MCP SERVER · NPX, 10 SECONDS
Your agent can code. Now it can see.
One command gives Claude Code, Claude Desktop or Cursor three new tools: score any URL with a predictive attention model, score local mockups, and battle two pages. Your agent ships pages a human eye will actually follow.
> build the pricing page, then make sure the eye lands on the Pro plan
⏺ Writing app/pricing/page.tsx…
⏺ heatpoints · analyze_page(https://preview-x7f2.vercel.app/pricing)
ATTENTION SCORE 54/100 · hierarchy 31/100 · primary zone: Top-left (your logo)
⏺ The eye lands on the logo, not the Pro card. Adding contrast to Pro, muting the side plans…
⏺ heatpoints · analyze_page(https://preview-x7f2.vercel.app/pricing)
ATTENTION SCORE 73/100 · hierarchy 78/100 · primary zone: Center (the Pro card)
✓ Shipped at 73/100 — the Pro plan now owns the first glance.
Deterministic model: same pixels, same score. The deltas your agent measures are real.
Install in 10 seconds
claude mcp add heatpoints -- npx -y heatpoints-mcp # full metrics (free account = 10 analyses/month): claude mcp add heatpoints \ -e HEATPOINTS_API_KEY=hp_xxx -- npx -y heatpoints-mcp
No key = free scanner mode (3/day, score + top insight). Free account = 10 full analyses/month.
The three tools
analyze_page{ url }
Attention score /100 + focus clarity, hierarchy, coverage, primary gaze zone + guidance. Works on staging and preview URLs — score before merging.
analyze_image{ path }
Same scoring for a local screenshot, Figma export, thumbnail or ad creative. The agent can score designs that aren't deployed anywhere yet.
compare_pages{ url_a, url_b }
Attention battle: two URLs, one winner. Redesign vs current, you vs competitor, A vs B.
The eval layer for pages, in your agent's toolbox.
Free account = API key + 10 analyses/month. No card.
Get your API key →