screenhand 0.2.0 → 0.3.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (212) hide show
  1. package/README.md +165 -446
  2. package/bin/darwin-arm64/macos-bridge +0 -0
  3. package/dist/mcp-desktop.js +3615 -400
  4. package/dist/scripts/export-help-center.js +112 -0
  5. package/dist/scripts/marketing-loop.js +117 -0
  6. package/dist/scripts/observer-daemon.js +288 -0
  7. package/dist/scripts/orchestrator-daemon.js +399 -0
  8. package/dist/scripts/threads-campaign.js +208 -0
  9. package/dist/src/community/fetcher.js +109 -0
  10. package/dist/src/community/index.js +6 -0
  11. package/dist/src/community/publisher.js +191 -0
  12. package/dist/src/community/remote-api.js +121 -0
  13. package/dist/src/community/types.js +3 -0
  14. package/dist/src/community/validator.js +95 -0
  15. package/dist/src/context-tracker.js +489 -0
  16. package/dist/src/ingestion/coverage-auditor.js +233 -0
  17. package/dist/src/ingestion/doc-parser.js +164 -0
  18. package/dist/src/ingestion/index.js +8 -0
  19. package/dist/src/ingestion/menu-scanner.js +152 -0
  20. package/dist/src/ingestion/reference-merger.js +186 -0
  21. package/dist/src/ingestion/shortcut-extractor.js +180 -0
  22. package/dist/src/ingestion/tutorial-extractor.js +170 -0
  23. package/dist/src/ingestion/types.js +3 -0
  24. package/dist/src/jobs/manager.js +82 -14
  25. package/dist/src/jobs/runner.js +138 -15
  26. package/dist/src/learning/engine.js +356 -0
  27. package/dist/src/learning/index.js +9 -0
  28. package/dist/src/learning/locator-policy.js +120 -0
  29. package/dist/src/learning/pattern-policy.js +89 -0
  30. package/dist/src/learning/recovery-policy.js +116 -0
  31. package/dist/src/learning/sensor-policy.js +115 -0
  32. package/dist/src/learning/timing-model.js +204 -0
  33. package/dist/src/learning/topology-policy.js +90 -0
  34. package/dist/src/learning/types.js +9 -0
  35. package/dist/src/logging/timeline-logger.js +4 -1
  36. package/dist/src/memory/playbook-seeds.js +200 -0
  37. package/dist/src/memory/recall.js +60 -8
  38. package/dist/src/memory/service.js +30 -5
  39. package/dist/src/memory/store.js +34 -5
  40. package/dist/src/native/bridge-client.js +253 -31
  41. package/dist/src/observer/state.js +199 -0
  42. package/dist/src/observer/types.js +43 -0
  43. package/dist/src/orchestrator/state.js +68 -0
  44. package/dist/src/orchestrator/types.js +22 -0
  45. package/dist/src/perception/ax-source.js +162 -0
  46. package/dist/src/perception/cdp-source.js +162 -0
  47. package/dist/src/perception/coordinator.js +771 -0
  48. package/dist/src/perception/frame-differ.js +287 -0
  49. package/dist/src/perception/index.js +22 -0
  50. package/dist/src/perception/manager.js +199 -0
  51. package/dist/src/perception/types.js +47 -0
  52. package/dist/src/perception/vision-source.js +399 -0
  53. package/dist/src/planner/deterministic.js +298 -0
  54. package/dist/src/planner/executor.js +870 -0
  55. package/dist/src/planner/goal-store.js +92 -0
  56. package/dist/src/planner/index.js +21 -0
  57. package/dist/src/planner/planner.js +520 -0
  58. package/dist/src/planner/tool-registry.js +71 -0
  59. package/dist/src/planner/types.js +22 -0
  60. package/dist/src/platform/explorer.js +213 -0
  61. package/dist/src/platform/help-center-markdown.js +527 -0
  62. package/dist/src/platform/learner.js +257 -0
  63. package/dist/src/playbook/engine.js +296 -11
  64. package/dist/src/playbook/mcp-recorder.js +204 -0
  65. package/dist/src/playbook/recorder.js +3 -2
  66. package/dist/src/playbook/runner.js +1 -1
  67. package/dist/src/playbook/store.js +139 -10
  68. package/dist/src/recovery/detectors.js +156 -0
  69. package/dist/src/recovery/engine.js +327 -0
  70. package/dist/src/recovery/index.js +20 -0
  71. package/dist/src/recovery/strategies.js +274 -0
  72. package/dist/src/recovery/types.js +20 -0
  73. package/dist/src/runtime/accessibility-adapter.js +55 -18
  74. package/dist/src/runtime/applescript-adapter.js +8 -2
  75. package/dist/src/runtime/cdp-chrome-adapter.js +1 -1
  76. package/dist/src/runtime/executor.js +23 -3
  77. package/dist/src/runtime/locator-cache.js +24 -2
  78. package/dist/src/runtime/service.js +59 -15
  79. package/dist/src/runtime/session-manager.js +4 -1
  80. package/dist/src/runtime/vision-adapter.js +2 -1
  81. package/dist/src/state/app-map-types.js +72 -0
  82. package/dist/src/state/app-map.js +1974 -0
  83. package/dist/src/state/entity-tracker.js +108 -0
  84. package/dist/src/state/fusion.js +96 -0
  85. package/dist/src/state/index.js +21 -0
  86. package/dist/src/state/ladder-generator.js +236 -0
  87. package/dist/src/state/persistence.js +156 -0
  88. package/dist/src/state/types.js +17 -0
  89. package/dist/src/state/world-model.js +1456 -0
  90. package/dist/src/util/atomic-write.js +19 -4
  91. package/dist/src/util/sanitize.js +146 -0
  92. package/dist-app-maps/com.figma.Desktop.json +959 -0
  93. package/dist-app-maps/com.hnc.Discord.json +1146 -0
  94. package/dist-app-maps/notion.id.json +2831 -0
  95. package/dist-playbooks/canva-screenhand-carousel.json +445 -0
  96. package/dist-playbooks/codex-desktop.json +76 -0
  97. package/dist-playbooks/competitor-research-stack.json +122 -0
  98. package/dist-playbooks/davinci-color-grade.json +153 -0
  99. package/dist-playbooks/davinci-edit-timeline.json +162 -0
  100. package/dist-playbooks/davinci-render.json +114 -0
  101. package/dist-playbooks/devto.json +52 -0
  102. package/dist-playbooks/discord.json +41 -0
  103. package/dist-playbooks/google-flow-create-project.json +59 -0
  104. package/dist-playbooks/google-flow-edit-image.json +90 -0
  105. package/dist-playbooks/google-flow-edit-video.json +90 -0
  106. package/dist-playbooks/google-flow-generate-image.json +68 -0
