ScreenHand

Native desktop control for MCP agents.

An open-source MCP server for macOS and Windows that gives Claude, Cursor, Codex CLI, and OpenClaw fast desktop control via Accessibility/UI Automation, OCR, and Chrome CDP.

Website | Quick Start | Why ScreenHand | Tools | FAQ

Why ScreenHand?

• ~50ms native UI actions via Accessibility APIs and Windows UI Automation
• 0 extra AI calls for native clicks, typing, and UI element lookup
• 70+ tools across desktop apps, browser automation, OCR, memory, sessions, jobs, and playbooks
• macOS + Windows behind the same MCP interface
• Multi-agent safe — session leases prevent conflicts between Claude, Cursor, and Codex
• Background worker — queue jobs and let the daemon process them continuously

What is ScreenHand?

ScreenHand is a desktop automation bridge for AI. It connects AI assistants like Claude to your operating system so they can:

• See your screen via screenshots and OCR
• Read UI elements via Accessibility APIs (macOS) or UI Automation (Windows)
• Click buttons, menus, and links
• Type text into any input field
• Control Chrome tabs via DevTools Protocol
• Run AppleScript commands (macOS)
• Queue & execute multi-step jobs via playbooks with a background worker daemon
• Coordinate multiple AI agents with session leases and stall detection

It works as an MCP (Model Context Protocol) server, meaning any MCP-compatible AI client can use it out of the box.

| Problem | ScreenHand Solution | |---|---| | AI can't see your screen | Screenshots + OCR return all visible text | | AI can't click UI elements | Accessibility API finds and clicks elements in ~50ms | | AI can't control browsers | Chrome DevTools Protocol gives full page control | | AI can't automate workflows | 70+ tools for cross-app automation | | Only works on one OS | Native bridges for both macOS and Windows | | Multiple agents conflict | Session leases with heartbeat and stall detection | | Jobs need manual triggering | Worker daemon processes the queue continuously |

Quick Start

Source install (recommended today)

ScreenHand currently builds a native bridge locally for Accessibility/UI Automation, so the fastest reliable setup is still from source:

git clone https://github.com/manushi4/screenhand.git
cd screenhand
npm install
npm run build:native   # macOS — builds Swift bridge
# npm run build:native:windows   # Windows — builds .NET bridge

Then connect ScreenHand to your AI client.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "screenhand": {
      "command": "npx",
      "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
    }
  }
}

Claude Code

Add to your project .mcp.json or ~/.claude/settings.json:

{
  "mcpServers": {
    "screenhand": {
      "command": "npx",
      "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project (or ~/.cursor/mcp.json for global):

{
  "mcpServers": {
    "screenhand": {
      "command": "npx",
      "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
    }
  }
}

OpenAI Codex CLI

Add to ~/.codex/config.toml:

[mcp.screenhand]
command = "npx"
args = ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
transport = "stdio"

OpenClaw

Add to your openclaw.json:

{
  "mcpServers": {
    "screenhand": {
      "command": "npx",
      "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
    }
  }
}

Why? OpenClaw's built-in desktop control sends a screenshot to an LLM for every click (~3-5s, costs an API call). ScreenHand uses native Accessibility APIs — press('Send') runs in ~50ms with zero AI calls. See the full integration guide.

Any MCP Client

ScreenHand is a standard MCP server over stdio. It works with any MCP-compatible client — just point it at mcp-desktop.ts.

Replace /path/to/screenhand with the actual path where you cloned the repo.

Tools

ScreenHand exposes 70+ tools organized by category.

See the Screen

| Tool | What it does | Speed | |------|-------------|-------| | screenshot | Full screenshot + OCR — returns all visible text | ~600ms | | screenshot_file | Screenshot saved to file (for viewing the image) | ~400ms | | ocr | OCR with element positions and bounding boxes | ~600ms |

Control Any App (Accessibility / UI Automation)

| Tool | What it does | Speed | |------|-------------|-------| | apps | List running apps with bundle IDs and PIDs | ~10ms | | windows | List visible windows with positions and sizes | ~10ms | | focus | Bring an app to the front | ~10ms | | launch | Launch an app by bundle ID or name | ~1s | | ui_tree | Full UI element tree — instant, no OCR needed | ~50ms | | ui_find | Find a UI element by text or title | ~50ms | | ui_press | Click a UI element by its title | ~50ms | | ui_set_value | Set value of a text field, slider, etc. | ~50ms | | menu_click | Click a menu bar item by path | ~100ms |

