Skip to main content

See Corexi run on your own product, live, in 15 minutes.

Book a demo
corexi
How our AI works

One engine. Every signal, verified.

Corexi does not guess. It walks your live product, reads it with vision AI, checks every accessibility claim with axe-core, fuses your behavioural data to find the friction that costs conversions, and hands over the fix. Here is how.

  • Live in production
  • 10 MCP tools across 6 IDEs
  • Listed on cursor.directory
The pipeline

From login to fix-ready code.

Every run takes the same five steps, in order. The walk decides what gets read; the rest turns each screen it reaches into findings you can act on.

  1. 00Walk~minutes

    Autopilot signs in with the login you saved and walks the product like a careful new user: it looks at the screen, decides the next click, acts, checks what changed, and records the step. Menus open on hover, forms get sample values, loops are cut short, and it never saves, submits, pays, deletes, invites, signs out, or leaves your domain. Every screen it reaches goes to the pipeline below.

  2. 01Capture~3–8s

    Playwright renders your product at desktop (1440px) and mobile (390px). Multi-state probes test focus indicators, modals and form validation. Bot-protection bypass with fallback chains. Not a static screenshot: a living render.

  3. 02Analyze~8–15s

    Claude Sonnet reads the rendered UI like a UX expert across 9 categories. axe-core runs in parallel, cross-checking every accessibility finding against WCAG 2.2: 27 rule IDs verified. No hallucination passes unchecked.

  4. 03Fuse~1–3s

    Behavioral signals from up to 7 analytics providers merge with visual findings. The Pulse UX Engine computes the weighted UX Score, neurodiversity scores, and ties every finding to conversion impact. Each scan accumulates context, compounding accuracy over time.

  5. 04Ship~2–5s

    Every finding ships with fix-ready code that respects your design tokens. Choose your output: direct fix code, IDE-tailored prompts, bulk export, or MCP delivery: plus a report with annotated screenshots and bounding boxes.

What makes it different

Six things no other tool combines.

  • Verified, not guessed

    Every accessibility finding is cross-checked against axe-core. Verified claims carry a badge; LLM-only claims stay flagged. You always know the confidence level.

  • Gets smarter over time

    Each scan feeds the engine six signals: tech stack, design tokens, growth goal, UI state, behavioral snapshots and prior history. The layer learns your product.

  • Design-token-aware

    Upload CSS variables, Tailwind theme or Style Dictionary. Fix code outputs var(--your-token), not hardcoded values. Colors matched by perceptual distance (ΔE ≤ 6).

  • Not a black box

    Every score traces to a named research source: Nielsen Norman, Baymard, WCAG 2.2, Fitts, Miller, BDA, CUD. 200+ peer-reviewed sources, all auditable.

  • Proactive, not just reactive

    get_ux_rules gives the agent UX patterns before writing code; review_code checks the result before commit. No other tool offers pre-coding UX guidance inside the IDE.

  • Neurodiversity as first-class

    Color Vision Deficiency and Dyslexia Readability are scored as dedicated categories. WCAG covers compliance; Corexi covers whether real people can convert.

From the research base

A few of the 200+ sources behind every score.

The base is not frozen: new peer-reviewed research is folded into the engine with every release.

  • Nielsen Norman GroupUSABILITY

    10 heuristics, still the usability baseline

  • WCAG 2.2 · W3CACCESSIBILITY

    Accessibility success criteria · 2023

  • ISO 9241-110ERGONOMICS

    Interaction principles standard

  • Baymard InstituteCONVERSION

    E-commerce UX, 150k+ hours of testing

  • Fitts's LawMOTOR

    1954 law → today's 44px tap-target rule

  • Miller's LawCOGNITIVE

    7±2 memory limit → cognitive-load scoring

By the numbers

Measured, not marketing.

  • 200+
    Research sources
    NN/g, Baymard, WCAG 2.2, Fitts, Miller, BDA, CUD
  • 9
    Core categories
    + neurodiversity & CVD lenses
  • 7
    Analytics providers
    GA4, PostHog, Clarity, Hotjar, Mixpanel, Amplitude, Firebase
  • 10
    MCP tools
    Across 6 IDEs, live on cursor.directory
  • 27
    axe-core rules
    Cross-checked on every accessibility finding
  • 92.9%
    Eval recall
    Must-detect findings, CI-gated baseline
Honest about limitations

What it can and cannot do.

The questions teams ask once they have seen a run.

Dynamic content
Autopilot opens menus, modals and forms as it walks, and each screen is captured once it has settled. Content that only appears after long scrolling or many lazy loads may take more than one run to reach.
AI accuracy
Strong on structural issues (contrast, spacing, hierarchy). Subjective design quality is harder: every finding shows evidence so you can verify. Eval baseline: 92.9% recall.
Behind a login
Give it a test account, or a ready session copied from your own signed-in browser when the login has 2FA or a captcha. Autopilot never solves captchas, never saves, pays, deletes, invites or signs out, and never leaves your domain.
Data handling
Cookieless capture, no tracking scripts required. Screenshots processed in memory, encrypted at rest. Analytics via read-only OAuth. We never modify your analytics or codebase.

Try it on your product.

Give it your product's URL. We set up the login with you once; from then on it goes round on its own, after every release.

14-day free trial, set up with us on a 15-minute call · no credit card