  107. package/dist-playbooks/google-flow-generate-video.json +191 -0
  108. package/dist-playbooks/google-flow-open-project.json +48 -0
  109. package/dist-playbooks/google-flow-open-scenebuilder.json +64 -0
  110. package/dist-playbooks/google-flow-search-assets.json +64 -0
  111. package/dist-playbooks/instagram.json +57 -0
  112. package/dist-playbooks/linkedin.json +52 -0
  113. package/dist-playbooks/n8n.json +43 -0
  114. package/dist-playbooks/reddit.json +52 -0
  115. package/dist-playbooks/threads.json +59 -0
  116. package/dist-playbooks/x-twitter.json +59 -0
  117. package/dist-playbooks/youtube.json +59 -0
  118. package/dist-references/canva.json +646 -0
  119. package/dist-references/codex-desktop.json +305 -0
  120. package/dist-references/davinci-resolve-keyboard.json +594 -0
  121. package/dist-references/davinci-resolve-menu-map.json +1139 -0
  122. package/dist-references/davinci-resolve-menus-batch1.json +116 -0
  123. package/dist-references/davinci-resolve-menus-batch2.json +372 -0
  124. package/dist-references/davinci-resolve-menus-batch3.json +330 -0
  125. package/dist-references/davinci-resolve-menus-batch4.json +297 -0
  126. package/dist-references/davinci-resolve-shortcuts.json +333 -0
  127. package/dist-references/devpost.json +186 -0
  128. package/dist-references/devto.json +317 -0
  129. package/dist-references/discord.json +549 -0
  130. package/dist-references/figma.json +1186 -0
  131. package/dist-references/finder.json +146 -0
  132. package/dist-references/google-ads-transparency.json +95 -0
  133. package/dist-references/google-flow.json +649 -0
  134. package/dist-references/instagram.json +341 -0
  135. package/dist-references/linkedin.json +324 -0
  136. package/dist-references/meta-ad-library.json +86 -0
  137. package/dist-references/n8n.json +387 -0
  138. package/dist-references/notes.json +27 -0
  139. package/dist-references/notion.json +163 -0
  140. package/dist-references/reddit.json +341 -0
  141. package/dist-references/threads.json +337 -0
  142. package/dist-references/x-twitter.json +403 -0
  143. package/dist-references/youtube.json +373 -0
  144. package/native/macos-bridge/Package.swift +22 -0
  145. package/native/macos-bridge/Sources/AccessibilityBridge.swift +482 -0
  146. package/native/macos-bridge/Sources/AppManagement.swift +339 -0
  147. package/native/macos-bridge/Sources/CoreGraphicsBridge.swift +537 -0
  148. package/native/macos-bridge/Sources/ObserverBridge.swift +120 -0
  149. package/native/macos-bridge/Sources/StreamCapture.swift +136 -0
  150. package/native/macos-bridge/Sources/VisionBridge.swift +238 -0
  151. package/native/macos-bridge/Sources/main.swift +498 -0
  152. package/native/windows-bridge/AppManagement.cs +234 -0
  153. package/native/windows-bridge/InputBridge.cs +436 -0
  154. package/native/windows-bridge/Program.cs +270 -0
  155. package/native/windows-bridge/ScreenCapture.cs +453 -0
  156. package/native/windows-bridge/UIAutomationBridge.cs +571 -0
  157. package/native/windows-bridge/WindowsBridge.csproj +17 -0
  158. package/package.json +12 -1
  159. package/scripts/postinstall.cjs +127 -0
  160. package/dist/.audit-log.jsonl +0 -55
  161. package/dist/.screenhand/memory/.lock +0 -1
  162. package/dist/.screenhand/memory/actions.jsonl +0 -85
  163. package/dist/.screenhand/memory/errors.jsonl +0 -5
  164. package/dist/.screenhand/memory/errors.jsonl.bak +0 -4
  165. package/dist/.screenhand/memory/state.json +0 -35
  166. package/dist/.screenhand/memory/state.json.bak +0 -35
  167. package/dist/.screenhand/memory/strategies.jsonl +0 -12
  168. package/dist/agent/cli.js +0 -73
  169. package/dist/agent/loop.js +0 -258
  170. package/dist/config.js +0 -9
  171. package/dist/index.js +0 -56
  172. package/dist/logging/timeline-logger.js +0 -29
  173. package/dist/mcp/mcp-stdio-server.js +0 -448
  174. package/dist/mcp/server.js +0 -347
  175. package/dist/mcp-entry.js +0 -59
  176. package/dist/memory/recall.js +0 -160
  177. package/dist/memory/research.js +0 -98
  178. package/dist/memory/seeds.js +0 -89
  179. package/dist/memory/session.js +0 -161
  180. package/dist/memory/store.js +0 -391
  181. package/dist/memory/types.js +0 -4
  182. package/dist/monitor/codex-monitor.js +0 -377
  183. package/dist/monitor/task-queue.js +0 -84
  184. package/dist/monitor/types.js +0 -49
  185. package/dist/native/bridge-client.js +0 -174
  186. package/dist/native/macos-bridge-client.js +0 -5
  187. package/dist/npm-publish-helper.js +0 -117
  188. package/dist/npm-token-cdp.js +0 -113
  189. package/dist/npm-token-create.js +0 -135
  190. package/dist/npm-token-finish.js +0 -126
  191. package/dist/playbook/engine.js +0 -193
  192. package/dist/playbook/index.js +0 -4
  193. package/dist/playbook/recorder.js +0 -519
  194. package/dist/playbook/runner.js +0 -392
  195. package/dist/playbook/store.js +0 -166
  196. package/dist/playbook/types.js +0 -4
  197. package/dist/runtime/accessibility-adapter.js +0 -377
  198. package/dist/runtime/app-adapter.js +0 -48
  199. package/dist/runtime/applescript-adapter.js +0 -283
  200. package/dist/runtime/ax-role-map.js +0 -80
  201. package/dist/runtime/browser-adapter.js +0 -36
  202. package/dist/runtime/cdp-chrome-adapter.js +0 -505
  203. package/dist/runtime/composite-adapter.js +0 -205
  204. package/dist/runtime/executor.js +0 -250
  205. package/dist/runtime/locator-cache.js +0 -12
  206. package/dist/runtime/planning-loop.js +0 -47
  207. package/dist/runtime/service.js +0 -372
  208. package/dist/runtime/session-manager.js +0 -28
  209. package/dist/runtime/state-observer.js +0 -105
  210. package/dist/runtime/vision-adapter.js +0 -208
  211. package/dist/test-mcp-protocol.js +0 -138
  212. package/dist/types.js +0 -1
package/README.md CHANGED
@@ -2,9 +2,9 @@
2
2
 
