Claude Code and Codex create VMs, databases and deploys mid-task, then forget them. agent-leftovers reads the logs they already keep on your machine and lists every leftover with the exact command to turn it off. Read-only, local, free.

Hi Product Hunt!

I kept reading the same story: someone lets Claude Code or Codex finish a task, and a week later finds a GPU VM, a 250 GB disk or a batch script hammering a paid API with their key. The agent made it; nobody remembered it.

Your agent already writes down every command it runs in local session logs. agent-leftovers reads them and answers one question: what did my agent create that is still out there, and how do I turn it off?

• 59 rules across GCP, AWS, Azure, Cloudflare, Vercel, Fly, Supabase, Heroku, Terraform and more

• Skips what the agent already deleted and creates that failed

• Flags scripts the agent wrote that call OpenAI, Anthropic or Gemini with your key, even ones left running in the background

• Read-only and fully local: it runs nothing and sends nothing

Try it: npx github:ApahSaroj/agent-leftovers

Honest limit: it only sees the agent's shell. Things made through MCP tools or a web console won't show.

What should I add next: live "is it still running?" checks, or support for Cursor and Gemini CLI?

Update: v0.2 is out, built from today's feedback.

• It now catches resources your agent created through MCP servers (Supabase, Neon, Vercel, Cloudflare, DigitalOcean, E2B and more), not just shell commands.

• New optional --verify flag asks your own cloud CLIs which leftovers still exist. Read-only describe/list calls only, off by default.

Still free, zero dependencies, and nothing leaves your machine unless you pass --verify.

npx github:ApahSaroj/agent-leftovers

Thanks to everyone who pointed out the MCP blind spot. Keep it coming.