Stop Your AI Agent From Building What Already Exists

#ai #mcp #opensource #productivity

I wasted 6 hours building something that already had 847 GitHub repos

Last month I told Claude: "Build me an AI-powered food recommendation engine."

It did. Beautifully. Clean code, tests passing, README done.

Then I searched GitHub. 847 repos. Twelve of them had over 100 stars. Some were updated that same week.

I had just mass-produced another clone.

The problem isn't coding speed — it's decision correctness

Every AI coding tool in 2026 makes you build faster. Cursor, Claude Code, Copilot — they're all racing to write code at the speed of thought.

But none of them ask the one question that matters:

Should you build this at all?

So I built a reality check that lives inside the workflow

Idea Reality MCP is an MCP server — not a website, not a dashboard, not another SaaS validator.

It's a tool your AI agent calls before it starts building.

Install:

uvx idea-reality-mcp

Add to Claude Desktop config:

{
  "mcpServers": {
    "idea-reality": {
      "command": "uvx",
      "args": ["idea-reality-mcp"]
    }
  }
}

Then just tell Claude: "Check if this idea already exists before we build it."

What it returns

{
  "reality_signal": 82,
  "duplicate_likelihood": "high",
  "evidence": [
    {"source": "github", "type": "repo_count", "count": 847},
    {"source": "github", "type": "high_star_repos", "count": 12},
    {"source": "hn", "type": "mention_count", "count": 34}
  ],
  "top_similars": [
    {"name": "food-rec-ai", "stars": 2340, "updated": "2026-02-18"}
  ],
  "pivot_hints": [
    "Space is saturated. Consider vertical-specific targeting.",
    "Most existing tools are generic — niche wins."
  ]
}

An 82 means: stop. Research first. Pivot or differentiate.

A 15 means: green light. The space is open.

Why MCP, not a website?

Idea validators already exist as websites — IdeaProof, ValidatorAI, DimeADozen, FounderPal. There are dozens.

But they all require you to leave your workflow, open a browser, type your idea, wait for a report, then go back to coding.

That's the wrong architecture. The check should happen inside the moment you decide to build.

MCP makes this possible. Your AI agent calls idea_check() the same way it calls any other tool. No context switch. No extra tab. No friction.

IDEA → reality check → BUILD

Instead of:

IDEA → BUILD → discover competition → regret

The scoring is intentionally simple

v0.1.0 uses three signals:

GitHub repo count (keyword search across 3 query variants)
GitHub star/recency (are top repos actively maintained?)
Hacker News mentions (has this been discussed in the last 12 months?)

Weighted formula: (github_repos × 0.6) + (github_stars × 0.2) + (hn_mentions × 0.2)

Is it perfect? No. Is it better than zero signal? Absolutely.

What's next

This is v0.1.0. The roadmap includes ProductHunt scanning, deeper keyword extraction, and an opt-in "idea memory dataset" — a global record of what people have checked and what happened next.

If you're building with Claude, Cursor, or any MCP-compatible tool: