drissionpage-mcp 0.5.9

Professional browser automation for Codex, Claude Code, and MCP clients powered by DrissionPage

DrissionPage MCP Server

Professional browser automation for Codex, Claude Code, and MCP clients powered by DrissionPage

_{A new interaction layer for multimodal AI — vision coordinates in, natural pointer action chains out.}

Official Repositories: GitHub | GitCode

English Version | 中文版本

🖱️ Vision-Guided Human–Computer Interaction

DrissionPage MCP 0.5.9 provides a complete first-stage interaction layer for multimodal AI: it can turn a vision model's viewport coordinate into a complete, physically plausible pointer action chain—not just a raw teleport-and-click.

One MCP call connects visual understanding to real browser interaction. The model identifies where to act; DrissionPage MCP handles how the pointer gets there and performs the click.

Screenshot / page observation
        ↓
Multimodal model identifies viewport coordinates
        ↓
page_click_xy(profile="natural")
        ↓
Cubic Bézier motion → reaction pause → press → hold → release
        ↓
Observe and verify the resulting page state

What makes this interaction layer different?

• Natural pointer dynamics: 20–35 cubic Bézier movement steps instead of coordinate teleportation.
• Human-like timing: 8–25ms point intervals, smoothstep ease-in-out, and a 100–300ms reaction pause after arrival.
• Physical click semantics: 50–120ms press duration with correct Chromium CDP button state for left, right, and middle clicks.
• Controlled micro-motion: bounded ±0.5 CSS-pixel intermediate jitter while the final point remains exact.
• Failure-safe execution: a pressed button is always released if the action chain is interrupted.
• Model-readable evidence: results include the chosen profile, start and target coordinates, step count, reaction delay, hold duration, and planned duration.

This makes vision-guided operation practical for canvas controls, visual editors, maps, charts, non-semantic widgets, responsive interfaces, and other surfaces where selectors or accessibility metadata are incomplete. Structured DOM automation remains the preferred path when reliable selectors are available; the vision interaction layer expands what an MCP agent can operate when they are not.

{
  "x": 442,
  "y": 369,
  "start_x": 100,
  "start_y": 100,
  "profile": "natural",
  "button": "left",
  "element": "visually identified control"
}

Designed for legitimate UI automation, testing, accessibility workflows, and technical research. Security or anti-automation challenge completion is not offered as a guaranteed supported capability.

🧭 Client Setup Navigation

• Vision-guided human–computer interaction
• Install + screenshot walkthrough
• Codex CLI/IDE quick setup
• Codex CLI/IDE integration example
• Claude Code setup
• Cursor setup
• Claude Desktop setup
• Troubleshooting

---

🚀 What is DrissionPage MCP?

DrissionPage MCP Server is a local Model Context Protocol (MCP) server that brings DrissionPage browser automation tools to Codex CLI/IDE, Claude Code, Claude Desktop, and other MCP clients.

Structured, deterministic automation remains the default through 54 tools plus MCP Resources/Prompts. When selectors or accessibility metadata are insufficient, 0.5.9 also provides an optional vision-guided human–computer interaction layer that converts viewport coordinates into natural Chromium pointer action chains, powered by DrissionPage.

🌟 Why Choose DrissionPage MCP?

• Structured-First, Vision-Ready: Uses DOM structure when available and multimodal coordinates when visual interaction is the better tool
• Deterministic: Reliable element selection with CSS/XPath normalization for LLM-friendly selectors
• Vision-Ready Interaction: Converts multimodal model coordinates into natural pointer movement and physically timed clicks
• Fast & Lightweight: Built on DrissionPage's efficient engine with minimal overhead
• Type-Safe: Full type hints and Pydantic validation for all tools
• Open-source Friendly: Includes compatibility notes, troubleshooting, and CI checks for maintainable contributions
• Easy Integration: Simple pip install + Codex TOML or MCP JSON configuration

✅ Quality and Real-World Validation

DrissionPage MCP is backed by a strict regression suite and browser-backed scenario checks:

• Strict automated tests: unit, protocol, schema snapshot, response-contract, resource/prompt, release-metadata, security-policy, browser-integration, and coverage checks run in CI.
• 95% coverage floor: CI enforces the current 95% coverage threshold and uploads coverage reports.
• Real browser verification: Chrome/Chromium-backed integration tests exercise the same MCP tools exposed to clients.
• Scenario validation: the playground MCP Lab covers realistic forms, commerce pages, social feeds, timelines, dynamic waits, iframe cases, and recovery paths without depending on public demo websites.