Keyboard and Mouse

| Tool | What it does | |------|-------------| | click | Click at screen coordinates | | click_text | Find text via OCR and click it (fallback) | | type_text | Type text via keyboard | | key | Key combo (e.g. cmd+s, ctrl+shift+n) | | drag | Drag from point A to B | | scroll | Scroll at a position |

Chrome Browser (CDP)

| Tool | What it does | |------|-------------| | browser_tabs | List all open Chrome tabs | | browser_open | Open URL in new tab | | browser_navigate | Navigate active tab to URL | | browser_js | Run JavaScript in a tab | | browser_dom | Query DOM with CSS selectors | | browser_click | Click element by CSS selector (uses CDP mouse events) | | browser_type | Type into an input field (uses CDP keyboard events, React-compatible) | | browser_wait | Wait for a page condition | | browser_page_info | Get page title, URL, and content |

Anti-Detection & Stealth (CDP)

Tools for interacting with sites that have bot detection (Instagram, LinkedIn, etc.):

| Tool | What it does | |------|-------------| | browser_stealth | Inject anti-detection patches (hides webdriver flag, fakes plugins/languages) | | browser_fill_form | Human-like typing with random delays via CDP keyboard events | | browser_human_click | Realistic mouse event sequence (mouseMoved → mousePressed → mouseReleased) |

Tip: Call browser_stealth once after navigating to a protected site. Then use browser_fill_form and browser_human_click for interactions. The regular browser_type and browser_click also use CDP Input events now.

Smart Execution (fallback chain)

Tools that automatically choose the best method (Accessibility → CDP → OCR → coordinates):

| Tool | What it does | |------|-------------| | execution_plan | Generate an execution plan for a task | | click_with_fallback | Click using the best available method | | type_with_fallback | Type using the best available method | | read_with_fallback | Read content using the best available method | | locate_with_fallback | Find an element using the best available method | | select_with_fallback | Select an option using the best available method | | scroll_with_fallback | Scroll using the best available method | | wait_for_state | Wait for a UI state using the best available method |

Platform Playbooks (lazy-loaded)

Pre-built automation knowledge for specific platforms — selectors, URLs, flows, and error solutions.

| Tool | What it does | |------|-------------| | platform_guide | Get automation guide for a platform (selectors, URLs, flows, errors+solutions) | | export_playbook | Auto-generate a playbook from your session. Share it to help others. |

platform_guide({ platform: "devpost", section: "errors" })   # Just errors + solutions
platform_guide({ platform: "devpost", section: "selectors" }) # All CSS selectors
platform_guide({ platform: "devpost", section: "flows" })     # Step-by-step workflows
platform_guide({ platform: "devpost" })                       # Full playbook

Contributing playbooks: After automating any site, run:

export_playbook({ platform: "twitter", domain: "twitter.com" })

This auto-extracts URLs, selectors, errors+solutions from your session and saves a ready-to-share playbooks/twitter.json.

Available platforms: devpost. Add more by running export_playbook or creating JSON files in playbooks/.

Zero performance cost — files only read when platform_guide is called.

AppleScript (macOS only)

| Tool | What it does | |------|-------------| | applescript | Run any AppleScript command |

Memory (Learning) — zero-config, zero-latency

ScreenHand gets smarter every time you use it — no manual setup needed.

What happens automatically:

• Every tool call is logged (async, non-blocking — adds ~0ms to response time)
• After 3+ consecutive successes, the winning sequence is saved as a reusable strategy
• Known error patterns are tracked with resolutions (e.g. "launch times out → use focus() instead")
• On every tool call, the response includes auto-recall hints:
  • Error warnings if the tool has failed before
  • Next-step suggestions if you're mid-way through a known strategy

Predefined seed strategies:

• Ships with 12 common macOS workflows (Photo Booth, Chrome navigation, copy/paste, Finder, export PDF, etc.)
• Loaded automatically on first boot — the system has knowledge from day one
• Seeds are searchable via memory_recall and provide next-step hints like any learned strategy

Background web research:

• When a tool fails and no resolution exists, ScreenHand searches for a fix in the background (non-blocking)
• Uses Claude API (haiku, if ANTHROPIC_API_KEY is set) or DuckDuckGo instant answers as fallback
• Resolutions are saved to both error cache and strategy store — zero-latency recall next time
• Completely silent and fire-and-forget — never blocks tool responses or throws errors