3
3
  # ScreenHand
4
4
 
5
- **Native desktop control for MCP agents.**
5
+ **Let AI control your desktop — click buttons, fill forms, automate workflows in ~50ms with zero extra AI calls.**
6
6
 
7
- An open-source [MCP server](https://modelcontextprotocol.io/) for macOS and Windows that gives Claude, Cursor, Codex CLI, and OpenClaw fast desktop control via Accessibility/UI Automation, OCR, and Chrome CDP.
7
+ An open-source [MCP server](https://modelcontextprotocol.io/) for macOS and Windows. Works with Claude, Cursor, Codex CLI, and any MCP-compatible client.
8
8
 
9
9
  [![License: AGPL-3.0](https://img.shields.io/badge/License-AGPL--3.0-blue.svg)](LICENSE)
10
10
  [![npm: screenhand](https://img.shields.io/npm/v/screenhand)](https://www.npmjs.com/package/screenhand)
@@ -12,570 +12,289 @@ An open-source [MCP server](https://modelcontextprotocol.io/) for macOS and Wind
12
12
  [![Platform: macOS & Windows](https://img.shields.io/badge/Platform-macOS%20%7C%20Windows-green)]()
13
13
  [![MCP Compatible](https://img.shields.io/badge/MCP-Compatible-purple)]()
14
14
 
15
- [Website](https://screenhand.com) | [Quick Start](#quick-start) | [Why ScreenHand](#why-screenhand) | [Tools](#tools) | [FAQ](#faq)
15
+ [Quick Start](#quick-start) | [What It Does](#what-it-does) | [Example](#example) | [All 111 Tools](docs/tools.md) | [Architecture](docs/architecture.md) | [Website](https://screenhand.com)
16
16
 
17
17
  </div>
18
18
 
19
19
  ---
20
20
 
21
- ## Why ScreenHand?
21
+ <!-- TODO: Add demo GIF here — 15 sec showing Claude controlling a real app -->
22
22
 
23
- - `~50ms` native UI actions via Accessibility APIs and Windows UI Automation
24
- - `0` extra AI calls for native clicks, typing, and UI element lookup
25
- - `70+` tools across desktop apps, browser automation, OCR, memory, sessions, jobs, and playbooks
26
- - `macOS + Windows` behind the same MCP interface
27
- - **Multi-agent safe** — session leases prevent conflicts between Claude, Cursor, and Codex
28
- - **Background worker** — queue jobs and let the daemon process them continuously
23
+ ## The Problem
29
24
 
30
- ## What is ScreenHand?
25
+ AI assistants can write code but can't use your computer. Every click requires a screenshot → LLM interpretation → coordinate guess — **3-5 seconds and an API call per action**.
31
26
 
32
- ScreenHand is a **desktop automation bridge for AI**. It connects AI assistants like Claude to your operating system so they can:
27
+ ScreenHand gives AI direct access to native OS APIs. No screenshots needed for clicks. No AI calls for button presses.
33
28
 
34
- - **See** your screen via screenshots and OCR
35
- - **Read** UI elements via Accessibility APIs (macOS) or UI Automation (Windows)
36
- - **Click** buttons, menus, and links
37
- - **Type** text into any input field
38
- - **Control** Chrome tabs via DevTools Protocol
39
- - **Run** AppleScript commands (macOS)
40
- - **Queue & execute** multi-step jobs via playbooks with a background worker daemon
41
- - **Coordinate** multiple AI agents with session leases and stall detection
42
-
43
- It works as an [MCP (Model Context Protocol)](https://modelcontextprotocol.io/) server, meaning any MCP-compatible AI client can use it out of the box.
44
-
45
- | Problem | ScreenHand Solution |
46
- |---|---|
47
- | AI can't see your screen | Screenshots + OCR return all visible text |
48
- | AI can't click UI elements | Accessibility API finds and clicks elements in ~50ms |
49
- | AI can't control browsers | Chrome DevTools Protocol gives full page control |
50
- | AI can't automate workflows | 70+ tools for cross-app automation |
51
- | Only works on one OS | Native bridges for both macOS and Windows |
52
- | Multiple agents conflict | Session leases with heartbeat and stall detection |
53
- | Jobs need manual triggering | Worker daemon processes the queue continuously |
29
+ | | Without ScreenHand | With ScreenHand |
30
+ |---|---|---|
31
+ | Click a button | Screenshot → LLM → coordinate click (~3-5s) | Native Accessibility API (~50ms) |
32
+ | Cost per action | 1 LLM API call | 0 LLM calls |
33
+ | Accuracy | Coordinate guessing misses on layout shift | Exact element targeting by role/name |
34
+ | Browser control | Needs focus, screenshot per action | CDP in background (~10ms), no focus needed |
35
+ | Works across apps | One app at a time | Cross-app workflows, multi-agent coordination |
54
36
 
55
37
  ## Quick Start
56
38
 
57
- ### Source install (recommended today)
39
+ ### 1. Add to your AI client (one step)
58
40
 
59
- ScreenHand currently builds a native bridge locally for Accessibility/UI Automation, so the fastest reliable setup is still from source:
41
+ <details open>
42
+ <summary><b>Claude Code</b> (recommended)</summary>
60
43
 
61
44
  ```bash
62
- git clone https://github.com/manushi4/screenhand.git
63
- cd screenhand
64
- npm install
65
- npm run build:native # macOS — builds Swift bridge
66
- # npm run build:native:windows # Windows — builds .NET bridge
45
+ claude mcp add screenhand -- npx -y screenhand
67
46
  ```
68
47
 
69
- Then connect ScreenHand to your AI client.
48
+ Done. That's it.
49
+ </details>
70
50
 
71
- ### Claude Desktop
51
+ <details>
52
+ <summary><b>Claude Desktop</b></summary>
72
53
 
