What's Running MCP server

whats-running-mcp

Ground truth for agent fleets. An MCP server that tells your AI agent what is actually running on the machine — live agent sessions, listening TCP ports, loaded daemons, system load — read straight from the OS (ps / lsof / launchctl / df), never from docs, transcripts, or memory files.

Why

If you run more than one coding-agent session (Claude Code, Codex, Aider…), you've seen this failure mode: an agent "health-checks" a server that died weeks ago, or confidently reports a daemon as live because an old transcript said so. Model memory is a cache with no invalidation. The OS is the source of truth.

This server started as a shell script (whats-running) I wrote after my own fleet of Claude Code sessions kept reporting stale state as live. Wrapping it in MCP means every session grounds itself the same way, automatically.

Tools

Install

Claude Desktop (one-click, no terminal): download the latest whats-running-mcp-.dxt from Releases and open it with Claude Desktop (double-click, or Settings → Extensions → Install Extension…). The server and its dependencies ship inside the bundle — no npm, no Node install.

Register with Claude Code (available in every session):

claude mcp add --scope user whats-running -- npx -y whats-running-mcp

Or in any MCP client config:

{
  "mcpServers": {
    "whats-running": {
      "command": "npx",
      "args": ["-y", "whats-running-mcp"]
    }
  }
}

From source

git clone https://github.com/stcmain/whats-running-mcp.git
cd whats-running-mcp
npm install && npm run build
# then point your client at node /path/to/whats-running-mcp/dist/index.js

Published as whats-running-mcp on npm and as io.github.stcmain/whats-running-mcp in the MCP Registry.

Configuration

Design notes

• Read-only by construction. Every command is a fixed binary with fixed flags via execFile — no shell, and nothing derived from model input is ever executed. The only model-controlled inputs are a boolean and a substring filter applied in-process.
• Local information disclosure is the point. The model sees process names, pids, working directories, and ports of your own machine. Don't attach this server to an untrusted client.
• Best-effort, never blocking. Commands time out at 10s and degrade to empty results rather than erroring the whole snapshot.

Platform: macOS first (my daily driver), Linux best-effort (systemd --user in place of launchctl).

Related free tools

Same idea — read-only, MIT, no telemetry — pointed at different blind spots:

Shortened. The full README is on GitHub.