.9×
act:fill	11,690	62	188.5×
snapshot	11,690	29	403.1×
snapshot	11,690	29	403.1×
Total	93,485	806	116×

Conventional pays the full tree eight times. Hound pays once on open, then a 29-token "nothing changed" delta per follow-up.

Method

Conventional = what a snapshot-per-action server returns after each action: the full accessibility snapshot plus up to 40,000 characters of page text, with no memory of what the agent already saw.
Hound = the budgeted response for the same step, computed as a delta against the previous state (1,200-token budget, minimal profile for snapshots).
Both are measured on the same page in the same browser session, after the same action — never against a stale document.
If a page has no element matching a step (e.g. no search box), that step is dropped from both sides rather than crediting Hound for work it didn't do.
Tokens are estimated with Hound's own estimator, the same one it budgets against.

Honest caveats. This is a single run per page, not an average; page content changes over time, so absolute numbers drift. The estimator is a model, not a real tokenizer. And the comparison is against a snapshot-per-action server — a server that already sends diffs would narrow this gap. The claim is narrow on purpose: Hound removes the repeated full-tree cost.

Reproduce it

Two commands, no account, no API key:

pip install hound-browser-mcp
playwright install chromium

hound-bench                      # default pages
hound-bench --url https://your-page.example --json out.json
hound-bench --pages 3 --budget 1200 --profile balanced

Or from source: pip install -e ".[dev]" && python -m bench.run --url https://news.ycombinator.com

Raw output for the run above is committed at bench/results/2026-10-28.json. If your numbers differ, your page changed or your budget did — that is the benchmark working as intended, not breaking.