trevor-nichols

Computer Use MCP Server — Development skill for Claude Code

Development community

Have your agents control you Mac (comparable to claude-code computer-use tool).

How to install Computer Use MCP Server

This entry records only its repository, not the path inside it, so there is no exact command to give. Open trevor-nichols/computer-use-mcp-server and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Computer Use MCP Server does

Have your agents control you Mac (comparable to claude-code computer-use tool).

Alternatives in Development

  • Mempalace Init — Set up MemPalace — install the package, initialize a palace, register the MCP server with Cursor, and verify e 58.6k ★
  • Claude Context Mode — An MCP server that sits between Claude Code and these outputs 5.3k ★
  • 04 Fix MCP Server — Prompt 04: Fix MCP Server Build 2.2k ★

README

macOS Computer Use MCP Server

Standalone local `computer-use` MCP server for macOS.

The current capture flow is optimized for MCP clients that can consume inline image attachments. Saved local image paths remain available through `capture_metadata` when a client wants file-based follow-up such as `view_image(imagePath)`.

This repository implements:

  • standalone MCP server
  • stdio transport first
  • session-owned state
  • desktop lock
  • permission and app approval coordination
  • native host seam with Swift bridge services and a selectable input backend (swift, rust, fake)
  • fake mode for development and testing on non-macOS hosts

Capture contract

`screenshot` and `zoom` now do two things directly:

  • attach the captured image inline in the MCP content array
  • return a text item containing captureId=...

Geometry and saved-file metadata are retrieved separately through `capture_metadata`.

The intended consumer flow is:

  1. call screenshot or zoom
  2. inspect the inline MCP image attachment
  3. if geometry or file metadata is needed, call capture_metadata(captureId)

For clients that prefer file-based image viewing, the fallback flow is:

  1. call screenshot or zoom
  2. extract the captureId from the text item
  3. call capture_metadata(captureId)
  4. read structuredContent.imagePath
  5. call view_image(imagePath) if needed

The server does not expose capture delivery through MCP `resources/read`. Screenshot image data is delivered as an MCP `image` content item, not as plain base64 text in `structuredContent`.

What is included

Runnable in this environment

  • TypeScript MCP server
  • protocol-compatible stdio JSON-RPC transport
  • Streamable HTTP transport
  • full current tool surface:
    • request_access
    • screenshot
    • list_displays
    • select_display
    • zoom
    • capture_metadata
    • cursor_position
    • mouse_move
    • left_click
    • right_click
    • middle_click
    • double_click
    • triple_click
    • `lef