Skip to content

browser.md

Latest commit

 

History

History
46 lines (34 loc) · 5.2 KB

browser.md

File metadata and controls

46 lines (34 loc) · 5.2 KB
Raw

Browser

Driving a real browser, scraping the web, and end-to-end testing.

Which one: if you want the agent to click through a page during a coding session, start with Playwright CLI — it's the token-efficient default. Use Playwright MCP instead when a long exploratory loop needs persistent state. Use Chrome DevTools MCP when you're debugging a live page rather than automating it. Use browser-use or browser-harness when you're building a standalone automation. Use Firecrawl when you want content out of the web at scale rather than interaction.

Driving a browser from a session

  • Playwright CLI [CLI] (Microsoft, 12.4k★) — a command-line interface for browser automation built specifically for coding agents. A persistent daemon keeps browsers warm, and ref-based accessibility snapshots make interactions deterministic.

    • Use when: default choice for browser work inside a coding session. CLI invocations avoid loading large tool schemas and verbose accessibility trees into context, which matters when you're already juggling a big codebase.
    • Covers: open click fill type check screenshot pdf eval · navigation · press hover drag upload · tabs · cookies and localStorage · route network mocking · console, tracing, video capture.
    • Needs: npm install -g @playwright/cli. Chrome, Firefox, WebKit and Edge; multiple isolated sessions with separate state.
    • Note: the official comparison is explicit — CLI for high-throughput agents on limited context, MCP for specialized agentic loops where token cost matters less.

  • Playwright MCP [MCP] (Microsoft) — cross-browser automation through the accessibility tree: navigate, click, fill, screenshot, assert.

    • Use when: you want repeatable end-to-end flows, or a long-running exploratory loop that benefits from structured parameter passing and persistent introspection.
    • claude mcp add playwright npx @playwright/mcp@latest
    • Note: the official successor to the deprecated Puppeteer MCP — don't use the old one.

  • Chrome DevTools MCP [MCP] (Google) — a live, inspectable Chrome: DOM, console, network, performance traces.

    • Use when: debugging a real rendered page or a perf issue that only reproduces in a browser.
    • npx chrome-devtools-mcp@latest
    • Overlaps Playwright: DevTools for debugging and perf, Playwright for repeatable test flows. You rarely need both in one task.

Building an automation

  • browser-use [Library] (108k★) — Python library for building agents that drive a browser: click, type, fill forms, extract data. Playwright underneath, works with Claude/GPT/Gemini or their own tuned models.

    • Use when: you're writing a scraper, a QA suite or a filing bot as its own program. Ranks #1 on the Odysseys long-horizon benchmark at 87.4% over 200 tasks, ahead of the CV-based computer-use agents.
    • Needs: Python 3.11+. Optional CLI, registers as a skill for Claude Code / Cursor / Codex. A hosted cloud version exists for production.
    • Note: not an MCP server. For driving a browser from inside a session, use the tools above instead.

  • browser-harness [CLI] [Library] (browser-use, 16.5k★) — a deliberately thin CDP harness: the agent writes whatever helper code is missing during execution, and the harness keeps the result. It improves every run and accumulates site-specific domain skills.

    • Use when: the target site keeps breaking your selectors, or you want automation that hardens itself instead of being re-patched by hand.
    • Needs: Python 3.12+, Chrome/Chromium with remote debugging enabled. uv install browser-harness, then register the skill via browser-harness skill. Optional Browser Use Cloud key (free tier).
    • Versus browser-use: browser-use is the batteries-included library; browser-harness is minimal on purpose and grows into the shape of your target.

Getting content out of the web

  • Firecrawl [Library] [MCP] (162k★) — scrape, crawl, search, map and extract at scale. URLs to markdown, HTML, screenshots or structured JSON; an agent mode gathers data from a natural-language prompt.

    • Use when: you want the content, not the interaction — research, ingestion, competitive sweeps, populating a dataset.
    • Modes: scrape · crawl (whole site) · map (discover URLs) · search (results with full page content) · batch scrape · extract.
    • Needs: self-host under AGPL-3.0, or the paid cloud at firecrawl.dev for the extras. pip install firecrawl-py / npm install firecrawl. SDKs for Python, Node, Go, Java, Rust, Ruby, .NET, PHP, Elixir. Ships a Firecrawl MCP.
    • Note: AGPL-3.0 on the self-hosted version is a real consideration if you're embedding it in a closed product.

  • scrapfly-scrapers [Library] (1.1k★) — maintained Python scrapers for 40+ popular sites.

    • Use when: your target is already covered and you'd rather not rediscover its anti-bot quirks. Read the relevant scraper as a reference even if you don't run it.