Give your agent a real browser.

Headless turns browser QA into a small, inspectable command surface, so an agent can navigate, test, record, and explain without screen coordinates or a brittle script stack.

See the workflow Read the docs

localhost:3000

private session

HWorkspaceDocumentation

New auditCheck a page with your agent.

Enter a local URL. Headless keeps the browser control plane private.

http://localhost:3000/dashboardRun audit

$headless inspect --context actions --task "run an audit"

>8 elements · semantic refs ready

P2 benchmark / 27 Aug 2026

Smallest agent surface.
Measured, not claimed.

5 fresh Linux ARM64 containers per case. The table reports medians for the same dashboard workflow.

Tokens · vs Selenium + Python58%

fewer estimated agent tokens

Tokens · vs Puppeteer65%

fewer estimated agent tokens

CPU time · vs Selenium + Python32%

less median CPU time

CPU time · vs Puppeteer34%

less median CPU time

Agent token footprint / lower is better

Headless warm = 1×

Headless cold1.25x

Headless warm1x

Selenium with Python2.36x

Puppeteer2.87x

Wall time vs CPU time per run

The gap is waiting, not computing

cold3.41s1.79s

warm3.24s1.22s

Python2.83s1.79s

Puppeteer2.40s1.85s

Workflow	Estimated tokens	Wall time	CPU time	Peak memory
Headlesscold	218	3.413 s	1.794 s	381 MiB
Headlesswarm	174	3.239 s	1.218 s	379 MiB
SeleniumPython	410	2.828 s	1.788 s	378 MiB
Puppeteer	499	2.400 s	1.849 s	367 MiB

Warm starts with the Headless host and session ready. Cold includes starting both.Estimated tokens measure workflow source size, not billed LLM tokens. Point-in-time snapshot; repeat before comparing a change.

Read the method

The control loop

From intent to evidence—

without touching the desktop.

Visit

Open a local or authenticated page.

Act

Click, fill, press, scroll, and wait by meaning.

Prove

Return recordings, screenshots, reports, and diffs.

hermes-vm — headless

secure local session

$ headless start ✓ ready

$ headless session create qa

$ headless --session qa visit localhost:3000/designers/dashboard

$ headless --session qa record start --fps 10

$ headless --session qa click --role button --name Continue

$ headless --session qa report create --output pr-report.json ✓ evidence ready

Made for real QA

When the page breaks,
the agent can look closer.

Deep diagnostics are available on demand, not forced into every interaction.

Observe

Accessible elements, rendered media, console, network, CSS, cookies, and storage—only when the agent needs them.

Capture

Browser-pixel MP4 recordings and private PNG artifacts built for frontend QA and PR review.

Reproduce

Reusable flows, visual comparisons, throttling, offline mode, and exact API mocks.

Contain

A private Unix socket, no exposed DevTools port, bounded commands, and safe navigation rules.

Run it where your agent runs

Your VM has a browser
now. Use it.

One CLI protocol. The same commands over MCP. Purpose-built for the last mile of software QA.

Open Headless

Headless began as a fork of antiwork/chromeless . With thanks to Sahil Lavingia for the original foundation.