Computer Load for ChatGPT and Codex

Русская документация · MIT license · Validation

A local-first MCP plugin that helps explain why a computer feels slow. It takes a short CPU, memory, disk and network sample, then displays the busiest processes in an interactive dashboard. It runs on the computer being diagnosed.

This is an independent community project, not an official OpenAI product or a ChatGPT directory listing. The dashboard and current user-facing messages are in Russian. macOS has been exercised locally; real Windows UI operation and the widget inside a native ChatGPT/Codex chat remain unverified.

What it does

• Samples for 2–10 seconds and returns at most 50 processes (defaults: 3 seconds, 25 processes).

• Shows CPU activity, available RAM, swap growth, free disk space, disk throughput and network throughput, with labels when a metric is unavailable.

• Keeps reports and process-stop proposals in memory, with expiry and bounded request queues.

• Supports MCP over stdio and a loopback HTTP endpoint, plus an MCP Apps widget.

• Offers a separate, explicitly confirmed close request for a small allowlist of media applications. Diagnostics alone never close an application.

The project does not measure GPU, temperature, native macOS memory pressure or disk health, and does not scan files or identify malware. A short sample cannot prove sustained overload. Per-process RSS values must not be summed as exact machine RAM usage.

Related MCP server: MCP System Monitor Server

Run locally

Requirements: Node.js 20+, Python 3.9+, npm and access to npm/PyPI during setup. On Windows, install the Python launcher (py). Run as a regular user.

npm run setup
npm start

Open http://127.0.0.1:8798 and choose the diagnosis button. Keep the terminal open while using the agent. Alternatively, launch Start-macOS.command or Start-Windows.cmd; the first launch prepares project-local dependencies.

If the macOS launcher was downloaded without execute permission, run chmod +x Start-macOS.command from the project directory before opening it.

The standalone panel works without an AI account or API key. Exporting its JSON report and sharing that report is a separate user action. Installation downloads dependencies; normal sampling does not call an external API.

Connect an MCP client

Prepare dependencies first. Configure the client to launch:

{
  "command": "node",
  "args": ["/absolute/path/to/computer-load-chatgpt/server/main.mjs", "--stdio"],
  "cwd": "/absolute/path/to/computer-load-chatgpt"
}

Only one instance can use the default ports at a time. A stdio instance starts its own local confirmation panel, so stop a standalone instance that you started before launching the same plugin through a client. Custom instances can use COMPUTER_LOAD_MCP_PORT and COMPUTER_LOAD_LOCAL_PORT with different loopback ports.

The repository includes plugin.json, mcp.json, a local marketplace manifest, and a model skill. See the desktop and private-tunnel setup for the integration tested during development. Client support and tunnel access depend on the host and account. OpenAI usage is billed by the host; this project does not provide model access or reset account limits.

Privacy and process controls

Reports contain process names, PIDs and resource metrics. Command-line arguments, file contents, usernames, executable paths, hostnames and network addresses are not included. Executable identity and ownership are checked locally. Process names are untrusted input and must never be treated as instructions.

The control panel binds to 127.0.0.1:8798, verifies local address, Host, Origin and CSRF, and cannot be embedded in a frame. The MCP endpoint binds to 127.0.0.1:8797. Do not expose either port to the public Internet. Public authentication and multi-user agent isolation are not implemented.

A human must confirm a proposed close in the local panel by entering the exact application name and confirming saved work. The agent rechecks identity, owner, creation time, policy and expiry before acting. macOS uses SIGTERM; Windows uses WM_CLOSE for a visible window. There is no forced-kill fallback. Even SIGTERM can lose unsaved work. System, business, remote-access, security, unknown and other-user processes are protected or left to manual handling. ChatGPT and Codex are protected. See SECURITY.md for the full boundary.

Development and validation

npm run setup
npm run check
npm test
npm run verify
npm run verify:stdio
npm run pack:source

The verification commands use separate loopback test ports. Tests use fake processes and one disposable child created by the test itself; they do not close user applications. Live diagnostic reports and screenshots stay in ignored artifacts/. Source archives are built from an explicit file allowlist into ignored dist/.

See VALIDATION.md for dated results and unverified cases, CONTRIBUTING.md for contributions, and CHANGELOG.md for release notes. The package is marked private to prevent accidental npm publication; its source is available under MIT.

Architecture and attribution

agent/ contains Python/psutil sampling and the process policy; server/ contains the Node MCP bridge and local confirmation panel; public/ contains the shared dashboard; skills/ contains host guidance; tests/ and scripts/ contain verification and packaging tools.

Transport/resource wiring was adapted from the OpenAI MCP App Basics example and Apps SDK quickstart. Dependencies retain their own licenses; see THIRD_PARTY_NOTICES.md.