# Agent Browser

> Control websites through the local agent-browser CLI for navigation, screenshots, form interaction, login flows, text extraction, and website-access testing. Use when auditing whether websites are reachable or extractable, when the Codex in-app Browser plugin is unavailable/insufficient, or when a repeatable shell-driven browser workflow is needed. ANTI-PATTERN: Do not use for pure HTTP checks that curl can answer, source-code review, or restricted authenticated content without credentials.

- Skill: `cryptopafi/agent-browser` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cryptopafi/agent-browser`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cryptopafi/agent-browser/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: cryptopafi (https://skillmd.com/u/cryptopafi)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/cryptopafi/agent-browser

---


# Agent Browser

Use this skill for shell-driven browser automation with the local `agent-browser` CLI. It is most useful for website-access audits, form/login checks, screenshots, and extracting visible page text when `curl` is not enough.

Prefer the Codex Browser plugin when it is available and the target can be handled in-app. Use this skill when you need repeatable CLI automation, session persistence, downloads, or direct page text/screenshot artifacts from shell.

## Core Loop

1. Normalize the target URL.
2. Open the page.
3. Wait for load or a bounded timeout.
4. Capture status evidence: URL, title, screenshot, and visible text.
5. Classify the outcome.
6. Retry with a fallback method only when the first failure is inconclusive.

## Essential Commands

```bash
agent-browser open "https://example.com"
agent-browser wait --load networkidle
agent-browser get url
agent-browser get title
agent-browser get text body
agent-browser screenshot --full
agent-browser snapshot -i
```

For interactive pages:

```bash
agent-browser snapshot -i
agent-browser click @e1
agent-browser fill @e2 "value"
agent-browser press Enter
agent-browser wait --load networkidle
```

## Website Access Classification

Use these statuses for platform audits:

- `OK_PUBLIC`: page loads and meaningful public content is extractable.
- `OK_PARTIAL`: page loads but key data is hidden, paginated, script-heavy, or incomplete.
- `LOGIN_REQUIRED`: login wall blocks the needed information.
- `BOT_BLOCKED`: CAPTCHA, Cloudflare challenge, 403 anti-bot, or similar block.
- `GEO_BLOCKED`: explicit country/region restriction.
- `TIMEOUT`: page did not load within bounded retries.
- `DOWN`: DNS/TLS/connection failure or 5xx confirmed after retry.
- `UNKNOWN`: conflicting evidence; manual review needed.

## Safety And Credentials

- Never print passwords, cookies, tokens, or private session state.
- If credentials are provided, prefer encrypted `agent-browser auth` or session state files outside project deliverables.
- Do not bypass access controls, paywalls, CAPTCHAs, or account restrictions.
- Public extraction is allowed; authenticated extraction requires valid credentials and a legitimate use case.

## When to Use vs Alternatives

- `browser`: use first for normal Codex in-app website inspection and visible browser work.
- `agent-browser`: use for repeatable shell-based access tests, screenshots, downloads, or batch page extraction.
- `apify-ultimate-scraper`: use when platform-specific scraping actors exist and credentials/API tokens are available.
- `dogfood`: use for systematic UX/bug exploration of one web app, not for a broad list of affiliate platforms.