weber

name

weber

description

Use when the user wants web or browser automation but tool choice is unclear, including authorized browser interaction, local webapp UI testing, screenshots/traces, public crawling or scraping, batch data extraction, headless browser backends, anti-detection browsing, or choosing between web-access, Playwright, agent-browser, Camofox, Crawlee, Firecrawl/Browserless, Browser Use/Stagehand, and Obscura.

Weber Skill

Weber is a router, not a replacement for browser tools. Use it to choose the smallest reliable backend for an authorized web task, then load/follow the chosen tool's own skill or docs.

Hard Boundaries

• Work only on authorized systems, user-owned sessions, local apps, or public data that can be accessed normally.
• Do not frame the goal as bypassing access controls, rate limits, login requirements, bans, CAPTCHAs, or platform protections.
• Before adding dependencies, changing global config, installing a browser engine, or using a paid/cloud API, explicitly tell the user what will change, where, and why.
• Prefer existing local tools before installing anything.
• Do not export cookies, tokens, localStorage, credentials, or private page data into reusable artifacts.

Host Chrome And Isolated Runtimes

• If the task needs the user's logged-in Chrome, choose web-access CDP and repair that route before falling back to isolated backends.
• In MMS-launched sessions, prefer MMS_HOST_CONTEXT_JSON / MMS_OPS_ENV_SAFE_CONFIG for host path and WebAccess hints before guessing paths.
• MMS/Codex sandboxes may rewrite HOME/XDG_*; do not assume os.homedir() points at the real Chrome profile.
• Before declaring web-access unavailable from an isolated session, run the web-access dependency check with host-home hints; the current web-access scripts also read WEB_ACCESS_HOST_HOME, HOST_HOME, REAL_HOME, and os.userInfo().homedir:

WEB_ACCESS_HOST_HOME="$(python3 -c 'import os,pwd; print(pwd.getpwuid(os.getuid()).pw_dir)')" \
  node "${WEB_ACCESS_SKILL_DIR:-$WEB_ACCESS_HOST_HOME/.codex/skills/web-access}/scripts/check-deps.mjs"

• If the proxy is stale, kill only the 3456 listener and rerun check-deps; do not kill/restart the user's Chrome unless explicitly asked.
• Treat 3456 as the CDP proxy and 9222 as the real Chrome debug endpoint; if 9222 is already listening, keep Chrome running and reconnect the proxy.
• Do not switch to Playwright, agent-browser, Camofox, or Obscura for a logged-in-user task just because the isolated agent cannot see the host Chrome profile.
• If using the unified adapter for a logged-in-user task, pass requireLoggedInChrome: true so fallback cannot jump to an isolated backend.

Default Routing

Choose by task shape:

Need	Use
Search, source discovery, known URL extraction, official docs lookup	web-access with search/fetch/curl/Jina as appropriate
Logged-in user Chrome, internal sites, dynamic pages, social/content sites, exploratory browser navigation	web-access CDP
Local webapp verification, UI flow debugging, screenshots, traces, accessibility snapshots, deterministic CLI steps	playwright skill / Playwright CLI wrapper; for visual/UI evidence, create QA-ready red annotated screenshots with labels outside the target region and connector lines
Fast headless interaction with ref-based CLI, simple extraction, isolated sessions, no user Chrome login needed	agent-browser
Large URL queues, site crawl, retries, concurrency, structured datasets in a Node project	Crawlee or project-native crawler code
Managed crawling/scraping/search API is acceptable and API keys/cost are approved	Firecrawl or Browserless
Production automation mixing code with natural-language page handling	Stagehand
Autonomous browser agent experiments with an LLM loop	Browser Use
Authorized pages blocked by bot-detection heuristics, geo-specific public scraping with approved proxy	Camofox — fingerprint-consistent browser backend, REST API, compact a11y snapshots
Python crawling, adaptive element tracking, large-scale structured scraping with checkpoint/resume	Scrapling — Python framework, anti-detect fetchers, MCP built-in, install-on-demand
Experimental lightweight CDP-compatible engine for high-volume isolated headless work	Obscura, installed only when a concrete task justifies it

When uncertain, start with the least invasive option that can prove progress, then escalate only when evidence shows it is insufficient.