Fingerprint matching & feedback loop:

• Each strategy is fingerprinted by its tool sequence (e.g. apps→focus→ui_press)
• O(1) exact-match lookup when the agent follows a known sequence
• Success/failure outcomes are tracked per strategy — unreliable strategies are auto-penalized and eventually skipped
• Keyword-based fuzzy search with reliability scoring for memory_recall

Production-grade under the hood:

• All data cached in RAM at startup — lookups are ~0ms, disk is only for persistence
• Disk writes are async and buffered (100ms debounce) — never block tool calls
• Sync flush on process exit (SIGINT/SIGTERM) — no lost writes
• Per-line JSONL parsing — corrupted lines are skipped, not fatal
• LRU eviction: 500 strategies, 200 error patterns max (oldest evicted automatically)
• File locking (.lock + PID) prevents corruption from concurrent instances
• Action log auto-rotates at 10 MB
• Data lives in .screenhand/memory/ as JSONL (grep-friendly, no database)

| Tool | What it does | |------|-------------| | memory_snapshot | Get current memory state snapshot | | memory_recall | Search past strategies by task description | | memory_save | Manually save the current session as a strategy | | memory_record_error | Record an error pattern with an optional fix | | memory_record_learning | Record a verified pattern (what works/fails) | | memory_query_patterns | Search learnings by scope and method | | memory_errors | View all known error patterns and resolutions | | memory_stats | Action counts, success rates, top tools, disk usage | | memory_clear | Clear actions, strategies, errors, or all data |

Session Supervisor — multi-agent coordination

Lease-based window locking with heartbeat, stall detection, and automatic recovery. Prevents multiple AI agents from fighting over the same app window.

| Tool | What it does | |------|-------------| | session_claim | Claim exclusive control of an app window | | session_heartbeat | Keep your lease alive (call every 60s) | | session_release | Release your session lease | | supervisor_status | Active sessions, health metrics, stall detection | | supervisor_start | Start the supervisor background daemon | | supervisor_stop | Stop the supervisor daemon | | supervisor_pause | Pause supervisor monitoring | | supervisor_resume | Resume supervisor monitoring | | supervisor_install | Install supervisor as a launchd service (macOS) | | supervisor_uninstall | Uninstall supervisor launchd service | | recovery_queue_add | Add a recovery action to the supervisor's queue | | recovery_queue_list | List pending recovery actions |

The supervisor runs as a detached daemon that survives MCP/client restarts. It monitors active sessions, detects stalls, expires abandoned leases, and queues recovery actions.

Jobs & Worker Daemon

Queue multi-step automation jobs and let a background worker process them continuously. Jobs can target specific apps/windows and execute via playbook engine or free-form steps.

| Tool | What it does | |------|-------------| | job_create | Create a job with steps (optionally tied to a playbook + bundleId/windowId) | | job_status | Get the status of a job | | job_list | List jobs by state (queued, running, done, failed, blocked) | | job_transition | Transition a job to a new state | | job_step_done | Mark a job step as done | | job_step_fail | Mark a job step as failed | | job_resume | Resume a blocked/waiting job | | job_dequeue | Dequeue the next queued job | | job_remove | Remove a job | | job_run | Execute a single queued job through the runner | | job_run_all | Process all queued jobs sequentially | | worker_start | Start the background worker daemon | | worker_stop | Stop the worker daemon | | worker_status | Get worker daemon status and recent results |

Job state machine: queued → running → done | failed | blocked | waiting_human

Worker daemon features:

• Runs as a detached process — survives MCP/client restarts
• Continuously polls the job queue and executes via JobRunner
• Playbook integration — jobs with a playbookId execute through PlaybookEngine
• Focuses/validates the target bundleId/windowId before each step
• Persists status and recent results to ~/.screenhand/worker/state.json
• Single-instance enforcement via PID file
• Graceful shutdown on SIGINT/SIGTERM

# Start the worker daemon directly
npx tsx scripts/worker-daemon.ts
npx tsx scripts/worker-daemon.ts --poll 5000 --max-jobs 10

