acrawl

  █████╗  ██████╗██████╗  █████╗ ██╗    ██╗██╗     
 ██╔══██╗██╔════╝██╔══██╗██╔══██╗██║    ██║██║     
 ███████║██║     ██████╔╝███████║██║ █╗ ██║██║     
 ██╔══██║██║     ██╔══██╗██╔══██║██║███╗██║██║     
 ██║  ██║╚██████╗██║  ██║██║  ██║╚███╔███╔╝███████╗
 ╚═╝  ╚═╝ ╚═════╝╚═╝  ╚═╝╚═╝  ╚═╝ ╚══╝╚══╝ ╚══════╝

LLM-powered web crawler. Describe what you want in plain English — get structured data back.

Single binary. No Python runtime. 21 tools. 25 LLM providers. MCP server built-in.

Why acrawl?

Most web scraping still means writing code: XPath selectors, pagination logic, retry handling, anti-bot workarounds. LLMs can read pages like humans do, but wiring one up to a browser is a project in itself.

acrawl is that wiring, packaged as a single Rust binary. You describe a goal; the agent figures out which pages to visit, what to click, what to extract, and when it's done.

• No code required. Describe the goal in English. The agent plans and executes.
• One binary, zero runtimes. cargo build --release produces a self-contained executable. No Python, no Node runtime — just Rust and a Chromium download for browser automation.
• Smart fetching. Static pages are served over HTTP (fast). When JavaScript or interaction is needed, acrawl detects JS framework markers (__next_data__, __nuxt, __vue, ng-app, React roots), auth redirects, and short <noscript> bodies — then transparently escalates to a headless browser.
• 21 tools, not a chatbot. The agent has real tools — navigate, click, fill forms, run JS, take screenshots, manage tabs — plus a fork/join layer to spawn parallel sub-agents across multiple browser tabs.
• 25 LLM providers. Anthropic, OpenAI, Google Gemini, DeepSeek, AWS Bedrock, Azure OpenAI, Vertex AI, GitHub Copilot, Groq, Mistral, xAI, Cohere, Alibaba DashScope, OpenRouter, and more. Or bring your own via any OpenAI-compatible endpoint.
• MCP client. Extend the agent with custom tools via Model Context Protocol servers (stdio, SSE, HTTP, WebSocket).
• MCP server. acrawl mcp exposes all 17 browser tools plus an autonomous run_goal agent to any MCP-compatible client — Claude Code, Cursor, Windsurf, VS Code, Zed, JetBrains, TRAE, Gemini CLI, and more. Install with acrawl mcp install.

How does it compare?

vs. AI web agents and scraping tools

Notes:

• browser-use (85k+ GitHub stars): Python + Playwright, DOM + screenshots, supports GPT/Claude/Gemini/Ollama via LiteLLM, 89.1% WebVoyager. No single binary — requires Python and pip install. Every action calls an LLM: 2-5s/step, ~$0.02-0.30/task. Cloud tier adds stealth; self-hosted is bare Playwright.
• Stagehand (Browserbase, 21k+ stars): TypeScript + CDP (v3), mixes deterministic Playwright with AI primitives (act(), extract(), observe()). Action caching reuses successful clicks without re-calling the LLM. Requires Node and, for production, Browserbase cloud hosting.
• Skyvern (21k+ stars, Apache 2.0): vision-first (screenshot-only, no DOM), handles legacy portals and government forms that DOM tools struggle with. No-code cloud UI available. Each step costs vision-model tokens — ~$0.10-0.50/task. 85.85% WebVoyager.
• Firecrawl (82k+ stars): managed scraping API. Returns LLM-ready Markdown, JSON extraction, site-wide crawl. Not an agentic tool — minimal multi-step interaction. Ships an official MCP server. Per-page pricing from $19/month.
• Playwright MCP (Microsoft, 29k+ stars): MCP server that exposes browser control via the accessibility tree. Sub-100ms actions, zero vision tokens. Drives an LLM client's browser rather than having its own reasoning — no autonomous goal navigation. Used in GitHub Copilot Agent.

vs. native LLM provider browsing

Most AI providers offer some form of browsing, but it is designed for conversational information retrieval, not programmatic web automation. Key constraints:

Notes:

• ChatGPT Agent (OpenAI, July 2025): runs in a sandboxed cloud virtual machine with its own Chromium instance. Can browse, click, and fill forms but pauses for user confirmation on sensitive actions (purchases, logins). Uses two modes: a fast text browser for research queries and a visual browser for interaction. Cannot run code in the browser, install extensions, or access your local file system. Susceptible to prompt injection. Available to Plus/Pro/Business subscribers.
• ChatGPT Atlas (OpenAI, October 2025): a full Chromium browser with ChatGPT integrated as a sidebar + agent. Agent mode drives the same sandboxed cloud VM as ChatGPT Agent; core limitations are identical.
• Claude Computer Use (Anthropic API, beta since October 2024): screenshot + mouse/keyboard API for any desktop application, not just browsers. Vision-only — no DOM access. Developers must provide and manage the entire computing environment (typically a Docker container with Xvfb + Firefox). Not a ready-to-use binary. Requires significant infrastructure to operate in production.
• Claude in Chrome (Anthropic Chrome extension, beta November 2025+): lets Claude operate within your existing Chrome session using your real cookies and logins. Available to Max plan subscribers. Not an open API — no programmatic control. Good for interactive personal tasks; not suitable for batch automation.
• Gemini / Deep Research (Google): browsing is grounded via Google Search API calls, not a live browser session. Deep Research synthesizes across many searches but cannot interact with pages (click, fill forms, navigate dynamically). Project Mariner (experimental computer use) is a separate, limited research preview.
• Copilot / Edge (Microsoft): Edge's Copilot Mode uses Bing retrieval with some ability to navigate pages. Real-world tests show high latency (6+ minutes for multi-page comparison tasks) and frequent interruptions for user confirmation. Not a developer API.

Quick Start

Install

Linux / macOS (x64 / ARM64):

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

Windows (x64, PowerShell):

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

This downloads the latest binary, verifies its SHA256 checksum, and sets up CloakBrowser for stealth browser automation. Requires Node.js 20+ for browser features.

acrawl checks for updates on startup and shows a notification when a new version is available.

git clone https://github.com/Mingye-Lu/AgenticCrawler.git
cd AgenticCrawler
cargo build --release

# Install CloakBrowser (required for browser automation — binary auto-downloads on first use)
npm install

Browser Extension (optional)

The acrawl Bridge extension lets acrawl control your real browser (with your sessions, cookies, and existing extensions) instead of a headless CloakBrowser instance. Download acrawl-extension.zip from the latest release, unzip it, then load it into your browser:

Enable Developer mode, click Load unpacked, and select the unzipped folder. Then run /extension in the acrawl REPL to connect. See extension/README.md for full setup details.

Configure

# Set up your LLM provider (interactive prompt)
./target/release/acrawl auth anthropic   # or: openai, other

Credentials are stored in ~/.acrawl/credentials.json. Override the config directory with ACRAWL_CONFIG_HOME.

Run

# Interactive REPL
./target/release/acrawl

# One-shot mode
./target/release/acrawl prompt "scrape all book titles and prices from books.toscrape.com"

# Resume a saved session
./target/release/acrawl --resume session.json /status /compact

Examples

Scrape a product catalog:

acrawl > scrape all book titles, prices, and ratings from books.toscrape.com

The agent navigates to the site, reads the page, extracts the data, paginates through all 50 pages, and returns structured JSON.

Fill and submit a form:

acrawl > go to example.com/contact, fill in name "Jane Doe", email "jane@example.com",
         message "Hello", and submit the form

The agent locates form fields, fills them in, clicks submit, and confirms the result.

Monitor a price:

acrawl > check the current price of "Rust in Action" on books.toscrape.com

Single-page extraction — the agent fetches, reads, and returns the price without unnecessary navigation.

Extract from JS-rendered pages:

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

Static HTTP won't work here. acrawl detects React/Next.js markers and automatically escalates to a headless browser to render the JavaScript.

Parallel multi-page crawl:

acrawl > scrape the title, author, and price of every book across all 50 pages on books.toscrape.com.
         Fork a sub-agent for each page to speed this up.