73
54
  Add to `~/Library/Application Support/Claude/claude_desktop_config.json`:
74
-
75
55
  ```json
76
56
  {
77
57
  "mcpServers": {
78
58
  "screenhand": {
79
59
  "command": "npx",
80
- "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
60
+ "args": ["-y", "screenhand"]
81
61
  }
82
62
  }
83
63
  }
84
64
  ```
65
+ </details>
85
66
 
86
- ### Claude Code
87
-
88
- Add to your project `.mcp.json` or `~/.claude/settings.json`:
67
+ <details>
68
+ <summary><b>Cursor</b></summary>
89
69
 
70
+ Add to `.cursor/mcp.json`:
90
71
  ```json
91
72
  {
92
73
  "mcpServers": {
93
74
  "screenhand": {
94
75
  "command": "npx",
95
- "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
76
+ "args": ["-y", "screenhand"]
96
77
  }
97
78
  }
98
79
  }
99
80
  ```
81
+ </details>
100
82
 
101
- ### Cursor
102
-
103
- Add to `.cursor/mcp.json` in your project (or `~/.cursor/mcp.json` for global):
104
-
105
- ```json
106
- {
107
- "mcpServers": {
108
- "screenhand": {
109
- "command": "npx",
110
- "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
111
- }
112
- }
113
- }
114
- ```
115
-
116
- ### OpenAI Codex CLI
83
+ <details>
84
+ <summary><b>OpenAI Codex CLI</b></summary>
117
85
 
118
86
  Add to `~/.codex/config.toml`:
119
-
120
87
  ```toml
121
88
  [mcp.screenhand]
122
89
  command = "npx"
123
- args = ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
90
+ args = ["-y", "screenhand"]
124
91
  transport = "stdio"
125
92
  ```
93
+ </details>
126
94
 
127
- ### OpenClaw
95
+ <details>
96
+ <summary><b>Any MCP Client</b></summary>
128
97
 
129
- Add to your `openclaw.json`:
98
+ ScreenHand is a standard MCP server over stdio. Run with `npx -y screenhand`.
99
+ </details>
130
100
 
131
- ```json
132
- {
133
- "mcpServers": {
134
- "screenhand": {
135
- "command": "npx",
136
- "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
137
- }
138
- }
139
- }
140
- ```
101
+ ### 2. Grant permissions
141
102
 
142
- > **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](docs/openclaw-integration.md).
103
+ **macOS**: System Settings > Privacy & Security > Accessibility > enable your terminal app.
143
104
 
144
- ### Any MCP Client
105
+ **Windows**: No special permissions needed.
145
106
 
146
- ScreenHand is a standard MCP server over stdio. It works with any MCP-compatible client — just point it at `mcp-desktop.ts`.
107
+ ### 3. Browser control (optional)
147
108
 
148
- Replace `/path/to/screenhand` with the actual path where you cloned the repo.
109
+ Launch Chrome with remote debugging to enable browser tools:
110
+ ```bash
111
+ open -a "Google Chrome" --args --remote-debugging-port=9222
112
+ ```
149
113
 
150
- ## Tools
114
+ That's it. Your AI client now has 111 tools for desktop automation.
151
115
 
152
- ScreenHand exposes 70+ tools organized by category.
116
+ <details>
117
+ <summary><b>Building from source</b> (contributors only)</summary>
153
118
 
154
- ### See the Screen
119
+ ```bash
120
+ git clone https://github.com/manushi4/screenhand.git
121
+ cd screenhand && npm install && npm run build:native
122
+ ```
155
123
 
156
- | Tool | What it does | Speed |
157
- |------|-------------|-------|
158
- | `screenshot` | Full screenshot + OCR — returns all visible text | ~600ms |
159
- | `screenshot_file` | Screenshot saved to file (for viewing the image) | ~400ms |
160
- | `ocr` | OCR with element positions and bounding boxes | ~600ms |
124
+ On Windows, use `npm run build:native:windows` instead.
125
+ </details>
161
126
 
162
- ### Control Any App (Accessibility / UI Automation)
127
+ ---
163
128
 
164
- | Tool | What it does | Speed |
165
- |------|-------------|-------|
166
- | `apps` | List running apps with bundle IDs and PIDs | ~10ms |
167
- | `windows` | List visible windows with positions and sizes | ~10ms |
168
- | `focus` | Bring an app to the front | ~10ms |
169
- | `launch` | Launch an app by bundle ID or name | ~1s |
170
- | `ui_tree` | Full UI element tree — instant, no OCR needed | ~50ms |
171
- | `ui_find` | Find a UI element by text or title | ~50ms |
172
- | `ui_press` | Click a UI element by its title | ~50ms |
173
- | `ui_set_value` | Set value of a text field, slider, etc. | ~50ms |
174
- | `menu_click` | Click a menu bar item by path | ~100ms |
129
+ ## What It Does
175
130
 
176
- ### Keyboard and Mouse
131
+ ScreenHand gives AI agents seven capabilities:
177
132
 
178
- | Tool | What it does |
179
- |------|-------------|
180
- | `click` | Click at screen coordinates |
181
- | `click_text` | Find text via OCR and click it (fallback) |
182
- | `type_text` | Type text via keyboard |
183
- | `key` | Key combo (e.g. `cmd+s`, `ctrl+shift+n`) |
184
- | `drag` | Drag from point A to B |
185
- | `scroll` | Scroll at a position |
133
+ ### Desktop Control 19 tools
134
+ Click buttons, type text, read UI trees, navigate menus, drag, scroll — all via native Accessibility APIs in ~50ms. Works with any app: Finder, Notes, VS Code, Xcode, System Settings, etc.
186
135
 
187
- ### Chrome Browser (CDP)
136
+ ### Browser Automation — 15 tools
137
+ Full Chrome control via DevTools Protocol. Navigate, click, type, run JavaScript, fill forms — all in the background at ~10ms. Built-in anti-detection (`browser_stealth`, `browser_human_click`) for sites with bot protection.
188
138
 
189
- | Tool | What it does |
190
- |------|-------------|
191
- | `browser_tabs` | List all open Chrome tabs |
192
- | `browser_open` | Open URL in new tab |
193
- | `browser_navigate` | Navigate active tab to URL |
194
- | `browser_js` | Run JavaScript in a tab |
195
- | `browser_dom` | Query DOM with CSS selectors |
196
- | `browser_click` | Click element by CSS selector (uses CDP mouse events) |
197
- | `browser_type` | Type into an input field (uses CDP keyboard events, React-compatible) |
198
- | `browser_wait` | Wait for a page condition |
199
- | `browser_page_info` | Get page title, URL, and content |
139
+ ### Smart Fallbacks 8 tools
140
+ `click_with_fallback`, `type_with_fallback`, etc. automatically try Accessibility → CDP → OCR → coordinates. You don't have to pick the right method — ScreenHand figures it out.
200
141
 
