acrawl

io.github.Mingye-Lu由 GitHub 帳號 Mingye-Lu 發布v0.17.0更新於 Oct 30, 2026

Autonomous web crawler. 17 browser tools + goal-driven run_goal agent. Single binary, stealth.

概覽

AI 產生的概覽

自主網頁爬蟲與瀏覽器代理,讓助理瀏覽網站、擷取資料,並檢視網路、主控台與效能資訊。

可以做什麼
  • 導覽頁面、點擊、輸入、選取、懸停、按鍵並模擬手機,也可透過橋接擴充功能在你已登入的瀏覽器中操作
  • 從網站擷取結構化資料,靜態頁面走 HTTP,需要 JavaScript 時切換到真實瀏覽器
  • 檢視網路請求、主控台日誌、WebSocket 訊框、效能計時、Cookie、儲存空間與程式碼覆蓋率
  • 執行 axe-core 無障礙稽核,並排查頁面流程中的失敗原因
  • 執行目標導向的代理,自動判斷分頁並回傳 JSON,還可把爬取存成確定性腳本,無需模型呼叫即可重播
適用情境
適合讓助理端到端抓取或爬取網站、偵錯網頁的網路與主控台行為,或在探索一次後從 cron 或 CI 低成本重複執行爬取。
執行需求
以本機程序方式執行單一二進位檔(Linux、macOS 或 Windows)。瀏覽器功能需要 Node.js 20+ 及隨附的隱匿瀏覽器。需透過 acrawl auth 指令設定 LLM 供應商並使用自己的 API 金鑰,支援 25 家供應商。
安裝前請注意
它會驅動真實瀏覽器,可在線上網站送出表單、點擊和輸入,可能改變遠端狀態。它使用你自己的 LLM API 金鑰,模型呼叫會產生費用。橋接擴充功能使其能在你已登入的瀏覽器工作階段中操作,可能觸及你的帳號。安裝腳本會下載並執行發行版二進位檔。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 acrawl,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

[acrawl]

Browser automation agent that acts and observes. Navigate, click, and extract — or inspect network traffic, debug console errors, profile performance, and audit accessibility. One Rust binary.

[CI] [License: MIT] [Rust] [Website]

Website · Docs · Install · Quick start · MCP server · Comparisons

Single Rust binary. Full DevTools observability. 43 tools. 25 LLM providers. MCP server built-in.

[acrawl architecture: the TUI, one-shot prompt, and MCP entry points feed a runtime turn loop that streams from an LLM provider and runs 43 tools, which fetch over HTTP, escalate to CloakBrowser, or drive the user's Chrome through a WebSocket extension bridge.]


text
acrawl > get the name, price and rating of every book on books.toscrape.com, all 50 pages

acrawl reads the goal, opens the site, works out the pagination and hands back JSON. Along the way the agent can watch the network, read the console, profile performance and audit accessibility, so it can debug a page as well as scrape it. It also runs as an MCP server, so Claude Code, Cursor, VS Code and 14 other clients can drive the same browser.

Install

Linux / macOS (x64, ARM64)

bash
curl -fsSL https://raw.githubusercontent.com/Mingye-Lu/AgenticCrawler/main/install.sh | bash

Windows (x64, PowerShell)

powershell
irm https://raw.githubusercontent.com/Mingye-Lu/AgenticCrawler/main/install.ps1 | iex

The script downloads the latest release, verifies its SHA256 checksum and sets up CloakBrowser, the stealth browser acrawl uses for JavaScript-heavy pages. Browser features need Node.js 20+. To build from source, see the install guide.

Quick start

bash
acrawl auth anthropic          # or openai, google, other (Ollama, LM Studio, ...), 25 providers in allacrawl                         # interactive terminal UIacrawl prompt "list the titles of the top 10 stories on news.ycombinator.com"   # one goal, then exit

Then go further:

bash
acrawl mcp install                     # add acrawl to Claude Code, Cursor, VS Code and other MCP clientsacrawl run books_catalog -o books.json # replay a saved script with no LLM calls

The quick start walks through a first crawl, getting JSON out, and writing goals that work.