The agent spawns up to 5 concurrent sub-agents, each on its own browser tab, to crawl pages in parallel. Results are merged when all sub-agents finish.

Features

21-Tool Toolbox

Navigation

Content Formats

The navigate tool's format parameter controls how the page is returned:

fit_markdown works in two passes:

1. Hard-block removal, elements whose class or id attribute contains any of these strings are removed immediately: nav, footer, header, sidebar, ads, comment, promo, advert, social, share.

2. Score-based pruning, remaining elements are scored; anything below 0.48 is removed. The score is:

   0.4 × text_density
   + 0.2 × (1 − link_density)
   + 0.2 × tag_weight
   + 0.1 × class_id_score
   + 0.1 × ln(text_length + 1)
   

   Tag weights: article = 1.5 · h1 = 1.2 · h2 = 1.1 · h3/p/section = 1.0 · h4 = 0.9 · h5/table = 0.8 · h6 = 0.7 · span = 0.3 · div/li/ul/ol = 0.5.

Use markdown instead of fit_markdown when: the page has important content inside elements named sidebar, nav, or similar, for example, metadata panels, related-article links, or author info stored in a sidebar div.

If fit_markdown prunes all content (empty result), the tool automatically falls back to plain text.

Interaction

Content Extraction

Agent Control

page_state Reference

Interaction tools (click, click_at, fill_form, hover, press_key, go_back, scroll, switch_tab, select_option) all return a page_state object. There are two variants:

Full page_state, returned after the first interaction on a URL, or when changes are too extensive to diff:

{
  "url": "https://example.com/page",
  "title": "Page Title",
  "page_map": {
    "headings": [{ "level": 1, "text": "...", "id": "...", "selector": "...", "char_count": 0, "preview": "..." }],
    "landmarks": [{ "tag": "nav", "role": "navigation", "id": "...", "selector": "...", "text_preview": "..." }],
    "links": [{ "text": "...", "href": "...", "selector": "..." }],
    "interactive": { "counts": { "buttons": 0, "inputs": 0, "selects": 0, "textareas": 0, "total": 0 }, "elements": [] },
    "meta": { "title": "...", "url": "...", "description": "..." },
    "truncated_links": false,
    "truncated_forms": false,
    "truncated_landmarks": false
  }
}

Diff page_state, returned on subsequent interactions on the same URL, showing only what changed:

{
  "url": "https://example.com/page",
  "title": "Page Title",
  "changed": true,
  "changes": {
    "added_headings": [],
    "removed_headings": [],
    "added_links": [],
    "removed_links": [],
    "added_landmarks": [],
    "removed_landmarks": [],
    "added_interactive": [{ "selector": "...", "tag": "...", "text": "..." }],
    "removed_interactive": [],
    "modified_interactive": [{ "selector": "...", "tag": "...", "text": "...", "state_changes": { "aria_expanded": "true" } }]
  }
}

If nothing changed, { "url": "...", "title": "...", "changed": false } is returned. On bridge failure, { "url": "unknown", "title": "unknown", "page_map": null }. Caps: max 50 links, 20 landmarks per page_state.

Sub-Agent Parallelism

The agent can fork child agents to crawl multiple pages concurrently. Each child gets its own browser tab, step budget, and independent state.

URL Claiming

Before a child agent is spawned, its scope is registered in a shared claim registry. This prevents two sibling agents from crawling the same URL simultaneously.

Rules:

• First-claimer-wins, if a second agent tries to claim a URL already claimed by a sibling, the fork fails immediately with a conflict message naming the owner agent. The parent LLM sees this conflict and can adjust scope.
• Three scope types: SinglePage (exact URL), UrlList (all-or-nothing batch), UrlPattern (regex, checked for overlap with all existing claims).
• RAII lifetime, a claim is held for the life of the child. When the child finishes, is cancelled, or its parent aborts setup, the claim is released automatically and the URL becomes available again.
• Cross-type checking, an exact URL conflicts with any already-claimed regex that matches it, and a new regex conflicts with any already-claimed exact URL it would match.
• Intra-list deduplication, if the same URL appears twice in a UrlList, it is silently deduplicated (the LLM sometimes produces duplicates).