201
- ### Anti-Detection & Stealth (CDP)
142
+ ### Memory & Learning — 14 tools
143
+ Gets smarter every session. Logs tool calls, saves winning strategies, tracks error patterns with fixes. Zero config, zero latency overhead (in-memory cache, async disk writes). Ships with 12 seed strategies for common macOS workflows. 6 learning policies: locator stability, sensor effectiveness, recovery ranking, pattern recognition, adaptive timing, and topology (navigation edge reliability).
202
144
 
203
- Tools for interacting with sites that have bot detection (Instagram, LinkedIn, etc.):
145
+ ### App Mastery Map automatic per-app spatial understanding
146
+ Builds a persistent reverse-engineered blueprint of every app from normal tool usage. 8 features record automatically: page zones, navigation graph (BFS pathfinding), hierarchy, I/O contracts, state machine, element visibility, timing profiles, and ready signals. Mastery levels (beginner → pro → expert → grandmaster) honestly reflect how well ScreenHand knows each app. Maps stored at `~/.screenhand/app-maps/`.
204
147
 
205
- | Tool | What it does |
206
- |------|-------------|
207
- | `browser_stealth` | Inject anti-detection patches (hides webdriver flag, fakes plugins/languages) |
208
- | `browser_fill_form` | Human-like typing with random delays via CDP keyboard events |
209
- | `browser_human_click` | Realistic mouse event sequence (mouseMoved → mousePressed → mouseReleased) |
148
+ ### Jobs & Orchestration 34 tools
149
+ Queue multi-step jobs, run them via background worker daemon, coordinate multiple AI agents with session leases, detect stalls, auto-recover. Survives client restarts.
210
150
 
211
- > **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.
151
+ ### Perception & Planning 17 tools
152
+ Continuous screen awareness (3-rate perception loop at 100ms/300ms/1000ms), real-time world model with entity tracking, goal-oriented planning with auto-decomposition, recovery engine with self-healing. The system always knows what's on screen and feeds observations into the App Mastery Map.
212
153
 
213
- ### Smart Execution (fallback chain)
154
+ > **Full reference**: See all [111 tools with descriptions](docs/tools.md).
214
155
 
215
- Tools that automatically choose the best method (Accessibility → CDP → OCR → coordinates):
156
+ ---
216
157
 
217
- | Tool | What it does |
218
- |------|-------------|
219
- | `execution_plan` | Generate an execution plan for a task |
220
- | `click_with_fallback` | Click using the best available method |
221
- | `type_with_fallback` | Type using the best available method |
222
- | `read_with_fallback` | Read content using the best available method |
223
- | `locate_with_fallback` | Find an element using the best available method |
224
- | `select_with_fallback` | Select an option using the best available method |
225
- | `scroll_with_fallback` | Scroll using the best available method |
226
- | `wait_for_state` | Wait for a UI state using the best available method |
158
+ ## Example
227
159
 
228
- ### Platform Playbooks (lazy-loaded)
160
+ **Browser** Claude controls Chrome in the background while you work:
229
161
 