# Or via MCP tools
worker_start → worker_status → worker_stop

Architecture

┌─────────────────────────────────────────────────────┐
│   MCP Client (Claude, Cursor, Codex CLI, OpenClaw)  │
└────────────────────────┬────────────────────────────┘
                         │ stdio JSON-RPC
┌────────────────────────▼────────────────────────────┐
│               mcp-desktop.ts                         │
│          (MCP Server — 70+ tools)                    │
├───────────┬──────────┬──────────────────────────────┤
│ Native    │  Chrome  │  Memory / Supervisor / Jobs   │
│ Bridge    │  CDP     │  / Playbooks / Worker         │
└─────┬─────┴────┬─────┴──────────────────────────────┘
      │          │
┌─────▼─────┐ ┌──▼──────┐  ┌──────────────┐  ┌──────────────┐
│macos-bridge│ │ Chrome  │  │  Supervisor  │  │   Worker     │
│(Swift, AX) │ │DevTools │  │   Daemon     │  │   Daemon     │
└────────────┘ └─────────┘  └──────────────┘  └──────────────┘

Key modules

| Path | Purpose | |---|---| | mcp-desktop.ts | MCP server entrypoint — all tool definitions | | src/native/bridge-client.ts | TypeScript ↔ native bridge communication | | native/macos-bridge/ | Swift binary using Accessibility API + OCR | | native/windows-bridge/ | C# binary using UI Automation + SendInput | | src/memory/ | Persistent memory service (strategies, errors, learnings) | | src/supervisor/ | Session leases, stall detection, recovery | | src/jobs/ | Job queue, runner, worker state persistence | | src/playbook/ | Playbook engine and store | | src/runtime/ | Execution contract, accessibility adapter, fallback chain | | scripts/worker-daemon.ts | Standalone worker daemon process | | scripts/supervisor-daemon.ts | Standalone supervisor daemon process |

State files

All persistent state lives under ~/.screenhand/:

~/.screenhand/
├── memory/        # strategies, errors, learnings (JSONL)
├── supervisor/    # supervisor daemon state
├── locks/         # session lease files
├── jobs/          # job queue persistence
├── worker/        # worker daemon state, PID, logs
└── playbooks/     # saved playbook definitions

How It Works

ScreenHand has three layers:

AI Client (Claude, Cursor, etc.)
    ↓ MCP protocol (stdio)
ScreenHand MCP Server (TypeScript)
    ↓ JSON-RPC (stdio)
Native Bridge (Swift on macOS / C# on Windows)
    ↓ Platform APIs
Operating System (Accessibility, CoreGraphics, UI Automation, SendInput)

1. Native bridge — talks directly to OS-level APIs:
   • macOS: Swift binary using Accessibility APIs, CoreGraphics, and Vision framework (OCR)
   • Windows: C# (.NET 8) binary using UI Automation, SendInput, GDI+, and Windows.Media.Ocr
2. TypeScript MCP server — routes tools to the correct bridge, handles Chrome CDP, manages sessions, runs jobs
3. MCP protocol — standard Model Context Protocol so any AI client can connect

The native bridge is auto-selected based on your OS. Both bridges speak the same JSON-RPC protocol, so all tools work identically on both platforms.

Use Cases

App Debugging

Claude reads UI trees, clicks through flows, and checks element states — faster than clicking around yourself.

Design Inspection

Screenshots + OCR to read exactly what's on screen. ui_tree shows component structure like React DevTools but for any native app.

Browser Automation

Fill forms, scrape data, run JavaScript, navigate pages — all through Chrome DevTools Protocol.

Cross-App Workflows

Read from one app, paste into another, chain actions across your whole desktop. Example: extract data from a spreadsheet, search it in Chrome, paste results into Notes.

Multi-Agent Coordination

Run Claude, Cursor, and Codex simultaneously — each claims its own app window via session leases. The supervisor detects stalls and recovers.

Background Job Processing

Queue automation jobs with job_create, start the worker daemon with worker_start, and let it process tasks continuously — even after you close your AI client.

UI Testing

Click buttons, verify text appears, catch visual regressions — all driven by AI.