Claiming is automatic, it happens inside fork before the child starts, not inside navigate. The agent does not call it explicitly.

Smart Fetch Routing

Every navigate call goes through a two-tier fetch router:

1. HTTP first — fast reqwest-based fetch (30s timeout, follows up to 10 redirects).
2. Auto-escalation — if any of the following are detected, the request is transparently replayed in a headless browser:
   • HTTP 403, 429, or 503 responses
   • JS framework markers: __next_data__, __nuxt, __vue, ng-app, _react, data-reactroot
   • Auth redirects: URLs containing /login, /signin, /auth, /oauth, accounts.google.com
   • Short response body (< 500 chars) with a <noscript> tag

When --no-headless / --headed is set, all fetches go directly through the browser.

25 LLM Providers

Custom / Local Providers

To use any OpenAI-compatible endpoint (Ollama, LMStudio, vLLM, a local proxy, etc.):

acrawl auth other

You'll be prompted for a base URL and an optional API key. Examples:

This creates a credentials.json entry with auth_method: "api_key" and your base_url. Leave the API key blank if your server doesn't require one.

Models use the provider/model-id format: anthropic/claude-sonnet-4-6, openai/gpt-4o, amazon-bedrock/anthropic.claude-sonnet-4-6-20250514-v1:0, etc.

Providers With Non-Standard Auth

GitHub Copilot, device code flow:

acrawl auth copilot

1. acrawl prints a URL (https://github.com/login/device) and an 8-character user code, and attempts to open your browser automatically.
2. Paste the code at the GitHub page and authorize the app.
3. acrawl polls GitHub until authorization completes, then exchanges the GitHub token for a short-lived Copilot API token, which is stored in credentials.json.

No API key is needed, the entire flow is interactive. If authorization succeeds, credentials are stored automatically.

GitLab Duo, API key:

acrawl auth gitlab

Prompts for a GitLab Personal Access Token (PAT). Paste your token and press Enter. No browser redirect required. Note: GitLab Duo does not support tool calling.

Amazon Bedrock, AWS credentials:

acrawl auth amazon-bedrock

Prompts for AWS Access Key ID, AWS Secret Access Key, and region (default: us-east-1). These are stored directly in credentials.json and used to sign requests with AWS SigV4.

Azure OpenAI:

acrawl auth azure

Prompts for Resource Name (e.g. myresource), Deployment Name (e.g. gpt-4o), and API key.

Interactive TUI

The default interface is a full terminal UI with:

• Markdown rendering with syntax highlighting and streaming output
• Slash command overlay — type / to see all commands with Tab completion
• Model picker — /model opens a searchable list grouped by provider category
• Auth modal — /auth walks through provider setup interactively
• Session header — shows current model, session ID, cost, and context usage in real time
• Debug mode — /debug toggles raw tool call input/output in the transcript
• Reasoning effort — Ctrl+T cycles through high/medium/low for reasoning models (o3, o4-mini)

Keybindings:

Running acrawl without a TTY on stdout (e.g. piped or redirected) exits with an error pointing at acrawl prompt for one-shot use and acrawl --resume for session maintenance.

Session Management

• Auto-save — sessions are saved automatically on exit.
• Resume — --resume session.json reloads a conversation. Resume-safe slash commands (/status, /compact, /cost, /config, /version, /export, /help, /clear) can be appended to the command line.
• Export — /export [file] writes a human-readable markdown transcript.
• Auto-compaction, when cumulative input tokens exceed the threshold (default 200K), acrawl compacts the session:
  • Preserved verbatim: the most recent ~80K tokens of messages (always at least 2 messages), with tool call pairs kept intact, no ToolResult is ever left without its matching ToolUse.
  • Preserved as metadata: message counts, deduplicated tool list, last 3 user requests (160 chars each), pending work items (inferred from messages containing "todo", "next", "pending", "follow up", "remaining"), up to 10 key URLs, and the most recent non-empty message as "current work".
  • Tool output pruning: tool outputs older than the innermost 40K-token window are truncated to 2,000 chars with a [… output truncated from N chars] marker before the preserved window is calculated.
  • Summary generation: template-based by default (no LLM call). Opt in to LLM summarization via compaction_llm_summarization: true in settings.json, this sends the removed messages to the model and uses its output as the summary, with a fallback to the template if the LLM fails.
  • Continuation prompt: the compacted session prepends a system message instructing the agent to resume directly without recapping or asking questions.
  • Browser state is unaffected, compaction only modifies message history; the current browser tab, URL, cookies, and page state are unchanged.
• Multiple sessions — /session list to browse, /session switch <id> to switch.