230
- Pre-built automation knowledge for specific platforms — selectors, URLs, flows, and **error solutions**.
162
+ ```
163
+ You: Search for "screenhand" on Instagram
231
164
 
232
- | Tool | What it does |
233
- |------|-------------|
234
- | `platform_guide` | Get automation guide for a platform (selectors, URLs, flows, errors+solutions) |
235
- | `export_playbook` | Auto-generate a playbook from your session. Share it to help others. |
165
+ browser_tabs() # ~10ms
166
+ [34DF5DE1] Instagram — https://www.instagram.com/
236
167
 
237
- ```
238
- platform_guide({ platform: "devpost", section: "errors" }) # Just errors + solutions
239
- platform_guide({ platform: "devpost", section: "selectors" }) # All CSS selectors
240
- platform_guide({ platform: "devpost", section: "flows" }) # Step-by-step workflows
241
- platform_guide({ platform: "devpost" }) # Full playbook
242
- ```
168
+ → browser_js({ code: "/* click Search icon */" }) # ~10ms
169
+ → browser_fill_form({ selector: "input", text: "screenhand" }) # ~50ms (human-like)
170
+ → browser_js({ code: "/* extract results */" }) # ~10ms
243
171
 
244
- **Contributing playbooks:** After automating any site, run:
245
- ```
246
- export_playbook({ platform: "twitter", domain: "twitter.com" })
172
+ Found @screenhand_ as the top result.
247
173
  ```
248
- This auto-extracts URLs, selectors, errors+solutions from your session and saves a ready-to-share `playbooks/twitter.json`.
249
-
250
- Available platforms: `devpost`. Add more by running `export_playbook` or creating JSON files in `playbooks/`.
251
-
252
- Zero performance cost — files only read when `platform_guide` is called.
253
-
254
- ### AppleScript (macOS only)
255
-
256
- | Tool | What it does |
257
- |------|-------------|
258
- | `applescript` | Run any AppleScript command |
259
-
260
- ### Memory (Learning) — zero-config, zero-latency
261
-
262
- ScreenHand gets smarter every time you use it — **no manual setup needed**.
263
-
264
- **What happens automatically:**
265
- - Every tool call is logged (async, non-blocking — adds ~0ms to response time)
266
- - After 3+ consecutive successes, the winning sequence is saved as a reusable strategy
267
- - Known error patterns are tracked with resolutions (e.g. "launch times out → use focus() instead")
268
- - On every tool call, the response includes **auto-recall hints**:
269
- - Error warnings if the tool has failed before
270
- - Next-step suggestions if you're mid-way through a known strategy
271
-
272
- **Predefined seed strategies:**
273
- - Ships with 12 common macOS workflows (Photo Booth, Chrome navigation, copy/paste, Finder, export PDF, etc.)
274
- - Loaded automatically on first boot — the system has knowledge from day one
275
- - Seeds are searchable via `memory_recall` and provide next-step hints like any learned strategy
276
-
277
- **Background web research:**
278
- - When a tool fails and no resolution exists, ScreenHand searches for a fix in the background (non-blocking)
279
- - Uses Claude API (haiku, if `ANTHROPIC_API_KEY` is set) or DuckDuckGo instant answers as fallback
280
- - Resolutions are saved to both error cache and strategy store — zero-latency recall next time
281
- - Completely silent and fire-and-forget — never blocks tool responses or throws errors
282
-
283
- **Fingerprint matching & feedback loop:**
284
- - Each strategy is fingerprinted by its tool sequence (e.g. `apps→focus→ui_press`)
285
- - O(1) exact-match lookup when the agent follows a known sequence
286
- - Success/failure outcomes are tracked per strategy — unreliable strategies are auto-penalized and eventually skipped
287
- - Keyword-based fuzzy search with reliability scoring for `memory_recall`
288
-
289
- **Production-grade under the hood:**
290
- - All data cached in RAM at startup — lookups are ~0ms, disk is only for persistence
291
- - Disk writes are async and buffered (100ms debounce) — never block tool calls
292
- - Sync flush on process exit (SIGINT/SIGTERM) — no lost writes
293
- - Per-line JSONL parsing — corrupted lines are skipped, not fatal
294
- - LRU eviction: 500 strategies, 200 error patterns max (oldest evicted automatically)
295
- - File locking (`.lock` + PID) prevents corruption from concurrent instances
296
- - Action log auto-rotates at 10 MB
297
- - Data lives in `.screenhand/memory/` as JSONL (grep-friendly, no database)
298
-
299
- | Tool | What it does |
300
- |------|-------------|
301
- | `memory_snapshot` | Get current memory state snapshot |
302
- | `memory_recall` | Search past strategies by task description |
303
- | `memory_save` | Manually save the current session as a strategy |
304
- | `memory_record_error` | Record an error pattern with an optional fix |
305
- | `memory_record_learning` | Record a verified pattern (what works/fails) |
306
- | `memory_query_patterns` | Search learnings by scope and method |
307
- | `memory_errors` | View all known error patterns and resolutions |
308
- | `memory_stats` | Action counts, success rates, top tools, disk usage |
309
- | `memory_clear` | Clear actions, strategies, errors, or all data |
310
-
311
- ### Session Supervisor — multi-agent coordination
312
-
313
- Lease-based window locking with heartbeat, stall detection, and automatic recovery. Prevents multiple AI agents from fighting over the same app window.
314
-
315
- | Tool | What it does |
316
- |------|-------------|
317
- | `session_claim` | Claim exclusive control of an app window |
318
- | `session_heartbeat` | Keep your lease alive (call every 60s) |
319
- | `session_release` | Release your session lease |
320
- | `supervisor_status` | Active sessions, health metrics, stall detection |
321
- | `supervisor_start` | Start the supervisor background daemon |
322
- | `supervisor_stop` | Stop the supervisor daemon |
323
- | `supervisor_pause` | Pause supervisor monitoring |
324
- | `supervisor_resume` | Resume supervisor monitoring |
325
- | `supervisor_install` | Install supervisor as a launchd service (macOS) |
326
- | `supervisor_uninstall` | Uninstall supervisor launchd service |
327
- | `recovery_queue_add` | Add a recovery action to the supervisor's queue |
328
- | `recovery_queue_list` | List pending recovery actions |
329
-
330
- 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.
331
-
332
- ### Jobs & Worker Daemon
333
-
334
- 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.
335
-
336
- | Tool | What it does |
337
- |------|-------------|
338
- | `job_create` | Create a job with steps (optionally tied to a playbook + bundleId/windowId) |
339
- | `job_status` | Get the status of a job |
340
- | `job_list` | List jobs by state (queued, running, done, failed, blocked) |
341
- | `job_transition` | Transition a job to a new state |
342
- | `job_step_done` | Mark a job step as done |
343
- | `job_step_fail` | Mark a job step as failed |
344
- | `job_resume` | Resume a blocked/waiting job |
345
- | `job_dequeue` | Dequeue the next queued job |
346
- | `job_remove` | Remove a job |
347
- | `job_run` | Execute a single queued job through the runner |
348
- | `job_run_all` | Process all queued jobs sequentially |
349
- | `worker_start` | Start the background worker daemon |
350
- | `worker_stop` | Stop the worker daemon |
351
- | `worker_status` | Get worker daemon status and recent results |
352
-
353
- **Job state machine:** `queued → running → done | failed | blocked | waiting_human`
354
-
355
- **Worker daemon features:**
356
- - Runs as a detached process — survives MCP/client restarts
357
- - Continuously polls the job queue and executes via JobRunner
358
- - Playbook integration — jobs with a `playbookId` execute through PlaybookEngine
359
- - Focuses/validates the target `bundleId`/`windowId` before each step
360
- - Persists status and recent results to `~/.screenhand/worker/state.json`
361
- - Single-instance enforcement via PID file
362
- - Graceful shutdown on SIGINT/SIGTERM
363
174
 
364
- ```bash
365
- # Start the worker daemon directly
366
- npx tsx scripts/worker-daemon.ts
367
- npx tsx scripts/worker-daemon.ts --poll 5000 --max-jobs 10
175
+ **Desktop** — native app control without screenshots:
368
176
 
369
- # Or via MCP tools
370
- worker_start worker_status worker_stop
177
+ ```
178
+ apps() # List running apps ~10ms
179
+ → focus("com.apple.Notes") # Bring Notes to front ~10ms
180
+ → ui_tree() # Read full UI element tree ~50ms
181
+ → ui_press("New Note") # Click "New Note" button ~50ms
182
+ → type_text("Hello world") # Type text ~30ms
371
183
  ```
372
184
 
373
- ## Architecture
185
+ **Cross-app** — chain actions across your whole desktop:
374
186
 
375
187
  ```
376
- ┌─────────────────────────────────────────────────────┐
377
- │ MCP Client (Claude, Cursor, Codex CLI, OpenClaw) │
378
- └────────────────────────┬────────────────────────────┘
379
- stdio JSON-RPC
380
- ┌────────────────────────▼────────────────────────────┐
381
- │ mcp-desktop.ts │
382
- │ (MCP Server — 70+ tools) │
383
- ├───────────┬──────────┬──────────────────────────────┤
384
- │ Native │ Chrome │ Memory / Supervisor / Jobs │
385
- │ Bridge │ CDP │ / Playbooks / Worker │
386
- └─────┬─────┴────┬─────┴──────────────────────────────┘
387
- │ │
388
- ┌─────▼─────┐ ┌──▼──────┐ ┌──────────────┐ ┌──────────────┐
389
- │macos-bridge│ │ Chrome │ │ Supervisor │ │ Worker │
390
- │(Swift, AX) │ │DevTools │ │ Daemon │ │ Daemon │
391
- └────────────┘ └─────────┘ └──────────────┘ └──────────────┘
188
+ → browser_js(...) # Extract data from Chrome
189
+ focus("com.apple.Notes") # Switch to Notes
190
+ → type_text(extractedData) # Paste it in
191
+ key("cmd+s") # Save
392
192
  ```
393
193
 