Execution Loop

1. Define success: the exact data, UI state, screenshot, trace, or artifact needed.
2. Check available local tools before installing: command -v agent-browser, command -v npx, command -v obscura, and project package files.
3. Pick one primary backend and keep the first attempt small.
4. Validate with evidence: extracted rows/counts, current URL/title, snapshot, screenshot, trace, or output file.
5. For visual/UI evidence, annotate full-page screenshots by marking the exact changed, broken, or verified region in red; prefer browser/Playwright locator bounding boxes, keep labels outside important content, and use connector lines.
6. If blocked, re-route based on the blocker instead of repeatedly retrying the same backend.

Installed Baseline Checks

Use these commands to refresh assumptions; versions change over time:

agent-browser --version
npx --yes --package @playwright/cli@latest playwright-cli --version
npm view playwright @playwright/cli @playwright/mcp agent-browser version --json
curl -s http://localhost:9377/health  # Camofox

The ClawSkills thesethrose/agent-browser entry is the same tool family as the local agent-browser skill/CLI. Do not install it again if agent-browser --version works.

Optional Backend: Scrapling

Python scraping framework with adaptive element tracking. Install-on-demand.

Use when:

• Crawlee insufficient: sites change structure frequently, need elements to auto-relocate
• Python AI pipeline needs crawling with MCP integration
• Large-scale crawling with checkpoint pause/resume needed

Install:

pip install "scrapling[all]" && scrapling install

Smoke test:

python3 -c "from scrapling import Fetcher; print(Fetcher('https://example.com').status)"

Key API: Fetcher (HTTP), StealthyFetcher (anti-detect), DynamicFetcher (Playwright), Adaptor.find_similar() (adaptive tracking).

Limitations: Python 3.10+ required, heavier than Node alternatives for simple tasks.

Optional Backend: Camofox

Anti-detection headless browser server wrapping Camoufox (Firefox fork with C++ fingerprint spoofing). REST API on localhost:9377.

Use when:

• An authorized page is inaccessible to simpler automation because of bot-detection heuristics, and the task does not require bypassing login, CAPTCHA, bans, or access controls
• Need geo-specific public scraping with an approved proxy/locale/timezone setup
• Need compact a11y snapshots instead of raw HTML dumps

Install:

npm install @askjo/camofox-browser
npx camoufox-js fetch   # download ~300MB binary on first run
node node_modules/@askjo/camofox-browser/server.js  # start server

Smoke test:

curl -s http://localhost:9377/health
curl -s -X POST http://localhost:9377/tabs -H 'Content-Type: application/json' -d '{"url":"https://example.com"}'

Key API: POST /tabs (create), GET /tabs/:id/snapshot (a11y), POST /tabs/:id/click (by ref), POST /tabs/:id/navigate, DELETE /tabs/:id (close).

Limitations: Single instance, no existing login state; do not export or inject user cookies unless the user explicitly approves that flow.

Optional Backend: Obscura

Treat Obscura as experimental and install-on-demand only.

Use it when all are true:

• The task is authorized public or user-owned data extraction.
• The workload is batch/headless enough that Chrome/Playwright overhead matters.
• No existing local backend is sufficient.
• The user has been told what binary or source build will be installed.

Safe smoke test after install:

obscura fetch https://example.com --eval "document.title"
obscura fetch https://example.com --dump links

Do not make Obscura the default backend until it passes task-specific smoke tests on the user's machine.

Unified Browser Adapter

lib/ 目录提供跨后端统一接口，屏蔽 Camoufox/Playwright/agent-browser/web-access 差异。

import { createBrowser } from './lib'

// 自动选后端 + 降级
const browser = await createBrowser({ backend: 'auto' })
const session = await browser.exec(a => a.open('https://example.com'))
await browser.exec(a => a.click(session, 'button.submit'))
const shot = await browser.exec(a => a.screenshot(session, { path: '/tmp/result.png' }))
await browser.exec(a => a.close(session))

// 登录态 Chrome 任务：禁止降级到 isolated backend
const browser = await createBrowser({ requireLoggedInChrome: true })

// 指定后端
const browser = await createBrowser({ backend: 'camoufox' })

降级链：web-access → playwright → agent-browser → camoufox（可配置）

各 adapter 详见：lib/adapters/ 目录。

References

Read references/backend-map.md when comparing new tools, deciding whether to install Obscura or a crawler, or updating this router.