Tool Allowlist

Use --allowedTools to restrict which tools the agent can invoke (comma-separated, flag is repeatable):

acrawl prompt "scrape titles" --allowedTools navigate,read_content,screenshot

Omit --allowedTools to allow all 21 tools. Useful for locking down a crawl to read-only tools or excluding fork/wait_for_subagents when sub-agent parallelism is not desired.

MCP Extensibility

acrawl supports Model Context Protocol servers as a client, allowing you to extend the agent with custom tools. MCP tools are namespaced as server_name__tool_name and available alongside the built-in 21.

Supported transports: stdio, SSE, HTTP, WebSocket.

MCP Server (expose acrawl as a tool)

acrawl mcp starts a built-in MCP server that exposes acrawl's browser automation capabilities to external agents like Claude Code, Cursor, VS Code, Zed, JetBrains, TRAE, Gemini CLI, or any MCP-compatible client.

The server provides 18 tools in two modes:

Direct browser tools (17) — fine-grained control for clients that orchestrate themselves: navigate, click, click_at, fill_form, page_map, read_content, screenshot, go_back, scroll, wait, select_option, execute_js, hover, press_key, switch_tab, list_resources, save_file

Autonomous agent (1) — delegate a full crawl task:

• run_goal — Execute a high-level crawl goal autonomously. The agent plans, navigates, and extracts data using its own LLM loop. Requires ~/.acrawl/credentials.json configured with a model.

Transport: stdio only (no SSE / HTTP / WebSocket in this release).

Quick install

acrawl mcp install

Interactive installer that auto-detects your IDEs, lets you toggle which to configure (Space to select, Enter to confirm), and writes the correct config for each. Supports global (user-level) and project-level scopes.

Supported clients: Claude Code, Claude Desktop, Cursor, Windsurf, VS Code (Copilot), OpenCode, Zed, TRAE, JetBrains IDEs, Gemini CLI, Qwen Code, Codex CLI, Hermes, OpenClaw, Goose, Crush, Aider.

Manual configuration

If you prefer to configure manually, add this to your IDE's MCP config file:

{
  "mcpServers": {
    "acrawl": {
      "command": "acrawl",
      "args": ["mcp"]
    }
  }
}

{
  "servers": {
    "acrawl": {
      "command": "acrawl",
      "args": ["mcp"]
    }
  }
}

{
  "mcp": {
    "acrawl": {
      "type": "local",
      "command": ["acrawl", "mcp"]
    }
  }
}

{
  "context_servers": {
    "acrawl": {
      "command": {
        "path": "acrawl",
        "args": ["mcp"],
        "env": {}
      },
      "settings": {}
    }
  }
}

Or via the Claude Code CLI directly:

claude mcp add acrawl -- acrawl mcp

The browser tools share a persistent session across calls. run_goal creates its own isolated agent and browser.

Requirements: The 17 browser tools work without any configuration. run_goal requires ~/.acrawl/credentials.json (via acrawl auth) for its internal LLM.

Usage

acrawl [OPTIONS] [COMMAND]

Commands:
  prompt <text>      Run a single goal non-interactively
  mcp                Start MCP server (stdio transport)
  mcp install        Install MCP config into your IDEs interactively
  auth [provider]    Configure provider credentials
  system-prompt      Print the system prompt (for debugging)

Options:
  --model MODEL            Model in provider/id format (e.g. anthropic/claude-sonnet-4-6)
  --output-format FORMAT   text | json
  --resume FILE            Resume a saved session (with optional /commands)
  --compact                Compact history on resume
  --headless[=BOOL]        Force browser headless on/off
  --no-headless, --headed  Launch browser in visible mode
  --allowedTools TOOLS     Restrict available tools (comma-separated, repeatable)
  -p TEXT                  Shorthand for prompt mode
  -V, --version            Print version

Slash Commands

Configuration

All config lives in ~/.acrawl/ (override with ACRAWL_CONFIG_HOME).

credentials.json

Managed via acrawl auth. Stores per-provider:

Azure additionally requires resource_name and deployment_name. Bedrock requires aws_access_key_id, aws_secret_access_key, and region. Vertex requires gcp_project_id and gcp_region.

settings.json

Created with defaults on first run.

