Skip to content

Is Claude Code slowing down your Mac?

By Junaid, maker of Taskglass · Updated

Summarize this guide with

An AI coding agent itself uses little CPU, but what it starts can use a lot: builds, test runs, dev servers, language servers and MCP servers, often several sessions at once. Those processes show up as dozens of node or python rows in Activity Monitor, and some keep running after the agent quits.

How to find what your AI agents are costing

  1. 1

    Count your sessions

    Each terminal tab running Claude Code, Codex or Gemini CLI is a separate session with its own memory (often 300–500 MB each) and its own MCP servers. Exit the ones you're done with by typing /exit or pressing Ctrl+C twice.

  2. 2

    Look for heavy child processes

    In Activity Monitor, sort by % CPU. node, python, swift-frontend, cargo or esbuild near the top during agent work usually means a build or test run the agent started. It should drop when the task finishes.

  3. 3

    Find leftover servers

    Dev servers and MCP servers can outlive the agent that started them. In Terminal, lsof -i :3000 shows what holds port 3000; kill <pid> stops it.

  4. 4

    Unload local models

    Ollama keeps a model in memory for a while after use. ollama ps lists loaded models and ollama stop <model> frees the memory. In LM Studio, eject the model.

What agents start, and what it costs

ProcessWhat it isTypical cost
claude, codex, geminiThe agent session itself300–500 MB, little CPU while waiting
node / python … mcpMCP servers the agent connects to20–100 MB each, idle most of the time
next build, tsc, cargo, swift-frontendBuilds the agent runsSeveral cores for seconds to minutes
next dev, vite, uvicornDev serversMemory that stays until stopped
sourcekit-lsp, tsserverLanguage serversTens to hundreds of MB
ollama, llama-serverLocal modelsSeveral GB of memory while a model is loaded
caffeinateKeeps the Mac awake while the agent worksNone, but your Mac won't sleep

Why does Claude Code use so much memory?

Each session holds its conversation and tools in memory, and each one starts its own MCP servers. Five open sessions with four MCP servers each is 25 processes. Close sessions you aren't using, and remove MCP servers you never call.

Why won't my Mac sleep while Claude Code is open?

Claude Code runs caffeinate while it's working so a long task isn't interrupted. It stops when the agent goes idle; if your Mac still won't sleep, check for sessions that are stuck mid-task.

How do I know if an agent left processes behind?

Look for node or python processes whose parent is launchd (PID 1) after you closed the terminal: in Terminal, ps -o pid,ppid,command -p <pid> shows the parent. Dev servers and MCP servers found this way are usually safe to stop.

How Taskglass helps

Taskglass's Agents tab shows one row per agent session (Claude Code, Codex, Gemini CLI, Cursor, Ollama, LM Studio and more) with its project folder, whether it's working or idle, its CPU, memory and power, and every process it started. It flags dev and MCP servers left behind, with a Clean Up button, and charts today's CPU time per agent.

See how it works

Questions

Does Claude Code use a lot of CPU?

The agent itself uses little CPU, mostly while it streams a response. The heavy CPU comes from the builds, tests and tools it runs on your behalf.

Is it safe to kill node processes on a Mac?

Usually, if you know what they are: most are dev servers, build tools or MCP servers. Quitting one an agent is using makes that tool fail until it restarts. Check the command first with ps -o command -p <pid>.

How much memory does Ollama use?

Roughly the size of the loaded model: a 7–8B model takes about 5 GB, larger ones much more. ollama ps shows what's loaded and for how long it will stay.

Processes you might see

More guides

See the cause on your own Mac.

Taskglass shows which app is slowing, heating or draining your Mac, and explains it in plain English. One-time purchase, 30-day money-back guarantee.