npx -y -p mantis-browser mantis install
claude mcp add mantis -- npx -y -p mantis-browser mantis-mcp
npx -y -p mantis-browser mantis read https://example.com   # try it without an agent

Windows, Linux servers and other agents: SETUP.

Side-by-side replay of the same task with Mantis and Playwright MCP: Mantis finishes in 10 calls and about 292k input tokens, Playwright MCP in 22 calls and about 733k

One real task (Luna, logged runs replayed): about a third less time, over 2x fewer input tokens, about half the calls. This is a selected example from 16 long tasks; medians across all 16 are in Results. Full video.

Security and threat model: SECURITY.md.

How it works

Agent sends open/click to Mantis; Mantis reads the page in Chromium, numbers each control and returns only what changed; CAPTCHA, 2FA and payment go to you

What the agent sees

A shop page with a cookie popup becomes 8 numbered lines; after click #3 only the changes come back

Results

Same tasks, same pass criteria, three ways to browse: Mantis, the accessibility tree (Playwright MCP) and screenshots. Models: gpt-6-luna and claude-haiku-5-5. Full tables, method and caveats: bench/RESULTS.md.

Long local workflows (long-local-v1: 16 fixture tasks, 3 runs each, code-graded). Correct out of 48, Mantis / tree / screenshots: Luna 48 / 48 / 42, Haiku 48 / 48 / 48. Mantis uses 22-29% less time and 35-38% fewer input tokens than the tree.

Long real-site workflows (long-live-v1: 16 tasks on public sites, 48 runs per cell, blind judge). Medians over all runs, failures included. Mantis vs tree:

	Luna Mantis	Luna tree	Haiku Mantis	Haiku tree
Correct	41/48	37/48	47/48	44/48
Time	50s	75s	90s	135s
Input tokens	262k	362k	432k	746k
Tool calls	18	25	26	32

Short real-site tasks (live-v1 v1.8: 30 tasks, 90 runs per cell, blind judge or validators):

	Luna Mantis	Luna tree	Luna shots	Haiku Mantis	Haiku tree	Haiku shots
Success /90	89	90	86	89	89	90
Tool calls (median)	3	4	6	4	7	7.5
Input tokens (mean)	87k	122k	122k	72k	142k	75k
Tool output (median chars)	9.2k	54.9k	0.1k	9.5k	27.0k	0.1k
Time (median)	44s	23s	28s	45s	43s	33s

Haiku tree and screenshot runs were redone on the same CLI build (2.1.161, local) as the Mantis re-run, so tokens are comparable; their times come from a 16-way parallel run and may be inflated. The screenshot baseline is our own minimal tool, not a product's.

Reading it:

Long tasks: about a third less time, 28-42% fewer input tokens, 19-28% fewer calls, with accuracy equal or better.
Short tasks: accuracy is within one task of the tree on both models (Luna 89 vs 90, Haiku 89 vs 89; Luna screenshots trail at 86, Haiku screenshots lead at 90); Mantis returns about 3-6x less page text than the tree and uses 28-49% fewer input tokens (Haiku screenshots are close, 75k vs 72k), but is not faster: on Haiku it is about even with the tree (45s vs 43s) and slower than screenshots (33s).
Task swap: v1.8 replaced data-gov-climate with EIA About after scoring. It removed 7 Haiku failures (Mantis 2, tree 3, screenshots 2) and none for Luna, so Haiku was 87 / 86 / 88 in v1.7 and is 89 / 89 / 90 now; Mantis no longer leads the tree on Haiku.
The judge is not calibrated: 79% agreement and 6/48 false successes against deterministic keys, so treat model-judged rows as indicative (details).
Scope: read/navigate tasks on public sites, two models; live-v1 swapped one task (data-gov-climate to eia-about) in v1.8. The long-live-v1 task set was adjusted before scoring; swaps are listed in bench/RESULTS.md and bench/live-v1.md.