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

Mantis vs the accessibility tree (Playwright MCP) on public sites, gpt-6-luna and claude-haiku-5-5. Every cell is a per-run median; time is end-to-end wall time per run:

Long tasks (16 tasks, 48 runs per cell):

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

Short tasks (30 tasks, 90 runs per cell):

	Luna Mantis	Luna tree	Luna screenshot	Haiku Mantis	Haiku tree	Haiku screenshot
Correct	88/90	86/90	90/90	90/90	88/90	88/90
Input tokens (median)	63k	99k	104k	43k	88k	33k
Tool calls (median)	2	4	7	3	5	6
Time (median)	27s	31s	35s	24s	31s	23s

Long tasks: about a third less time and 28-42% fewer input tokens, accuracy equal or better. Short tasks (clean re-run, one build, all approaches interleaved): accuracy equal or better than the tree, 37-51% fewer input tokens and 13-22% less time than the tree. Against screenshots: Luna screenshots score 2 runs higher and use 67% more tokens and 29% more time; Haiku screenshots use about 22% fewer tokens and are about even on time, with 2x the tool calls.

Short-task scores are graded by deterministic answer keys on the captured final state. Long-task rows are model-judged by an uncalibrated judge (79% agreement with deterministic keys), so they are indicative; one short task was swapped after scoring. Screenshot baseline, local workflows, method and all caveats: bench/RESULTS.md.

Limitations

Canvas and WebGL apps, including drawing tools, maps and charts, are described only as canvas. data, ocr and look are available, but there is no coordinate click, drag or draw tool.
Login and bot walls need ask_human, which needs a display.
Benchmarks cover read, navigate and form-fill tasks on public and demo sites with Luna and Haiku; production authenticated and write-heavy workflows are not measured.
Long real-site rows were judged by an uncalibrated judge (79% agreement). long-local-v1 measures local fixtures with a deterministic full-ledger grader; long-live-v1 measures real sites with that judge.
The screenshot baseline is this project's own tool, not a third-party product.
Some live-v1 tasks were swapped after results; this is disclosed in bench/RESULTS.md.
Page content is untrusted data; see SECURITY.md.
This is a small v0.1.x beta project.