Provides visual computer-use controls for macOS, enabling an agent to list and bind windows, capture screenshots, and perform native click, scroll, keyboard, and typing actions.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@VelaCU @VelaCU Click the 'Submit' button in the Safari window."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
VelaCU is a small macOS visual Computer Use MCP server. It gives an MCP-capable agent a screenshot of one window, a visible 0..10 ruler, and native keyboard/pointer controls without taking over the user's physical mouse.
VelaCU can be used with WebGPT, Codex, and other MCP-capable agents.
Release: VelaCU v0.4.2 Build: native-20260820-05 Click backend: VelaClick v0.3.0 SkyLight
velacu_list(limit=20)
velacu_bind(query | window_id)
velacu_capture()
velacu_click(x, y, button="left", count=1)
velacu_scroll(direction, amount=3, x=5.0, y=5.0)
velacu_key(key)
velacu_type(text)
velacu_release()The usual visual loop is:
list -> bind -> capture -> click/scroll/key/type -> inspect the returned result -> repeat
task cleanup -> releasevelacu_capture and the post-action captures use a fixed 640px model width and draw a window-local ruler. The origin is the top-left; x increases right and y increases down. Click and targeted-scroll coordinates are read directly from the visible 0..10 ruler and use one decimal place. Agents should never convert them to pixels.
Related MCP server: MacWright
Click delivery is a standalone, stateless native helper:
MCP client
-> VelaCU MCP
-> VelaCU core
-> VelaClick (one process per click, CGEvent + SkyLight)
-> VelaCUHelper (capture, keyboard, targeted scroll)
-> VelaPointer (visual-only cursor)VelaClick receives the target pid, window id, and resolved screen/local coordinates, then exits. VelaCU no longer requires or runs a Cua daemon, Cua session, Unix socket, or Rust/Cua build. There is no AX/DOM target lookup and the real mouse cursor is not moved.
velacu_scroll sends native SkyLight/CGEvent wheel input at the requested normalized point and returns a post-scroll screenshot. It supports up, down, left, and right, so nested scroll regions can be targeted by their visible position.
System Settings RemoteView panes retain the verified compatibility path: for that specific cross-process topology VelaCU uses a short foreground HID fallback and restores the prior application and physical cursor position. It remains pixel-only and does not use AX or DOM lookup.
The visual pointer is independent of event delivery. Recent fixes include the physical-click flicker root fix, stationary fade while minimizing, short idle hover animation without an apparent long freeze, and faster contrast adaptation. These are visual effects only and do not debounce or patch click delivery.
On macOS with Python 3.9+:
./install.sh --target codexTo install without changing an agent configuration:
./install.sh --target none
velacu setup codexThe installer creates a local VelaCU installation, a velacu launcher, and a timestamped backup before changing the Codex MCP configuration. Run:
velacu doctorA release archive normally includes native binaries. A source checkout builds the native helper, VelaClick, VelaPointer, and status app with Xcode Command Line Tools. Source installation requires Python/Pillow and Swift; Rust and cargo are not required.
For another MCP-capable client, use generic-mcp.example.json or:
velacu setup genericpython3 -m pip install -r requirements.txt
./build.sh
python3 -m py_compile velacu_core.py velacu_mcp.py velacu_cli.py status_publisher.py
python3 -m unittest + test_velacu_mcp.py + test_parallel_isolation.py + test_velacu_keyboard.py + test_velacu_e2e.py + test_velacu_window_move.py + test_zero_prompt_contract.pyThe e2e and window-movement programs need a live macOS app/window and the relevant macOS permissions. The focused unit and contract tests do not need a live target window.
When an agent finishes its control task, it must close applications or windows opened only for that task, preserve user work windows that existed before the task, and then call velacu_release. velacu_release clears VelaCU's binding and visual pointer state; it does not close pre-existing user work.
velacu_mcp.py MCP tool surface
velacu_core.py binding, capture, coordinate conversion, native bridges
velacu_cli.py installer/doctor and agent setup commands
native/ VelaClick, VelaPointer, helper, and status app sources
resources/ public pointer source asset
scripts/ source/release packaging workflow
fixtures/ deterministic test pagesVelaCU is released under the MIT License. See LICENSE and THIRD_PARTY_NOTICES.md.
This server cannot be deployed
MCP server connecting AI agents to 100+ apps (Gmail, Slack, Notion, GitHub) via one-click OAuth.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
A paid remote MCP for AI agent browser approval MCP, built to return verdicts, receipts, usage logs,
Unified MCP Server is a remote MCP connector for AI agents and vertical AI products that provides access to 22,000+ authorized SaaS tools across 400+ integrations and 24 categories directly inside LLMs (Claude, GPT, Gemini, Cohere). Tools operate only on explicitly authorized customer connections, enabling agents to safely read and write against live third-party systems.