---

⚡ First Success Path

# Install from PyPI
python -m pip install -U drissionpage-mcp

# Verify package and environment
drissionpage-mcp --version
drissionpage-mcp doctor

Then add the Codex or MCP client configuration below and restart your client.

---

📦 Setup in Codex CLI/IDE (30 seconds)

Codex supports local stdio MCP servers through config.toml; the CLI and IDE extension share the same MCP configuration.

1. Edit Codex configuration:

   • User-level: ~/.codex/config.toml
   • Project-level: .codex/config.toml inside a trusted project

2. Add this configuration:

   [mcp_servers.drissionpage]
   command = "drissionpage-mcp"
   startup_timeout_sec = 20
   tool_timeout_sec = 60
   

3. Restart Codex. In the TUI, run /mcp; from a shell, run codex mcp list.

For Claude Code, Claude Desktop, and other JSON-based MCP clients, see Integration Examples.

---

🎯 Quick Examples

Navigate and Screenshot

"Visit https://example.com and take a screenshot for me"

Search and Extract

"Go to Wikipedia, search for Python, and get the first paragraph"

Form Automation

"Fill out the form at https://httpbin.org/forms/post and submit it"

Data Scraping

"Get the top 10 news headlines from news.ycombinator.com"

---

🛠️ 54 Powerful Tools + MCP Resources/Prompts

🌐 Navigation (4 tools)

• page_navigate - Navigate to any URL; optionally open it in a new tab with new_tab or return an observe change summary
• page_go_back / page_go_forward - Browser history
• page_refresh - Reload current page

🗂️ Tab Operations (3 tools)

• tab_list - List open browser tabs with stable MCP tab IDs
• tab_switch - Switch to a tab returned by tab_list
• tab_close - Close one tab without closing the whole browser

🎯 Element Interaction & Extraction (13 tools)

• element_find - Find one element by CSS selector or XPath; bare selectors like h1 are treated as CSS
• element_find_all - Extract bounded repeated elements with text, attributes, and recommended selectors
• element_click - Click any element
• element_type - Input text into elements
• element_upload_file - Upload files from DP_MCP_UPLOAD_ROOT to input[type=file]
• element_scroll_into_view - Bring an element into the viewport before acting
• element_hover - Hover an element to trigger menu/tooltip states
• element_select - Select an option by value, text, or index
• element_check - Check or uncheck checkbox/radio controls
• element_get_text - Get element or page text
• element_get_attribute - Get an HTML attribute
• element_get_property - Get a live DOM property such as an input value
• element_get_html - Get element or page HTML

🧾 Form Operations (1 tool)

• form_inspect - Inspect forms and controls with labels, selectors, requirements, options, and safe optional values

📸 Page Operations (13 tools)

• page_screenshot - Capture an inline full-page or viewport screenshot
• page_screenshot_save - Save a screenshot under DP_MCP_SCREENSHOT_ROOT
• page_snapshot - Return a bounded page outline with headings, links, buttons, inputs, forms, and selector recommendations
• page_observe - Return a compact page fingerprint with URL, title, counts, visible text samples, active element, and recent console summary
• page_evaluate - Run bounded JavaScript in the current page and return a JSON-safe result
• page_scroll - Scroll the page by direction or to a position
• keyboard_press - Send keys to the active element/page
• page_resize - Adjust browser window
• page_pointer_move - Move to vision-model viewport coordinates with a natural Bézier path without clicking
• page_pointer_drag - Perform one failure-safe natural drag between viewport coordinates
• page_click_xy - Convert vision-model viewport coordinates into natural Bézier pointer movement and physically timed clicks
• page_close - Close browser
• page_get_url - Get current URL

