tell-ai-ui-critic

tell-ai-ui-critic

An MCP server that captures rendered UI, diagnoses design tells and consistency drift, and generates redesign diffs from natural language art direction, all within Cursor.

Category
Visit Server

README

<div align="center">

Tell

Every AI-built UI has a tell.

Tell is an open-source design critic that reads the rendered UI of your product, names exactly what makes it look AI-generated, and helps you art-direct a distinctive redesign — without leaving Cursor.

Quick start · Features · How it works · Use it in Cursor · Project structure

License: MIT Built for Cursor TypeScript Next.js 14 MCP

<br/>

Tell capturing a page, naming its tells, and revealing a redesign

The shipped README demo asset lives in docs/media/tell-demo.gif. Raw Playwright recordings can be regenerated under output/playwright/, which stays ignored as local build output.

</div>


The problem

You can ship a whole product in a weekend now. That's the good news.

The bad news is that most of what gets shipped looks identical. Inter everywhere. One purple gradient in the hero. A soft shadow on every card. Rounded corners at exactly 8 pixels. It isn't ugly — it's forgettable. And as you and your AI agent keep iterating, the surface quietly drifts: six almost-identical grays, focus rings that only half the app bothers with, an empty state nobody designed.

You can feel that something is off, but you can't name it. "Add more whitespace" isn't a direction. Hiring a designer for a day is $800 and a two-week wait. Your demo is tomorrow.

Tell fixes the part that's actually hard: knowing what's wrong, and showing you a better direction you can ship yourself.


What Tell does

Tell looks at your product the way a person does — it opens the page in a real browser and reads what actually renders, not your source code. Then it does four things:

  1. Names the tells. It points at the specific patterns that read as generic, with evidence you can see on the screenshot.
  2. Catches the drift. It flags where your design has quietly become inconsistent across the surface.
  3. Uses taste, not lint rules. It tells the difference between a lazy default and a deliberate choice, and explains why.
  4. Closes the loop in Cursor. You describe a direction in plain English (or with your voice), and Tell drafts the redesign as a diff you apply right inside your editor.

No new dashboard to babysit. No design handoff. No leaving your workflow.


Features

Feature What it does
🔍 Rendered capture Opens your URL in a headless browser and records a full screenshot plus a computed-style fingerprint — fonts, colors, shadows, radii, spacing, contrast, and interactive states.
🎯 8 genericness detectors Deterministic checks for the classic AI tells: SystemFontTell, GradientCrutchTell, ShadowEverywhereTell, RadiusMonotoneTell, AcidAccentTell, EmojiChromeTell, CenteredEverythingTell, GrayMushTell.
📉 6 consistency-drift detectors Catches the slow decay: TokenBypass, NearDuplicateValues, FocusRingInconsistency, TypeScaleDrift, SpacingChaos, StateGap.
🧠 Taste engine Classifies every finding as generic, drift, or intentional with a plain-English rationale and a confidence score. A reflection pass rejects any reasoning that contradicts the measured facts.
🎙️ Voice art-direction Say "warmer, more editorial, less shadow" and Tell breaks the instruction into action items, maps it to a direction preset, and refines with Gemini when available. Text presets keep the demo safe when mic access fails.
🪄 Redesign as a diff Turns a chosen direction into a unified diff. It never auto-applies — you stay in control.
↔️ Before/after seam A draggable diagonal reveal between your captured page and a live reconciliation — same content, restyled from detected tokens with contrast ratios called out.
🐙 GitHub repo setup Paste github.com/owner/repo and Tell clones it, reads the README, installs deps, boots the dev server on a reachable free port, and captures localhost automatically.
📄 Multi-page scanning Discovers routes from the captured snapshot; scan each page to catch drift that only shows on some routes.
🔌 Cursor MCP server Run the whole pipeline from Cursor chat with tell_capture, tell_diagnose, tell_redesign, and tell_apply.

How it works

Tell is one pipeline, shared by both the web app and the Cursor MCP server. Everything up to the taste step is deterministic — no model, no network — so results are reproducible run to run. The model is only used for judgment and for drafting diffs.

flowchart LR
    url["Your URL or a\nlocal route"] --> capture["Capture\nrendered UI"]
    capture --> fingerprint["Build a\ndesign fingerprint"]
    fingerprint --> tells["Detect\ngenericness tells"]
    fingerprint --> drift["Detect\nconsistency drift"]
    tells --> taste["Taste engine\n(generic / drift / intentional)"]
    drift --> taste
    voice["Voice / text\nart-direction"] --> taste
    taste --> report["Tell Report +\nbefore/after seam"]
    report --> contrast["Contrast floor +\nreachable-state feedback"]
    taste --> diff["Redesign diff"]
    diff --> cursor["Apply in Cursor\nvia MCP"]

Why deterministic-first matters: the parts that must be trustworthy — reading the page and measuring it — never hallucinate. The model only weighs in where judgment is genuinely needed, and even then its rationale is checked against the hard facts before you see it.


