I use Claude Code and Cursor daily. The pattern I kept hitting: describe an idea → AI starts coding → hours later I search GitHub and realize there are already 10+ mature projects doing the same thing.

The fix was simple: search BEFORE coding. But I wanted it automated — not a manual GitHub search, but something the AI agent does by default as part of its workflow.

Technical details: - Keyword extraction uses a 3-stage dictionary pipeline (90+ intent anchors, 80+ synonym expansions, 150+ Chinese term mappings). No LLM needed for the MCP path — fast, free, deterministic. - Scoring is intentionally simple and explainable: weighted average across sources, not ML. You can read the formula in scoring/engine.py. - All sources are searched in parallel with asyncio.gather. Deep mode hits 5 APIs concurrently. - Zero storage — nothing is logged or saved. Protocol, not SaaS.

The most useful integration pattern: add "always run idea_check before building anything new" to your CLAUDE.md or Cursor rules. The agent will automatically check before writing code.

Honest limitation: dictionary-based keyword extraction has a coverage ceiling. Niche domains outside tech (e.g., "pet acupuncture booking app") may get less precise results. The web demo at mnemox.ai/check uses LLM extraction for better coverage.

Would love feedback on the scoring formula — it's the part I'm least confident about. Is a single 0-100 number useful, or would you prefer something more granular?