🧱 Frame / Shadow DOM (5 tools)

• frame_list - List iframe/frame contexts without changing global frame state
• frame_snapshot - Inspect a selected iframe with bounded outline data
• frame_find - Find an element inside a selected iframe
• shadow_find - Find one element inside an open shadow root
• shadow_find_all - Extract repeated elements inside an open shadow root

🍪 Cookies & Storage (4 tools)

• browser_cookies_get - Read normalized cookies with values redacted by default
• storage_get - Read localStorage/sessionStorage by key or as a map
• storage_set - Set one localStorage/sessionStorage item without echoing the value
• storage_clear - Clear one storage key or an entire storage area

🧪 Debug / Observability (1 tool)

• page_console_logs - Read bounded browser console messages with level filtering, cursor pagination, and limits

⏱️ Wait Operations (4 tools)

• wait_for_element - Wait for element to appear (with timeout)
• wait_for_url - Wait until the current URL contains text
• wait_until - Wait for observable conditions such as clickable, hidden, stable, text, or URL matches
• wait_time - Delay execution

🧩 MCP Resources and Prompts

• Resources: drissionpage://session/summary, drissionpage://session/history, drissionpage://session/state, drissionpage://session/config, drissionpage://guide/model-usage, drissionpage://page/current, drissionpage://tools/catalog, drissionpage://policy/summary
• Prompts: drissionpage_mcp_usage_playbook, browser_navigate_and_summarize, browser_extract_structured_data, browser_fill_form_safely, browser_vision_guided_interaction, browser_debug_page_issue

---

📚 Documentation

Guide	Description
README.md	Installation, tools, and architecture
docs/compatibility.md	Supported Python, DrissionPage, MCP, and browser versions
docs/tool-contract.md	Public MCP tool names, inputs, annotations, and response shape
docs/troubleshooting.md	Doctor command, browser startup, and client setup fixes
CHANGELOG.md	Release notes

---

🏗️ Architecture

Built with clean, modular design:

DrissionMCP/
├── drissionpage_mcp/
│   ├── cli.py              # Entry point
│   ├── server.py           # MCP server
│   ├── context.py          # Browser management
│   ├── response.py         # Response formatting
│   ├── tab.py              # Page operations
│   └── tools/              # 46 automation, tab/frame/shadow, page-understanding, form, debug, and session-state tools
├── tests/                  # Unit tests
└── playground/             # MCP Lab business-scenario playground

Key Principles:

• ✅ Type-safe Pydantic models for all tools
• ✅ Async/await throughout
• ✅ Clean separation of concerns
• ✅ Comprehensive error handling
• ✅ Unit and protocol test coverage for core tool registration/response behavior

---

🔧 Configuration

Codex CLI / IDE (Recommended)

[mcp_servers.drissionpage]
command = "drissionpage-mcp"
startup_timeout_sec = 20
tool_timeout_sec = 60

# Optional browser/runtime environment variables:
# [mcp_servers.drissionpage.env]
# CHROME_PATH = "/custom/path/to/chrome"
# DP_HEADLESS = "1"

You can also add it with the Codex CLI:

codex mcp add drissionpage -- drissionpage-mcp

If Codex/Cursor/Claude Desktop is launched from a GUI and cannot see your shell PATH or virtualenv, use the absolute Python executable instead:

[mcp_servers.drissionpage]
command = "/absolute/path/to/python"
args = ["-m", "drissionpage_mcp.cli"]
startup_timeout_sec = 20
tool_timeout_sec = 60

JSON MCP Clients

{
  "mcpServers": {
    "drissionpage": {
      "command": "drissionpage-mcp"
    }
  }
}

Advanced JSON Setup

{
  "mcpServers": {
    "drissionpage": {
      "command": "drissionpage-mcp",
      "args": ["--log-level", "DEBUG"],
      "env": {
        "CHROME_PATH": "/custom/path/to/chrome"
      }
    }
  }
}

Absolute-Python fallback for GUI clients:

{
  "mcpServers": {
    "drissionpage": {
      "command": "/absolute/path/to/python",
      "args": ["-m", "drissionpage_mcp.cli"],
      "env": {
        "CHROME_PATH": "/custom/path/to/chrome",
        "DP_HEADLESS": "1"
      }
    }
  }
}

---

📋 Requirements

• Python 3.10+ (3.11+ recommended)
• Chrome or Chromium browser
• Any MCP-compatible client: Codex CLI/IDE, Claude Code, Claude Desktop, Cursor, VS Code, etc.

