From 2 sources to 5: How I upgraded my "idea reality check" MCP server in one day

#ai #mcp #opensource #python

This is a follow-up to Stop Your AI Agent From Building What Already Exists.

v0.1 had a blind spot

Two weeks ago I shipped idea-reality-mcp — an MCP server that checks if your idea already exists before your AI starts coding.

It worked. But it only looked at two places: GitHub and Hacker News.

That meant it missed entire categories. npm has 297,000+ packages related to MCP alone. PyPI has its own ecosystem. Product Hunt has thousands of launched products that never made it to GitHub.

Two sources wasn't enough.

v0.2: five sources, one command

The new version scans GitHub, Hacker News, npm, PyPI, and Product Hunt in parallel:

uvx idea-reality-mcp

Two modes:

quick — GitHub + HN only (fast, same as v0.1)
deep — all five sources at once via asyncio.gather()

Here's a real test with depth="deep":

Metric	Result
Query	"AI trading bot for gold"
reality_signal	82 / 100
duplicate_likelihood	high
sources_used	GitHub + HN + npm + PyPI (PH skipped, no token)
GitHub repos	1,359
HN mentions	254 (across 3 keyword variants)
top_similars	GOLD_ORB (XAUUSD EA, 186 stars)
pivot_hints	"Consider niche differentiator or plugin for existing tools"

An 82 with 1,359 repos means: the space is crowded, but the tool also found a specific competitor (GOLD_ORB) that I could study before deciding whether to proceed.

What changed under the hood

New sources:

npm — hits the registry JSON API (/-/v1/search), free, no auth needed
PyPI — scrapes search HTML with regex fallback (no official search API exists)
Product Hunt — optional GraphQL v2, requires PRODUCTHUNT_TOKEN. No token? Gracefully skipped, zero config stays zero config.

Smarter keyword extraction:

v0.1 just sorted words by length. v0.2 detects compound terms ("machine learning", "web app"), prioritizes technical keywords (React, Docker, FastAPI), and generates a 4th query variant optimized for registry searches.

New scoring weights for deep mode:

GitHub repos:  25%    GitHub stars: 10%
HN mentions:   15%    npm packages: 20%
PyPI packages: 15%    Product Hunt: 15%

If Product Hunt is unavailable, its weight redistributes automatically across the other sources.

The numbers

v0.1: 2 sources, 31 tests
v0.2: 5 sources, 73 tests
Still zero config for basic usage
Still one install command

What's next

The scoring is better but still rule-based. v0.3 will likely add LLM-powered analysis for the "deep" mode — using the raw data from all five sources to generate a more nuanced assessment instead of just a weighted formula.

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