All fields are optional; omitting a field uses the default. The optimization block accepts a nested object with the following fields (all default to false/0/null, safe to omit entirely):

Environment Variables

Provider-specific env vars (see provider table above) are read as fallbacks when no credentials.json entry exists.

Performance Optimizations

acrawl ships 14 vendor-derived optimizations (sourced from browser-use, Stagehand, crawl4ai, Skyvern, Spider, nanobrowser, and ZeroClaw). All are disabled by default, enable selectively via settings.json.

Example settings.json with a cost-optimized profile:

{
  "optimization": {
    "html_diff_mode": true,
    "action_caching": true,
    "page_fingerprinting": true,
    "loop_detection": true,
    "self_healing": true,
    "budget_max_session_cost_usd": 0.50,
    "budget_enforcement": "warn"
  }
}

Known Limitations

acrawl works well on most public web content, but some situations are outside what the agent can reliably handle:

How It Works

flowchart LR
    Goal([Goal\nnatural language]) --> Plan
    Plan --> Navigate --> Observe --> Act --> Extract
    Extract -->|repeat until done| Plan
    Extract --> Output([Output\nJSON / CSV])

1. The agent receives a goal and builds a multi-step plan via a 7-section system prompt covering identity, operating procedure, data integrity, constraints, error recovery, completion protocol, and parallel exploration guidance.
2. Each turn, it picks from its 21 tools based on what it observes on the page.
3. navigate hits the FetchRouter, which tries HTTP first and auto-escalates to a headless Chromium browser when JavaScript, auth redirects, or framework markers are detected.
4. The browser is driven by an embedded Node.js subprocess (the PlaywrightBridge) speaking newline-delimited JSON over stdio — uses CloakBrowser for stealth browsing, not stock Playwright. Alternatively, acrawl can drive the user's real browser via a Chrome extension (/extension command) using CDP over a local WebSocket bridge.
5. For multi-page tasks, the agent can fork child agents onto separate browser tabs, each with independent state and step budgets. wait_for_subagents collects results; cancel_subagent aborts a running child; subagent_status polls without blocking.
6. When context grows large, auto-compaction summarizes older messages while preserving recent turns, tool usage, and pending work items.
7. The agent stops when the goal is met and all sub-agents have finished, or when the step limit is reached.

Architecture

crates/
  core/         Shared types, traits, error hierarchy (acrawl-core)
  api/          25 provider clients (Anthropic, OpenAI, Gemini, DeepSeek, Bedrock, Azure, ...), SSE streaming
  browser/      PlaywrightBridge, ExtensionBridge, FetchRouter, BrowserContext, WsBridgeServer
  agent/        21 tools, agent loop, sub-agent fork/join, CrawlState
  runtime/      ConversationRuntime, config, sessions, MCP client stack, OAuth PKCE
  render/       Markdown rendering, tool output formatting, OutputSink
  mcp-server/   Built-in MCP server (JSON-RPC over stdio), IDE installer
  tui/          Ratatui terminal UI (acrawl-tui)
  ui/           Shared application layer (LiveCli, session management, tool executor, auth)
  cli/          Thin binary entry point (main.rs, self_update.rs, uninstall.rs)
  commands/     17 slash commands with resume-safety annotations

11 crates, ~40K lines of Rust, 1,097 tests.

Development

cargo build --release                                     # build
cargo test --workspace                                    # run all tests
cargo clippy --workspace --all-targets -- -D warnings     # lint (pedantic)
cargo fmt --check                                         # format check

See CONTRIBUTING.md for the full development guide.

Changelog

See CHANGELOG.md.

Security

See SECURITY.md for the security policy and how to report vulnerabilities.

License

MIT