Web BrowserSAFE
These are commands I use with agents, mostly Claude
Overview
These are commands I use with agents, mostly Claude
4c7f30d883efOBSERVED · 2026-10-23What it tells the agent
The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.
---
name: web-browser
description: "Automate and interact with web pages through Chrome or Chromium using the Chrome DevTools Protocol (CDP): navigate, click, fill forms, inspect content, take screenshots, and debug console or network activity. Use when an agent needs a real browser. Prefer headless Chrome unless visible browser interaction is required."
license: Stolen from Mario
---
# Web Browser Skill
Minimal CDP tools for collaborative site exploration.
## Start Chrome (Prefer Headless)
```bash
./scripts/start.js --headless # Recommended: isolated reusable profile
./scripts/start.js --headless --profile # Headless with a copy of your profile
./scripts/start.js # Visible browser window when needed
./scripts/start.js --headless --reset-profile # Clear cached profile before launch
```
Starts Chrome with remote debugging (default port `:9222`). **Agents should use `--headless` by default** because it is less disruptive and supports navigation, evaluation, screenshots, emulation, and logging. User extensions are disabled for headless launches; a copied profile still provides its cookies and other browser state, but extension-based features are unavailable. Headed launches continue to load extensions normally. Use headed mode only when a person needs to see or interact with the browser, such as for `pick.js`, manual authentication, or debugging a headless-specific difference.
The start script only reuses a running browser when its profile and launch settings match. Stop the running skill browser before switching between headless and headed mode. Changing the debugging port does **not** isolate the profile: startup refuses to modify a profile already used on another port and preserves Chrome's profile locks.
Profile behavior:
- Default mode uses: `~/.cache/agent-web/browser/fresh-profile`
- `--profile` mode uses: `~/.cache/agent-web/browser/profile-copy`
- The skill **does not attach to your live Chrome profile directly**
- If `:9222` is already used by an unknown instance, start will fail instead of reusing it
If Chrome is installed in a non-standard location, set:
```bash
BROWSER_BIN=/path/to/chrome ./scripts/start.js --headless
```
Optional debug endpoint override:
```bash
BROWSER_DEBUG_PORT=9333 ./scripts/start.js --headless
```
## Stop Chrome
```bash
./scripts/stop.js # Stop skill Chrome on :9222
BROWSER_DEBUG_PORT=9333 ./scripts/stop.js # Stop skill Chrome on :9333
./scripts/stop.js --all # Stop all skill Chrome instances
```
Works for headless and headed browsers. Only processes using the skill's cached profiles are stopped, not your regular Chrome. Shutdown preserves cookies and other cached state; it never force-kills Chrome. Stop before using `--reset-profile`.
To switch to a visible browser:
```bash
./scripts/stop.js
./scripts/start.js
```
## Navigate
```bash
./scripts/nav.js https://example.com
./scripts/nav.js https://example.com --new
```
Navigate current tab or open new tab.
## Device Emulation (Mobile)
```bash
./scripts/emulate.js --list
./scripts/emulate.js iphone-14
./scripts/emulate.js pixel-7 --landscape
./scripts/emulate.js --reset
```
Set an active device emulation preference (viewport, DPR, touch, UA) for browser skill commands. Use `--reset` to clear.
Commands like `nav.js`, `eval.js`, `pick.js`, `dismiss-cookies.js`, and `screenshot.js` automatically apply the active preference.
## Evaluate JavaScript
```bash
./scripts/eval.js 'document.title'
./scripts/eval.js 'document.querySelectorAll("a").length'
./scripts/eval.js 'document.querySelector("button")?.click(); "clicked"'
./scripts/eval.js 'await Promise.resolve(document.title)'
./scripts/eval.js 'JSON.stringify(Array.from(document.querySelectorAll("a")).map(a => ({ text: a.textContent.trim(), href: a.href })).filter(link => !link.href.startsWith("https://")))'
```
Execute JavaScript in the active tab. Input can be an expression or statement list; the console-style completion value is printed and promises/top-level `await` are awaited. Be careful with string escaping, best to use single quotes.
## Screenshot
```bash
./scripts/screenshot.js
./scripts/screenshot.js --full-page
./scripts/screenshot.js --device iphone-14
./scripts/screenshot.js --device pixel-7 --full-page
```
Takes a screenshot and returns a temp file path.
- Default: current viewport
- `--full-page`: captures full document height
- `--device <preset>`: temporary mobile emulation for that screenshot only
## Pick Elements
```bash
./scripts/pick.js "Click the submit button"
```
Interactive element picker. Click to select, Cmd/Ctrl+Click for multi-select, Enter to finish. This requires headed Chrome; launch `start.js` without `--headless`.
## Dismiss Cookie Dialogs
```bash
./scripts/dismiss-cookies.js # Accept cookies
./scripts/dismiss-cookies.js --reject # Reject cookies (where possible)
```
Automatically dismisses EU cookie consent dialogs.
Run after navigating to a page:
```bash
./scripts/nav.js https://example.com && ./scripts/dismiss-cookies.js
```
## Quick Mobile Debug Flow
```bash
./scripts/start.js --headless
./scripts/nav.js https://example.com
./scripts/emulate.js iphone-14
./scripts/nav.js https://example.com # reload with mobile UA
./scripts/dismiss-cookies.js
./scripts/screenshot.js --full-page
```
## Background Logging (Console + Errors + Network)
Automatically started by `start.js` and writes JSONL logs to:
```
~/.cache/agent-web/logs/YYYY-MM-DD/<targetId>.jsonl
```
Manually start:
```bash
./scripts/watch.js
```
Tail latest log:
```bash
./scripts/logs-tail.js # dump current log and exit
./scripts/logs-tail.js --follow # keep following
```
Summarize network responses:
```bash
./scripts/net-summary.js
```Trust audit
SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | PASS |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- none-observed
- Network
- declared (5 observation(s))
- Shell
- declared (1 observation(s))
- Dependencies
- not all pinned
- Secrets in source
- none-found
Findings (1)
ws
Gates applied: no_behavioural_pass.
4c7f30d883effull audit observations/trust-audit/skill/mitsuhiko__web-browser.json · Report an issue / request a re-scanAudit history
Every audit this skill has had.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-23 | 4c7f30d883ef | SAFE | B | 89 | first audit |
Questions
What does the web-browser — mitsuhiko/agent-stuff skill do?
These are commands I use with agents, mostly Claude
Is web-browser — mitsuhiko/agent-stuff safe to install?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean skill reads B.
What can web-browser — mitsuhiko/agent-stuff access on my machine?
The audit observed that it reaches the network and runs shell commands. Each of those is consistent with what it says it does. Secrets in the source: none found.
How current is this page?
The grade is for one exact copy of the source (4c7f30d883ef), read on 2026-10-23. The repository is watched, and a new audit runs when it changes — this is the first audit.
Add this badge to your README
Shows this page's audit result and links back to it. It updates when the audit does.
[](https://modelsatlas.com/skills/mitsuhiko/web-browser/)
Not your listing, or something out of date? Claim or update your listing.