---

🧪 Testing

Verify Installation

# Environment diagnostics; add --launch-browser for a browser startup check
drissionpage-mcp doctor
drissionpage-mcp doctor --launch-browser

# Source checkout tests
python -m pip install -e ".[dev]"
python -m pytest tests/

# Coverage report (CI enforces the current 95% floor and uploads coverage.xml)
python -m pytest tests/ --cov=drissionpage_mcp --cov-report=term-missing --cov-report=xml

# Browser-backed MCP Lab scenario checks
DP_HEADLESS=1 python playground/run_mcp_lab.py --all --json

GitHub Actions runs lint, unit, protocol, package, browser integration, and coverage jobs. Codecov is configured through codecov.yml and the CI workflow.

Try It Out

# No-browser MCP registry check
python playground/run_mcp_lab.py --case registry

# Local deterministic site check
python playground/run_mcp_lab.py --case site

# Browser-backed form inspection scenario
DP_HEADLESS=1 python playground/run_mcp_lab.py --case form-inspect

---

🚀 Use Cases

✅ Automated Testing - Test web applications ✅ Data Scraping - Extract structured data from websites ✅ Form Automation - Fill and submit forms ✅ Monitoring - Check for updates or changes ✅ Screenshot Verification - Capture and verify page state ✅ Content Analysis - Analyze web content programmatically

---

🐛 Troubleshooting

Tools Not Loading?

drissionpage-mcp --version

Should output the installed package version, for example drissionpage-mcp 0.5.9.

Browser Issues?

# Check browser installation
which google-chrome    # Linux
which chromium         # macOS

Codex / MCP Client Not Finding Server?

• Codex: run codex mcp list; in the TUI, run /mcp
• JSON clients: verify config file path and JSON syntax
• Restart Codex or your MCP client after changes
• Check logs: drissionpage-mcp --log-level DEBUG

See docs/troubleshooting.md for the complete troubleshooting guide.

---

📊 Project Status

Component	Status
Core Features	✅ Complete
Testing	✅ Strict unit/protocol/schema checks plus browser-backed scenarios
Documentation	✅ Setup, compatibility, troubleshooting, and public tool contracts
Package	✅ PyPI metadata and build checks
Status	🟡 Beta; real browser behavior depends on local Chrome/Chromium and target sites

Version: 0.5.9 | License: Apache 2.0 | Maintained: ✅ Active

---

🗺️ Roadmap

Current (v0.5.9)

• 52 core automation, tab/frame/shadow, page-understanding, form-inspection, workflow, network-listener, session-state, and console-observability tools with removed alias surface
• stdio MCP server integration
• Doctor diagnostics for local setup
• Stable JSON mirror, structuredContent, and typed per-tool MCP outputSchema
• Structured recovery hints in error.details.hints for common failures
• Balanced page_snapshot output so link-heavy pages still expose controls and forms
• form_inspect read-only form inventory with labels, selectors, requirements, options, and safe optional values
• Tab management with tab_list, tab_switch, tab_close, and page_navigate(new_tab=true)
• Observable actions with page_observe, page_evaluate, wait_until, and optional observe=true changes on navigation, click, and type
• Console observability with page_console_logs, console summary in page_observe, and console change fields in observe=true
• Workflow helpers with browser_open_and_snapshot, browser_extract_links, and form_fill_preview
• Network listener beta with network_listen_start, network_listen_wait, and network_listen_stop for HTTP/XHR/Fetch observation
• Natural page_pointer_move, page_pointer_drag, and page_click_xy action chains with cubic Bézier motion, smoothstep easing, bounded jitter, reaction delay, and realistic button hold time
• File upload, scrolling, hover, select/check, keyboard, iframe, shadow DOM, cookie, and storage tools for DrissionPage 4.x
• Chrome sandbox remains enabled by default; DP_NO_SANDBOX=1 is reserved for restricted container/root environments
• Redacted session history resource and response size metadata for bounded outputs
• Opt-in local safety policy for navigation and screenshot paths
• Resources, prompts, eval harness, compatibility, and troubleshooting documentation
• PyPI distribution