394
- ### Key modules
395
-
396
- | Path | Purpose |
397
- |---|---|
398
- | `mcp-desktop.ts` | MCP server entrypoint — all tool definitions |
399
- | `src/native/bridge-client.ts` | TypeScript ↔ native bridge communication |
400
- | `native/macos-bridge/` | Swift binary using Accessibility API + OCR |
401
- | `native/windows-bridge/` | C# binary using UI Automation + SendInput |
402
- | `src/memory/` | Persistent memory service (strategies, errors, learnings) |
403
- | `src/supervisor/` | Session leases, stall detection, recovery |
404
- | `src/jobs/` | Job queue, runner, worker state persistence |
405
- | `src/playbook/` | Playbook engine and store |
406
- | `src/runtime/` | Execution contract, accessibility adapter, fallback chain |
407
- | `scripts/worker-daemon.ts` | Standalone worker daemon process |
408
- | `scripts/supervisor-daemon.ts` | Standalone supervisor daemon process |
194
+ ---
409
195
 
410
- ### State files
196
+ ## Claude Code Plugin
411
197
 
412
- All persistent state lives under `~/.screenhand/`:
198
+ If you use Claude Code, ScreenHand includes a plugin with **13 skills and 5 agents** that wrap all 111 tools into intent-oriented workflows.
413
199
 
