agent-browser
/agent-browser
Use when you need to screenshot or drive a dev server and the shared Chrome is unavailable or contended — e.g. parallel Conductor workspaces: headless per-workspace browser via npx agent-browser.
- Author
- itsvedantkumar ↗
- Repository
- vstack ↗
- Licence
- MIT
- Last repository change
- 06.09.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 1
No signals detected. Repository changed on 06.09.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Use when you need to screenshot or drive a dev server and the shared Chrome is unavailable or contended — e.g. parallel Conductor workspaces: headless per-workspace browser via npx agent-browser.
Original author description
Use when you need to screenshot or drive a dev server and the shared Chrome is unavailable or contended — e.g. parallel Conductor workspaces: headless per-workspace browser via npx agent-browser.
How it works
- Isolate the workspace first
- The ui-iterate loop
- Interact via accessibility-tree refs
- Done? Close it
- Which browser tool, and when
- Accessibility, without hand-rolling it
SKILL.md preview
---
name: agent-browser
description: "Use when you need to screenshot or drive a dev server and the shared Chrome is unavailable or contended — e.g. parallel Conductor workspaces: headless per-workspace browser via npx agent-browser."
---
# agent-browser — headless per-workspace browser
Browser automation CLI from Vercel Labs. Native Rust binary driving Chrome/Chromium over
CDP — no Playwright, no shared Chrome window. Each `--session` gets its own browser
instance, cookies, and history, so parallel Conductor workspaces never fight over tabs.
Needs node/npx. Run every command through `npx agent-browser ...` (or `agent-browser` if
installed globally with `npm i -g agent-browser`). First use on a machine may need
`npx agent-browser install` to download Chrome for Testing; existing Chrome, Brave,
Playwright, and Puppeteer installs are detected automatically.
## Isolate the workspace first
```bash
SESSION="$(npx agent-browser session id --scope worktree --prefix dev-loop)"
# Then pass --session "$SESSION" on every call, or export AGENT_BROWSER_SESSION="$SESSION"
```
The browser persists via a background daemon, so chaining commands with `&&` is safe.
## The ui-iterate loop
Screenshot the affected route at all three widths, then critique the images:
```bash
URL="http://localhost:${CONDUCTOR_PORT:-3000}/route-you-changed"
npx agent-browser --session "$SESSION" open "$URL"
npx agent-browser --session "$SESSION" wait --load networkidle
npx agent-browser --session "$SESSION" set viewport 375 812 # phone
npx agent-browser --session "$SESSION" screenshot /tmp/ui-375.png
npx agent-browser --session "$SESSION" set viewport 768 1024 # tablet
npx agent-browser --session "$SESSION" screenshot /tmp/ui-768.png
npx agent-browser --session "$SESSION" set viewport 1440 900 # desktop
npx agent-browser --session "$SESSION" screenshot /tmp/ui-1440.png
```
Read each PNG with the Read tool and critique it (hierarchy, spacing, states) per the
ui-iterate skill. Add `--full` to `screenshot` for the full page instead of the viewport.
## Interact via accessibility-tree refs
`snapshot` prints the accessibility tree with stable `@eN` element refs — more reliable
than CSS selectors for driving state (hover, open menus, filled forms) before a shot:
```bash
npx agent-browser --session "$SESSION" snapshot -i # interactive elements only, with refs
npx agent-browser --session "$SESSION" click @e2 # click by ref
npx agent-browser --session "$SESSION" fill @e3 "test@example.com" # clear + fill by ref
npx agent-browser --session "$SESSION" snapshot -i # re-snapshot after the page changes
```
Refs go stale on navigation or DOM changes — always re-run `snapshot` before reusing them.
If a click fails because another element covers the target (consent banner, modal), dismiss
the reported covering element and take a fresh snapshot before retrying.
## Done? Close it
```bash
npx agent-browser --session "$SESSION" close
```
More commands (`wait --text`, `hover`, `press`, `scroll`, `diff screenshot`, auth/state):
`npx agent-browser --help`, or `npx agent-browser skills get core` for the full workflow
guide served by the installed CLI. `skills get dogfood` is an adversarial exploratory-QA pass,
which complements a design review of a surface you already know about.
## Which browser tool, and when
Two browser lanes exist here and they must not fight over the same Chrome.
**claude-in-chrome** drives your real, logged-in browser. Use it when the session needs your
cookies, an authenticated app, or you want to watch what happens.
**agent-browser** runs its own headless Chrome. Use it whenever the shared browser would be
contended or is unavailable: parallel Conductor workspaces, background runs, anything
unattended. Always pass `--session <workspace>` and `--pin-tab`, which is what keeps two
workspaces from adopting each other's tabs.
Requires Node 24 or newer. On an older Node the CLI still runs but npm prints an engine
warning; if a command behaves strangely, check the Node version before debugging further.
## Accessibility, without hand-rolling it
```bash
npx agent-browser --session "$SESSION" a11y --tags wcag2a,wcag2aa
```
Runs axe-core in-page, offline and CSP-safe, scoped by selector and aware of iframes. Prefer it
to writing WCAG checks by hand in a review.
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · itsvedantkumar/vstack | 95 | MIT |
| vanducng/skills | 100 | MIT |
| mvschwarz/openrig | 100 | Apache-2.0 |
| mgiovani/cc-arsenal | 100 | MIT |
| desplega-ai/agent-swarm | 100 | MIT |
| nexu-io/open-design | 100 | Apache-2.0 |
| seaworld008/commonly-used-high-value-skills | 100 | MIT |
| slopus/happy | 100 | MIT |
| j62268781-alt/dsh-toolkit | 100 | MIT |
Database snapshot: 13.11.2026 · package version ab0be3149195 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100browser-handoff
Write a self-contained handoff block for a browser task that the user's local Claude Code session (connected to their Chrome with `claude --chrome`) will run. Use when an agent in a Conductor cloud workspace needs a setup or admin task done in the user's real browser (a… - 100use-browser
主浏览器技能:通过 browser_session / browser_tab / browser_page / browser_client 四个工具操作本机 Chrome(需 Ethan Browser 扩展)。嵌入本机真实 Cookie,能操作用户已登录的浏览器、接管当前 tab、做 snapshot/click/fill/screenshot/eval。支持多个浏览器同时连接,每个有自己的名字,agent 按需选择操作哪个端。当用户要求操作浏览器、网页自动化、点击输入、或需要用到本机真实 Chrome(含登录态)时触发。兜底用 agent-browser。 - 90remote-browser
Operate a remote/hosted browser session for agent automation when local browser tools are unavailable. Use for remote navigation, interaction, and observation. - 91browser
Drives the agent's own browser, or the operator's Chrome for their accounts. Use when WebFetch or WebSearch fails, times out, or returns a 403, an empty page or a JavaScript-only shell; when a page needs clicks, forms, scrolling or sign-in; or when a screenshot or exact rendered… - 100action-browser
Use Action Browser when the user wants an agent to open a URL in a real Chrome session, inspect a page, click or fill a DOM element, or capture a browser screenshot without setting up a headless browser. - 100agent-workspace-linux
Use when a task needs a hidden Linux desktop running in the background or a workspace-owned browser: GUI app QA, web/browser/shopping automation, app observation, or stale workspace cleanup. Routes agent-workspace-linux MCP tools on demand. Does NOT apply to host desktop/Chrome…