Requirements

macOS

• macOS 12+
• Node.js 18+
• Accessibility permissions: System Settings > Privacy & Security > Accessibility > enable your terminal
• Chrome with --remote-debugging-port=9222 (only for browser tools)

Windows

• Windows 10 (1809+)
• Node.js 18+
• .NET 8 SDK
• No special permissions needed — UI Automation works without admin
• Chrome with --remote-debugging-port=9222 (only for browser tools)
• Build: npm run build:native:windows

Skills (Slash Commands)

ScreenHand ships with Claude Code slash commands:

• /screenshot — capture your screen and describe what's visible
• /debug-ui — inspect the UI tree of any app
• /automate — describe a task and Claude does it

Install globally so they work in any project:

./install-skills.sh

Development

npm run dev               # Run MCP server with tsx (hot reload)
npm run check             # type-check (covers all entry files)
npm test                  # run test suite
npm run build             # compile TypeScript
npm run build:native      # build Swift bridge (macOS)
npm run build:native:windows  # build .NET bridge (Windows)

FAQ

What is ScreenHand?

ScreenHand is an open-source MCP server that gives AI assistants like Claude the ability to see and control your desktop. It provides 70+ tools for screenshots, UI inspection, clicking, typing, browser automation, session management, job queuing, and playbook execution on both macOS and Windows.

How does ScreenHand differ from Anthropic's Computer Use?

Anthropic's Computer Use is a cloud-based feature built into Claude. ScreenHand is an open-source, local-first tool that runs entirely on your machine with no cloud dependency. It uses native OS APIs (Accessibility on macOS, UI Automation on Windows) which are faster and more reliable than screenshot-based approaches.

How does ScreenHand work with OpenClaw?

ScreenHand integrates with OpenClaw as an MCP server — giving your Claw agent native desktop speed instead of screenshot-based clicking.

| | Without ScreenHand | With ScreenHand | |---|---|---| | Clicking a button | Screenshot → LLM interprets → coordinate click (~3-5s) | press('Send') via Accessibility API (~50ms) | | Cost per action | 1 LLM API call per click | 0 LLM calls — native OS API | | Accuracy | Coordinate guessing — can miss if layout shifts | Exact element targeting by role/name |

Setup — add to your openclaw.json:

{
  "mcpServers": {
    "screenhand": {
      "command": "npx",
      "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
    }
  }
}

Your Claw keeps its visual understanding for complex tasks, but now has 70+ fast native tools for clicks, typing, menus, scrolling, browser control, and more. See the full integration guide.

Does ScreenHand work on Windows?

Yes. ScreenHand supports both macOS and Windows. On macOS it uses a Swift native bridge with Accessibility APIs. On Windows it uses a C# (.NET 8) bridge with UI Automation and SendInput.

What AI clients work with ScreenHand?

Any MCP-compatible client: Claude Desktop, Claude Code, Cursor, Windsurf, OpenAI Codex CLI, and any other tool that supports the Model Context Protocol.

Does ScreenHand need admin/root permissions?

On macOS, you need to grant Accessibility permissions to your terminal app. On Windows, no special permissions are needed — UI Automation works without admin for most applications.

Is ScreenHand safe to use?

ScreenHand runs locally and never sends screen data to external servers. Dangerous tools (AppleScript, browser JS execution) are audit-logged. You control which AI client connects to it via MCP configuration.

Can ScreenHand control any application?

On macOS, it can control any app that exposes Accessibility elements (most apps do). On Windows, it works with any app that supports UI Automation. Some apps with custom rendering (games, some Electron apps) may have limited element tree support — use OCR as a fallback.

How fast is ScreenHand?

Accessibility/UI Automation operations take ~50ms. Chrome CDP operations take ~10ms. Screenshots with OCR take ~600ms. Memory lookups add ~0ms (in-memory cache). ScreenHand is significantly faster than screenshot-only approaches because it reads the UI tree directly.

Does the learning memory affect performance?

No. All memory data is loaded into RAM at startup. Lookups are O(1) hash map reads. Disk writes are async and buffered — they never block tool calls. The memory system adds effectively zero latency to any tool call.

Is the memory data safe from corruption?

Yes. JSONL files are parsed line-by-line — a single corrupted line is skipped without affecting other entries. File locking prevents concurrent write corruption. Pending writes are flushed synchronously on exit (SIGINT/SIGTERM). Cache sizes are capped with LRU eviction to prevent unbounded growth.

Contributing

Contributions are welcome! Please open an issue first to discuss what you'd like to change.

git clone https://github.com/manushi4/screenhand.git
cd screenhand
npm install
npm run build:native
npm test

Contact

• Email: [email protected]
• Issues: github.com/manushi4/screenhand/issues
• Website: screenhand.com

License

AGPL-3.0-only — Copyright (C) 2025 Clazro Technology Private Limited

screenhand.com | [email protected] | A product of Clazro Technology Private Limited