What you can do with it

Extract data from any site. Describe the fields you want; acrawl fetches static pages over HTTP and switches to a real browser only when a page needs JavaScript.

text
acrawl > get all repository names and star counts from github.com/trending

Debug a page like a developer would. Network requests, console logs, WebSocket frames, performance timings, cookies, storage, code coverage and axe-core accessibility audits are all agent tools.

text
acrawl > open localhost:3000/checkout, submit the form, and tell me which requests failed and why

Fill forms and work through flows. Click, type, select, hover, press keys and emulate phones. With the acrawl Bridge extension, it works inside your own logged-in Chrome, Edge or Brave.

Run it again for free. Save a crawl as a deterministic JSON script with loops, conditions and parallel branches, then replay it from cron or CI with acrawl run, with no model calls.

Split big jobs. The agent forks sub-agents onto separate tabs, and a URL-claiming registry stops two of them from crawling the same page.

Why acrawl

  • One binary. No Python environment, no Docker, no hosted service. Node.js is needed only for the bundled browser.
  • Acts and observes. Most browser agents navigate and click. acrawl's agent also sees the network, the console and performance, as first-class tools.
  • AI where you need it, scripts where you don't. Let the agent explore once; replay the result deterministically.
  • Your choice of model. 25 providers, including Anthropic, OpenAI, Gemini, DeepSeek, Bedrock, Azure, Vertex AI, GitHub Copilot, Groq, Mistral and OpenRouter, plus any OpenAI-compatible endpoint such as Ollama.
  • MCP both ways. Use other MCP servers as extra agent tools, or expose acrawl's 40 tools to your coding agent with acrawl mcp.
  • Open source. MIT-licensed, runs on your machine, with your own LLM key and no telemetry.

Bring your own model

[Anthropic, OpenAI, Google Gemini, DeepSeek, Amazon Bedrock, Azure OpenAI, Vertex AI, GitHub Copilot, GitLab Duo, Groq, Cerebras, Mistral AI, xAI, Cohere, Alibaba, OpenRouter, Together AI and Ollama]

Use it from your coding agent

acrawl mcp install sets acrawl up as an MCP server in any of these clients.

[Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, OpenCode, Zed, TRAE, JetBrains, Gemini CLI, Qwen Code, Codex CLI, Hermes, OpenClaw, Goose, Crush and Aider]

How it compares

acrawlbrowser-usePlaywright MCPFirecrawl
What it isBrowser agent, CLI and MCP serverPython agent library and hosted agentMCP server for browser controlHosted scrape and crawl API
Own reasoning loopYes, localYes, local or hostedNo, your client decides each stepYes, hosted
InstallSingle Rust binaryPython packageNode packageHosted, or self-host
DevTools tools for the agentNetwork, console, WebSocket, performance, cookies, storage, coverage, axe-coreNot a focusConsole, networkNot a focus
Replays with no LLM callsYes, JSON scriptsNoNoNo
LicenseMITMITApache 2.0AGPL-3.0 engine

Reviewed October 2026. The full comparison has a page for each tool.

Documentation

Everything else lives on the website, mingye-lu.github.io/AgenticCrawler.

Driving acrawl from a coding agent? npx skills add Mingye-Lu/AgenticCrawler installs the acrawl-mcp skill, which teaches the agent when to use direct tools, scripts or run_goal.

Development

acrawl is a Rust workspace of 11 crates.

bash
cargo build --releasecargo test --workspacecargo clippy --workspace --all-targets -- -D warningscargo fmt --check

See CONTRIBUTING.md to get started, CHANGELOG.md for release notes and SECURITY.md to report a vulnerability.

License

MIT

來源:README.md,提交 e7f62f9

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

8
  1. v0.17.0最新Oct 29, 2026
  2. v0.16.0Oct 13, 2026
  3. v0.15.2Oct 12, 2026
  4. v0.15.0Oct 6, 2026
  5. v0.14.3Oct 1, 2026
  6. v0.14.2Sep 21, 2026
  7. v0.14.0Sep 20, 2026
  8. v0.13.5Sep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分