agent-computer-use-mcp
Click on "Install 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., "@agent-computer-use-mcptake a screenshot of the current screen"
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.
Computer Use MCP
See the desktop. Understand the interface. Act with evidence.
agent-computer-use-mcp gives an MCP Host a compact, observable way to operate
local Windows applications. It combines semantic accessibility, secure
screenshot delivery, local PP-OCRv6, reliable Unicode input, and explicit
action verification behind a session-bound control lease.
Download the latest release · Read the changelog · Contribute
Preview:
0.xtargets Windows x64 and is evolving quickly. Actions remain Host-approved, application-visible, and fail closed when evidence is insufficient.
From intent to verified action
Natural-language task
↓
Host selects Computer Use from its semantic capability description
↓
Acquire a session-bound desktop control lease
↓
Observe: accessibility first → OCR → secure image understanding when needed
↓
Act through a semantic element, focus receipt, or observation-bound coordinate
↓
Verify the outcome and release control on success, failure, or cancellationThe Agent sees four focused tools:
Tool | Purpose |
| Select a foreground, visible, or restorable application and obtain a bounded lease. |
| Read semantic state, OCR, screenshots, focused controls, truncation, and coordinate metadata. |
| Activate, click, enter Unicode text, set values, or press keys with execution-path evidence. |
| End control and clean up the cursor, overlay, focus receipts, and session state. |
Four additional management tools—health, doctor, installation, and repair—stay with the Host and are not projected into the Agent's normal tool inventory.
Related MCP server: ClawdCursor
Why this implementation
Semantic first, pixels when necessary. Accessible controls are faster and more reliable than guessed coordinates. OCR handles low-latency text; image understanding is reserved for layouts, icons, and complex scenes.
Coordinates carry provenance. Screenshot dimensions, native-window scale, truncation, focus, observation identity, and expiry travel with the result.
Actions explain themselves. Every successful mutation reports the target path, provider path, delivery mode, and any fallback reason.
Chinese input is a first-class path. Coordinate-grounded Unicode entry uses verified native delivery, while semantic fields remain on the driver path.
Screenshots are assets, not file paths. The Host receives controlled MCP image content; the Agent never needs access to a temporary local file.
Control has a lifecycle. Leases are bound to the requesting Host session and are revoked on completion, cancellation, timeout, disconnect, or process shutdown.
No hidden self-update. Startup is offline. Installation, upgrade, downgrade, and rollback are explicit Host or operator actions.
Install
XiaozhiClaw
Open Plugins → Connectors, install Computer Use, review the requested desktop permissions, and enable it. Health, diagnostics, installation details, and repair planning are available from the connector's settings panel.
Any standard MCP Host
Download agent-computer-use-mcp-<version>-win32-x64.tar.gz from the
latest GitHub Release,
verify it against SHA256SUMS.txt, and extract it. The executable is:
artifact/bin/agent-computer-use-mcp.exeExample MCP configuration:
{
"mcpServers": {
"computer-use": {
"command": "C:\\Tools\\agent-computer-use-mcp\\artifact\\bin\\agent-computer-use-mcp.exe",
"args": []
}
}
}The release contains the exact Windows x64 driver, native overlay, ONNX runtime, PP-OCRv6 model pack, inventory, checksums, licenses, and SBOM used by CI. It requires no npm install and performs no network access at startup.
Safety model
Computer Use is intentionally visible and bounded:
the Host owns approval, process lifecycle, installation, and media assets;
the Agent cannot acquire another Host session's lease;
password, payment, credential, private, and denied-window policies fail closed;
the branded frame and cursor are excluded from screenshots, OCR, observations, traces, and benchmark artifacts;
uncertain mutations are never blindly replayed—they require a fresh observation;
repair reports a plan and never downloads or installs implicitly.
Develop
Requirements: Windows x64, Node.js 24, npm 11, and .NET 10 SDK.
git clone https://github.com/astraclawteam/agent-computer-use-mcp.git
cd agent-computer-use-mcp
npm ci
npm test
npm run mcpProtocol and packaging gates:
npm run phase:1.6
npm run phase:1.7
npm run phase:1.8
npm run artifact:windows:build -- --allow-networkReal desktop action gate:
npm run phase:1.4The full test suite is deliberately serial. It covers MCP interoperability, leases, policy, perception, Unicode delivery, cleanup races, overlay exclusion, and immutable Windows artifact assembly.
Release
A vX.Y.Z tag must match package.json, point to a commit reachable from
main, and have a matching CHANGELOG.md section. The release workflow then:
runs the full test and official MCP SDK gates;
assembles and verifies the Windows x64 SEA artifact;
writes
SHA256SUMS.txt;creates the GitHub Release and marks it Latest.
The workflow has no npm or Gitee credentials and publishes only verified GitHub Release assets. See the release pipeline specification and release gates.
Contributing
Real application edge cases are especially valuable: custom-drawn controls, tray restoration, high-DPI scaling, multilingual input, focus transitions, and post-action verification. Start with CONTRIBUTING.md, or open an app smoke report.
Security-sensitive findings should use GitHub private vulnerability reporting.
License
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- Alicense-qualityBmaintenanceMCP server that provides computer control capabilities including mouse movements, keyboard actions, screenshot capture with OCR, and window management through a unified API.Last updated155MIT
- Alicense-qualityAmaintenanceThe local MCP server that gives any AI agent safe desktop control. Provides 6 compact tools (computer, accessibility, window, system, browser, task) for cross-platform GUI automation with ground-truth verification.Last updated86391MIT
- AlicenseAqualityAmaintenanceA low-level computer-use MCP server that exposes raw desktop control (mouse, keyboard, shell, windows, processes, screenshots, etc.) to MCP clients like Claude Code and Codex.Last updated53MIT
- Alicense-qualityBmaintenanceGives AI agents and MCP clients direct control over native desktop apps, Chrome/Electron browsers, and Android devices with screenshots, OCR, accessibility-based element lookup, input simulation, window management, CDP, and ADB in one local server.Last updated120MIT
Related MCP Connectors
Hosted AgentLux MCP server for marketplace, identity, creator, services, and social flows.
Autopilot MCP server for GEO analyses, reports, content, audits, memories and agents.
An MCP server for deep research or task groups
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/astraclawteam/agent-computer-use-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server