# open-compute (computer-use) [Health: Active]

**Category:** 🖥️ OS Automation  
**Repository:** https://github.com/ellmos-ai/open-compute-mcp  
**GitHub Stars:** 1  
**npm Downloads (last month):** 216  
**Views:** 9  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/open-compute-computer-use

## Description
Model-agnostic computer-use MCP: screenshot, safety-gated actions, Windows UIA targeting.

## Tools
Capabilities this server exposes over MCP:

- **capture** — Return one-shot observation metadata plus an image (optionally one exact window).
- **do** — Execute a safety-gated action; coordinates require `observation_id` + issued window descriptor/token.
- **tree** — Return UIA elements and a one-shot observation ID for their coordinates.
- **click_name** — Exact-first, ambiguity-safe click in a required issued window, with score/alternatives.
- **invoke** — Exact-first, click-free UIA activation in a required issued window.
- **list_windows** — List stable window/process IDs, exact titles, issued tokens, rects and centers.
- **get_screen_size** — Virtual-desktop geometry + per-monitor breakdown (read-only).
- **watch_dir** — Watch directories for file-system changes.
- **push_status** — Feed-manager status (read-only).
- **rec_replay** — Replay a `.clirec` macro (needs the optional `clirec` package).
- **signal_show** — Show a configurable pre-action color/text countdown, then the mode-colored overlay, with owner/session lease and bounded TTL.
- **signal_hide** — Hide the signal overlay.
- **signal_status** — Owner/session/mode/visible/expires_at + pending abort message.
- **signal_abort** — Ask the human for a short abort reason; the message is returned for the model.
- **chat** — Human→model message about screen content, optionally with screenshot.
- **talk** — Push-to-talk voice note → WAV path (hold key, speak, release; STT/TTS model-side).

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "open-compute-computer-use": {
    "command": "npx",
    "args": ["-y","open-compute-mcp"]
  }
}
```

## Documentation

## What open-compute (computer-use) does

The open-compute (computer-use) MCP server gives an MCP client tools for observing a desktop, identifying application windows, inspecting Windows UI Automation elements, and performing controlled input. The client remains responsible for reasoning: it can capture an image, inspect the returned metadata or UIA tree, and then choose an action.

Its interaction model is intended for local computer use. The launcher communicates over MCP stdio and starts the open-compute engine through the local runtime. The engine connects to the desktop for screen capture, UI Automation, mouse and keyboard input, overlays, and voice-related operations.

## How it works

A typical workflow begins with `list_windows`, `capture`, or `tree`. Window listings provide stable process and window identifiers, titles, rectangles, centers, and issued tokens. Captures return an observation identifier, which is needed when coordinate-based actions are requested. UIA inspection provides elements and a one-shot observation identifier for their coordinates.

Actions are safety-gated. `do` requires the relevant observation identifier and issued window descriptor or token for coordinate operations. `click_name` uses exact-first matching and reports ambiguity, scores, and alternatives. `invoke` activates a UIA element without clicking it. Safety behavior can be configured with `OC_SAFETY_MODE`, while `OC_DENY` can provide a hard action blacklist.

The remaining tools cover supporting workflows. `watch_dir` reports filesystem changes, `push_status` reads feed-manager status, and `rec_replay` replays a `.clirec` macro when the optional `clirec` package is available. Signal tools display or hide a pre-action countdown and overlay, report its state, or request an abort reason from a human. `chat` sends a human message about screen content, optionally with a screenshot. `talk` records a push-to-talk voice note as a WAV path for model-side speech processing.

## Setup and configuration

Install and run the npm launcher in a local MCP client. The repository specifies Node.js 18 or newer and documents an `npx open-compute-mcp` launch path. The launcher starts the Python engine through `uvx`; `OPEN_COMPUTE_MCP_CMD` can be used by the launcher’s command-resolution configuration.

The MCP client and launcher must run on the host whose desktop is being controlled. An interactive Windows desktop session is required for real capture and input. Configuration also includes `OC_SAFETY_MODE`, with documented modes including confirmation, read-only, and allow-all, plus `OC_DENY` for prohibited actions.

## Limitations and notes

The open-compute (computer-use) MCP server is local stdio software, not a hosted service. A deployment flow that cannot access the client host’s desktop cannot provide real screen capture or input. Coordinate actions depend on fresh observation data and issued window credentials, so callers should not treat coordinates as context-free.

UIA-specific targeting is intended for Windows desktop applications. `rec_replay` has an extra dependency, and voice input requires the push-to-talk interaction exposed by `talk`. Read-only tools such as screen geometry, window listing, and feed-manager status do not replace the safety requirements for actions.

_Full upstream README: https://allmcps.com/mcp/open-compute-computer-use/readme_