How to kill leftover dev servers after Claude Code

Why agents leave processes running#

Claude Code, Codex CLI, Gemini CLI, Cursor's agent, Aider, OpenCode and other agents run commands to check their work: npm run dev, vite, next dev, uvicorn --reload, a test watcher. Long-running ones go in the background so the agent can keep going. Most of the time they stop when the session does. They survive when:

• the session ended abruptly: the terminal was closed, the agent crashed or the Mac slept mid-session
• the command detached itself from the terminal, as nohup, & with disown, or docker compose up -d do
• a server restarted itself in a child process that its parent didn't take with it

What you notice is “port 3000 is already in use” the next day, or a fan running for no reason. Each forgotten Node server holds a few hundred MB.

Step 1: find what's listening#

lsof -iTCP -sTCP:LISTEN -n -P

COMMAND   PID   USER   FD   TYPE   NODE NAME
node    48211   you    23u  IPv6   TCP *:3000 (LISTEN)
node    48377   you    19u  IPv4   TCP 127.0.0.1:5173 (LISTEN)
Python  48502   you     6u  IPv4   TCP 127.0.0.1:8000 (LISTEN)

lsof -nP -iTCP:3000 -sTCP:LISTEN

-n and -P show addresses and port numbers instead of names, which is faster and easier to read. The PID column is what you need next.

Step 2: check it's a leftover#

ps -o pid,ppid,etime,command -p 48211

  PID  PPID     ELAPSED COMMAND
48211     1    21:14:07 node /Users/you/code/storefront/node_modules/.bin/next dev --turbopack

The command shows the project folder, so you know which repo it belongs to. A PPID of 1 means the process's parent has exited and launchd adopted it: whatever started it, an agent or a terminal, is gone. An ELAPSED time of hours on a server you didn't start yourself is the other giveaway.

ps -axo pid,ppid,etime,command | awk '$2 == 1' | grep -E 'node|vite|next|uvicorn|python|ruby|rails' | grep -v grep

Containers are separate: docker ps lists the ones still running, and docker compose down in the project folder stops a project's set. Activity Monitor shows the same processes, but not their ports.

Step 3: stop it politely, then force if you must#

kill 48211

kill $(lsof -t -iTCP:3000 -sTCP:LISTEN)

kill sends SIGTERM, which lets the server close its connections and remove its own temporary files. Give it a few seconds, check with lsof again, and only if it's still there use kill -9, which can't be ignored but skips that tidy-up.

Making it happen less#

• Ask the agent to stop the servers it started before it finishes, or to use a fixed port so a leftover is easy to spot.
• End sessions from the agent rather than by closing the terminal window.
• Prefer commands that run in the foreground of the agent's own shell over nohup or detached containers.

The faster way: Agent sessions in iKnowMyMac#

iKnowMyMac's Agent sessions section, under Projects, tracks each agent session from the moment the agent starts: the project folder, how long it ran, every process it started and the ports they hold, and what's still running after it ended. Prompts and transcripts are never read.