vision-memory-mcp
Local cache of UI screenshots so a coding agent can recognise screens it has already seen
Connects MCP clients to the Ui.Vision RPA extension to write, run and debug browser and desktop macros
Official · Ui.Vision JavaScript Local · stdio Browser automationTesting
54 popularity score ★ 2.1k ↓ 251/wk
Opens the app, which asks before it adds the server.
claude mcp add ui-vision -- npx -y uivision-mcp-bridge{
"mcpServers": {
"ui-vision": {
"command": "npx",
"args": [
"-y",
"uivision-mcp-bridge"
]
}
}
}{
"mcpServers": {
"ui-vision": {
"command": "npx",
"args": [
"-y",
"uivision-mcp-bridge"
]
}
}
}{
"servers": {
"ui-vision": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"uivision-mcp-bridge"
]
}
}
}[mcp_servers.ui-vision]
command = "npx"
args = ["-y", "uivision-mcp-bridge"]gemini mcp add --scope user ui-vision npx -- -y uivision-mcp-bridgeOpens the app, which asks before it adds the server. Then replace each <NAME> in its settings with your value.
git clone https://github.com/A9T9/RPA.git <ABSOLUTE_PATH_TO_CLONE>
cd <ABSOLUTE_PATH_TO_CLONE>/mcp && npm installclaude mcp add ui-vision -- node '<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js'{
"mcpServers": {
"ui-vision": {
"command": "node",
"args": [
"<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js"
]
}
}
}{
"mcpServers": {
"ui-vision": {
"command": "node",
"args": [
"<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js"
]
}
}
}{
"servers": {
"ui-vision": {
"type": "stdio",
"command": "node",
"args": [
"<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js"
]
}
}
}[mcp_servers.ui-vision]
command = "node"
args = ["<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js"]gemini mcp add --scope user ui-vision node -- '<ABSOLUTE_PATH_TO_CLONE>/mcp/uivision-mcp-bridge.js'UIVISION_MCP_PORT |
Optional: Local WebSocket port the extension connects to (default 50888; same as --port); must match Settings > AI in the extension. |
|---|---|
UIVISION_MCP_TOKEN |
Optional: Fallback shared secret for connections that send no origin (same as --token); generated and saved in the home folder when not set. |
Ui.Vision MCP is a small Node.js bridge from the makers of the Ui.Vision RPA browser extension. Your MCP client starts it over stdio, and the extension's side panel connects to it over a WebSocket on 127.0.0.1 (port 50888 by default). Each tool call is passed to the extension and carried out there, inside the browser profile you already use, with its logins and cookies.
The assistant inspects the page as an accessibility tree or a screenshot, writes a macro (preferably JavaScript on Ui.Vision's uiv.* API), runs it, then reads the log and screenshots and fixes it. Macros it builds are kept in an "AI Generated" folder so you can review and replay them without the assistant. Besides DOM locators, macros can find targets by OCR text or image matching, and with the separate Desktop Automation app (XModules) they can click and type outside the browser: desktop programs, remote desktops, PDFs or canvas content.
The published 1.7.20 bridge registers 24 tools plus get_page, an alias kept for browser_snapshot. That includes tools the README's list skips: run_status, get_macro, send_chat and get_chat (drive the extension's own built-in AI chat), and get_ai_settings and set_ai_settings (change the extension's AI provider, model and OCR engine). Calls time out after 3 minutes by default, 15 minutes for run_macro and 11 for run_selftest (uivision-mcp-bridge.js, 1.7.20). An installer (npx uivision-mcp-bridge --setup) writes the server entry into Claude Code, Claude Desktop, Codex, Cursor, Windsurf and VS Code configs it finds, backing each one up first.
| Runtime | Node.js 18+ |
|---|
Node.js 18 or newer for npx. The Ui.Vision RPA extension (10.0.21 or newer) in Chrome, Edge or Firefox, with Settings > AI > MCP bridge switched on and its side panel open while you work. Desktop automation, whole-screen screenshots and local file access also need the separately installed Ui.Vision Desktop Automation app (XModules, version 2 for Ui.Vision V10), which the vendor sells in paid editions; check its pricing page.
Does nothing until the extension's side panel is connected: otherwise tool calls return "extension not connected" (open_panel can reopen it if the browser is running). Trusted CDP input (uiv.browser.*) and background Browser Vision work in Chrome and Edge, not Firefox. Only macros in the "AI Generated" folder can be deleted, and user macros are saved as copies rather than overwritten. click_at and type_at are described as for exploring only, not for saved automation. Long runs also need the MCP client's own tool timeout raised (the docs give 900 s for Codex). Only one bridge can hold the port; extra client windows relay through it (bridge 1.7.5+).
Suits people who already use Ui.Vision, or who need an assistant to automate work in their own signed-in browser and keep the result as a macro they can rerun. It also fits screens that have no usable DOM, such as remote desktops, legacy Windows apps or canvas, where OCR and image matching help (desktop work needs the paid XModules). Its own README contrasts it with DOM-only servers such as Playwright MCP and Chrome DevTools MCP: for headless runs, CI test suites or a browser with no extension installed, one of those is the better pick. It needs a desktop with a browser open, so it does not suit servers or containers.
bridge_statusrun_statusselect_browseropen_panellist_macrosopen_macroget_authoring_guideget_macrosend_chatget_chatreload_extensionrun_selftestget_ai_settingsset_ai_settingsset_macrocreate_macrodelete_macrorun_macrobrowser_snapshotget_pageclick_attype_atscreenshotsave_element_imagesave_relative_image