Future (v0.6+)

• Promote workflow/network beta contracts toward 0.6.0 after field testing
• Optional session persistence beyond redacted state summaries
• Proxy support
• Network interception

---

📖 Integration Examples

Codex CLI / IDE

[mcp_servers.drissionpage]
command = "drissionpage-mcp"
startup_timeout_sec = 20
tool_timeout_sec = 60

Verify with:

codex mcp list

Claude Code

{
  "mcpServers": {
    "drissionpage": {
      "command": "drissionpage-mcp"
    }
  }
}

Config file: ~/.config/claude-code/mcp_settings.json (macOS/Linux) or %APPDATA%\claude-code\mcp_settings.json (Windows).

Cursor

{
  "mcpServers": {
    "drissionpage": {
      "command": "drissionpage-mcp"
    }
  }
}

Config file: ~/.cursor/mcp.json (global) or .cursor/mcp.json (project). You can also add it from Cursor Settings → Tools & MCPs → New MCP Server.

Claude Desktop

{
  "mcpServers": {
    "drissionpage": {
      "command": "drissionpage-mcp"
    }
  }
}

Once connected, the tools load automatically:

---

🤝 Contributing

Contributions are welcome!

1. Fork the repository
2. Create a feature branch
3. Make focused changes
4. Run the relevant checks
5. Submit a pull request

See CONTRIBUTING.md for setup, validation, and compatibility expectations.

---

🔒 Security

• Runs locally in your environment
• Uses a local browser that may have access to authenticated sessions, cookies, downloads, and page content
• Can open and interact with any site reachable from the local machine
• Does not require external API credentials

Best Practices:

• Use a dedicated browser profile for sensitive workflows
• Review MCP client prompts before allowing actions on authenticated or production systems
• Respect website terms of service, robots.txt, and rate limits
• See SECURITY.md for reporting and safe-usage guidance

---

📄 License

Licensed under Apache License 2.0 - see LICENSE

---

📈 Statistics

---

🌟 Show Your Support

If you find this project useful, please consider:

• ⭐ Starring on GitHub
• 📤 Sharing with your network
• 💬 Leaving feedback or suggestions
• 🐛 Reporting issues to help improve

---

Made with ❤️ by Wukunyun

Ready to automate your workflows? Install now: python -m pip install -U drissionpage-mcp

---

🆕 Latest Version: v0.5.9

Released on 2026-07-12. This release completes the first vision-guided pointer action surface and improves how Claude/GPT select, execute, verify, and recover MCP interactions:

• Added page_pointer_move for natural hover, reveal, canvas positioning, and visual inspection without clicking.
• Added page_pointer_drag as one failure-safe start-to-end drag call: natural approach, button press, held-button Bézier movement, release delay, and guaranteed cleanup release on failure.
• page_click_xy, move, and drag share one motion planner, smoothstep easing, bounded jitter, pointer state, immutable action sequence, and natural / precise / direct profiles; natural clicks retain the 100–300 ms after arrival reaction delay and 50–120 ms button hold.
• MCP initialization, drissionpage://guide/model-usage, drissionpage://tools/catalog, and browser_vision_guided_interaction now teach selector-first routing, viewport coordinate conversion, move-vs-click-vs-drag intent selection, bounded verification, and stale-coordinate recovery.
• The compact tool catalog now exposes required inputs, optional inputs, defaults, enums, bounds, types, and field descriptions; tools/list remains the complete JSON Schema source.
• Pointer move, click, and drag return typed motion evidence, and drag never exposes persistent mouse_down / mouse_up state across MCP calls.
• The public registry now contains 54 tools while preserving the strict JSON_RESULT, structuredContent, typed outputSchema, DrissionPage 4.x, and Chrome sandbox contracts.