AI agents are quietly building a tool registry for each other

So I've been using Claude Code pretty heavily for the past few months and there's this thing that kept happening that drove me insane.

I'd ask it to do something with images. It would decide it needed a CLI tool. Then it would just... install one. Confidently. Except the package didn't exist. It made up a name that sounds like it should exist and ran with it.

The frustrating part is it's not really Claude's fault. Package names are weird and the tooling landscape changes constantly. Something that was the standard tool two years ago might be deprecated now. The model just doesn't know.

I ignored this for a while because the failures were fast, broken install, agent recovers, moves on. But it kept happening and I kept watching it waste time on dead ends.

So I built need. It's a searchable index of 10,000+ CLI tools that agents can query before they try to install something.

npx @agentneeds/need compress PNG images

Instead of guessing, the agent searches. Gets back real tools, real package managers, ranked by what's worked for other agents.

The ranking part is the thing I keep thinking about. Every time an agent installs something through need and it works, that gets recorded. Every failure too. So the index gets more accurate over time, shaped entirely by what agents report back. Not curated. Not maintained. Just self-correcting.

Anyway it works as an MCP server too so if you're using Claude Code or Cursor you can add it in about 30 seconds:

  {
    "mcpServers": {
      "need": {
        "command": "npx",
        "args": ["@agentneeds/need", "serve"]
      }
    }
  }

Give it a try:

npm i -g @agentneeds/need

Website
Github