414
- ```
415
- ~/.screenhand/
416
- ├── memory/ # strategies, errors, learnings (JSONL)
417
- ├── supervisor/ # supervisor daemon state
418
- ├── locks/ # session lease files
419
- ├── jobs/ # job queue persistence
420
- ├── worker/ # worker daemon state, PID, logs
421
- └── playbooks/ # saved playbook definitions
200
+ ```bash
201
+ ./install-plugin.sh # after npm install && npm run build:native
422
202
  ```
423
203
 
424
- ## How It Works
204
+ | Skill | What it does |
205
+ |-------|-------------|
206
+ | `/automate` | Control any desktop app |
207
+ | `/post-social` | Post to X, LinkedIn, Instagram, Reddit, Threads, Discord |
208
+ | `/run-campaign` | Multi-platform marketing campaigns |
209
+ | `/edit-video` | DaVinci Resolve automation |
210
+ | `/design-figma` | Figma design via Plugin API + browser |
211
+ | `/edit-canva` | Canva template editing |
212
+ | `/scrape-web` | Data extraction with anti-detection |
213
+ | `/fill-form` | Human-like form filling |
214
+ | `/qa-smoke-test` | Automated UI testing |
215
+ | `/record-workflow` | Record into reusable playbooks |
216
+ | `/learn-platform` | Discover how to automate a new app/site |
217
+ | `/run-jobs` | Job queues, background workers |
218
+ | `/manage-system` | Supervisor, memory, diagnostics |
219
+
220
+ 5 specialized agents: **marketing**, **design**, **QA**, **scraper**, **orchestrator**.
221
+
222
+ ---
425
223
 
426
- ScreenHand has three layers:
224
+ ## How It Works
427
225
 
428
226
  ```
429
- AI Client (Claude, Cursor, etc.)
227
+ AI Client (Claude, Cursor, Codex CLI)
430
228
  ↓ MCP protocol (stdio)
431
229
  ScreenHand MCP Server (TypeScript)
432
230
  ↓ JSON-RPC (stdio)
433
231
  Native Bridge (Swift on macOS / C# on Windows)
434
- Platform APIs
435
- Operating System (Accessibility, CoreGraphics, UI Automation, SendInput)
232
+ OS APIs
233
+ Accessibility, CoreGraphics, Vision, UI Automation, SendInput
436
234
  ```
437
235
 
438
- 1. **Native bridge** talks directly to OS-level APIs:
439
- - **macOS**: Swift binary using Accessibility APIs, CoreGraphics, and Vision framework (OCR)
440
- - **Windows**: C# (.NET 8) binary using UI Automation, SendInput, GDI+, and Windows.Media.Ocr
441
- 2. **TypeScript MCP server** — routes tools to the correct bridge, handles Chrome CDP, manages sessions, runs jobs
442
- 3. **MCP protocol** — standard Model Context Protocol so any AI client can connect
443
-
444
- 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.
445
-
446
- ## Use Cases
447
-
448
- ### App Debugging
449
- Claude reads UI trees, clicks through flows, and checks element states — faster than clicking around yourself.
450
-
451
- ### Design Inspection
452
- Screenshots + OCR to read exactly what's on screen. `ui_tree` shows component structure like React DevTools but for any native app.
453
-
454
- ### Browser Automation
455
- Fill forms, scrape data, run JavaScript, navigate pages — all through Chrome DevTools Protocol.
456
-
457
- ### Cross-App Workflows
458
- 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.
459
-
460
- ### Multi-Agent Coordination
461
- Run Claude, Cursor, and Codex simultaneously — each claims its own app window via session leases. The supervisor detects stalls and recovers.
462
-
463
- ### Background Job Processing
464
- 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.
236
+ ScreenHand reads the UI tree and DOM directly no screenshots needed for most operations. When screenshots are needed (canvas apps, visual verification), OCR runs in ~600ms via the native Vision framework.
465
237
 
466
- ### UI Testing
467
- Click buttons, verify text appears, catch visual regressions — all driven by AI.
238
+ ---
468
239
 
469
240
  ## Requirements
470
241
 
471
- ### macOS
472
-
473
- - macOS 12+
474
- - Node.js 18+
475
- - Accessibility permissions: System Settings > Privacy & Security > Accessibility > enable your terminal
476
- - Chrome with `--remote-debugging-port=9222` (only for browser tools)
477
-
478
- ### Windows
479
-
480
- - Windows 10 (1809+)
481
- - Node.js 18+
482
- - [.NET 8 SDK](https://dotnet.microsoft.com/download/dotnet/8.0)
483
- - No special permissions needed — UI Automation works without admin
484
- - Chrome with `--remote-debugging-port=9222` (only for browser tools)
485
- - Build: `npm run build:native:windows`
486
-
487
- ## Skills (Slash Commands)
488
-
489
- ScreenHand ships with Claude Code slash commands:
490
-
491
- - `/screenshot` — capture your screen and describe what's visible
492
- - `/debug-ui` — inspect the UI tree of any app
493
- - `/automate` — describe a task and Claude does it
494
-
495
- **Install globally** so they work in any project:
496
-
497
- ```bash
498
- ./install-skills.sh
499
- ```
500
-
501
- ## Development
502
-
503
- ```bash
504
- npm run dev # Run MCP server with tsx (hot reload)
505
- npm run check # type-check (covers all entry files)
506
- npm test # run test suite
507
- npm run build # compile TypeScript
508
- npm run build:native # build Swift bridge (macOS)
509
- npm run build:native:windows # build .NET bridge (Windows)
510
- ```
511
-
512
- ## FAQ
513
-
514
- ### What is ScreenHand?
515
- 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.
516
-
517
- ### How does ScreenHand differ from Anthropic's Computer Use?
518
- 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.
519
-
520
- ### How does ScreenHand work with OpenClaw?
521
-
522
- ScreenHand **integrates with** OpenClaw as an MCP server — giving your Claw agent native desktop speed instead of screenshot-based clicking.
523
-
524
- | | Without ScreenHand | With ScreenHand |
242
+ | | macOS | Windows |
525
243
  |---|---|---|
526
- | **Clicking a button** | Screenshot LLM interprets → coordinate click (~3-5s) | `press('Send')` via Accessibility API (~50ms) |
527
- | **Cost per action** | 1 LLM API call per click | 0 LLM calls — native OS API |
528
- | **Accuracy** | Coordinate guessing — can miss if layout shifts | Exact element targeting by role/name |
244
+ | OS | macOS 12+ | Windows 10 (1809+) |
245
+ | Runtime | Node.js 18+ | Node.js 18+ |
246
+ | Native | Swift (included) | [.NET 8 SDK](https://dotnet.microsoft.com/download/dotnet/8.0) |
247
+ | Permissions | Accessibility access for terminal | None (UI Automation works without admin) |
248
+ | Browser | Chrome with `--remote-debugging-port=9222` | Same |
249
+
250
+ ## Docs
251
+
252
+ | Document | What's in it |
253
+ |----------|-------------|
254
+ | [All 111 Tools](docs/tools.md) | Complete tool reference with descriptions and speeds |
255
+ | [Architecture](docs/architecture.md) | 7-layer design, app tiers, performance targets |
256
+ | [App Mastery Map](docs/app-mastery-map.md) | Layer 7: persistent spatial understanding, 8 auto-recording features |
257
+ | [Bug Tracker](docs/l2-bug-tracker.md) | 103 bugs found and fixed, 80-scenario validation results |
258
+ | [Testing Plan](docs/testing-plan.md) | L1/L2 test methodology and gate criteria |
529
259
 
530
- **Setup** — add to your `openclaw.json`:
260
+ ## FAQ
531
261
 
532
- ```json
533
- {
534
- "mcpServers": {
535
- "screenhand": {
536
- "command": "npx",
537
- "args": ["tsx", "/path/to/screenhand/mcp-desktop.ts"]
538
- }
539
- }
540
- }
541
- ```
262
+ <details>
263
+ <summary><b>How is this different from Anthropic's Computer Use?</b></summary>
542
264
 
543
- 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](docs/openclaw-integration.md).
265
+ Computer Use is cloud-based and screenshot-driven. ScreenHand is local-first, uses native OS APIs (50ms vs 3-5s per action), costs zero API calls for clicks/typing, and runs entirely on your machine.
266
+ </details>
544
267
 
545
- ### Does ScreenHand work on Windows?
546
- 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.
268
+ <details>
269
+ <summary><b>What apps can it control?</b></summary>
547
270
 
548
- ### What AI clients work with ScreenHand?
549
- Any MCP-compatible client: Claude Desktop, Claude Code, Cursor, Windsurf, OpenAI Codex CLI, and any other tool that supports the Model Context Protocol.
271
+ Any app with Accessibility support (most macOS/Windows apps). Chrome and Electron apps get full DOM access via CDP. Canvas-heavy apps (games, Photoshop viewport) use OCR as fallback.
272
+ </details>
550
273
 
551
- ### Does ScreenHand need admin/root permissions?
552
- 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.
274
+ <details>
275
+ <summary><b>Is it safe?</b></summary>
553
276
 
554
- ### Is ScreenHand safe to use?
555
- 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.
277
+ Runs locally, never sends screen data externally. PII is redacted from all persisted data (memory, playbooks, strategies). Dangerous protocols (`javascript:`, `data:`) are blocked. AppleScript and browser JS execution are audit-logged.
278
+ </details>
556
279
 
557
- ### Can ScreenHand control any application?
558
- 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.
280
+ <details>
281
+ <summary><b>Does it work with multiple AI agents at once?</b></summary>
559
282
 
560
- ### How fast is ScreenHand?
561
- 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.
283
+ Yes. Session leases with heartbeat prevent conflicts. The supervisor daemon detects stalls and recovers. Each agent claims its own app window.
284
+ </details>
562
285
 
563
- ### Does the learning memory affect performance?
564
- 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.
286
+ <details>
287
+ <summary><b>How fast is it?</b></summary>
565
288
 
566
- ### Is the memory data safe from corruption?
567
- 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.
289
+ Accessibility: ~50ms. Chrome CDP: ~10ms (background, no focus needed). OCR: ~600ms. Memory lookups: ~0ms (in-memory cache). All disk writes are async and non-blocking.
290
+ </details>
568
291
 
569
292
  ## Contributing
570
293
 
571
- Contributions are welcome! Please open an issue first to discuss what you'd like to change.
572
-
573
294
  ```bash
574
295
  git clone https://github.com/manushi4/screenhand.git
575
- cd screenhand
576
- npm install
577
- npm run build:native
578
- npm test
296
+ cd screenhand && npm install && npm run build:native
297
+ npm test # 1306 tests, 53 files
579
298
  ```
580
299
 
581
300
  ## Contact
@@ -586,7 +305,7 @@ npm test
586
305
 
587
306
  ## License
588
307
 
589
- AGPL-3.0-only — Copyright (C) 2025 Clazro Technology Private Limited
308
+ AGPL-3.0-only — Copyright (C) 2025-2026 Clazro Technology Private Limited
590
309
 
591
310
  ---
592
311