Quick start

You'll need Node 20+ and pnpm 9+.

git clone <your-repo-url> tell
cd tell
pnpm install

Then start the Tell app:

pnpm dev           # the Tell app → http://localhost:3000

The one-click way — point Tell at a GitHub repo. Paste a repo URL (e.g. github.com/owner/app) into the bar and hit Set up & run. Tell clones it, reads the README + package.json to figure out how to start it, installs, boots the dev server on a free port, verifies the URL responds, and captures localhost automatically. The UI blocks duplicate clicks, hides stale captures while work is running, and surfaces success or failure in plain language. If the run steps aren't clear, Tell shows what it found from the README and asks you to start it and paste the URL — no guessing.

The direct way — capture any URL. Paste a live URL (your deployed app, or a local http://localhost:3000) and hit Capture.

Either way: read the report, walk every page from the Pages strip, drag the before/after seam, art-direct a new direction, draft the diff, and copy it back into Cursor.

<details> <summary><strong>Prefer the seeded sample app?</strong></summary>

Run the deliberately bland fixture in a second terminal and capture it:

pnpm dev:fixture   # a deliberately bland sample app → http://localhost:3001

</details>

No time to wait on live capture? Tell ships with a committed report at fixtures/reports/tell-report.json and loads it automatically as an offline fallback, so the demo always works.

<details> <summary><strong>Environment variables (optional)</strong></summary>

Copy .env.example to .env and fill in what you have. Tell runs fully without keys — the taste and redesign steps simply fall back to their deterministic behavior.

GEMINI_API_KEY=            # powers richer taste/voice direction parsing
CURSOR_API_KEY=            # enables Cursor-SDK-backed redesign drafts
CURSOR_MODEL=composer-2.5  # optional; defaults to composer-2.5
CURSOR_AGENT_TIMEOUT_MS=75000
ANTHROPIC_API_KEY=         # reserved for richer source-aware diffs (optional)

</details>

<details> <summary><strong>Handy scripts</strong></summary>

pnpm test              # run the golden detector tests
pnpm typecheck         # strict type-check across every package
pnpm capture:fixture   # capture the sample app to a fresh report
pnpm diagnose:fixture  # diagnose the sample app from a capture

</details>


Use it in Cursor

Tell registers itself as an MCP server, so you can drive the whole pipeline from Cursor's Agent chat.

  1. Open this repo in Cursor — the tell server is already registered in .cursor/mcp.json.

  2. In Agent chat, ask for a diagnosis in plain English:

    "Run tell_diagnose on http://localhost:3001 and draft an editorial redesign."

  3. Review the findings and the proposed diff, then apply it — Tell hands you the patch, but you decide what lands.

Available tools

Tool What it returns
tell_capture The rendered screenshot + computed-style evidence for a URL.
tell_diagnose The full report: findings, taste verdicts, and the Tell score.
tell_redesign A redesign proposal (patch text) for one finding or the whole report.
tell_apply The unified diff plus instructions — it never writes files for you.

Project structure

Tell is a pnpm monorepo. Each package has one job.

tell/
├── packages/
│   ├── schema/      # zod contracts shared by everything
│   ├── core/        # capture + fingerprint + detectors (pure, no model)
│   ├── taste/       # taste engine + voice/text art-direction
│   ├── redesign/    # turns a direction into a diff
│   └── mcp/          # the Cursor-facing MCP server
├── apps/
│   └── web/         # the Tell app: report, before/after seam, voice director
├── fixtures/
│   ├── generic-app/ # a deliberately bland app used as demo input
│   └── reports/     # a committed report artifact for offline demos
└── docs/            # design system and reference notes

Tech stack

Layer Choice
Language TypeScript (strict)
Monorepo pnpm workspaces
Capture Playwright / Chrome DevTools Protocol
Web app Next.js 14 + Tailwind CSS
Taste reasoning Google Gemini with deterministic fallback
Redesign diffs Deterministic reconciliation CSS; Anthropic-ready interface
Editor integration Model Context Protocol (stdio)
Validation zod
Tests Vitest (golden detector tests)

Roadmap

  • [x] Deterministic capture → fingerprint → 14 detectors
  • [x] Taste engine with reflection + safe fallback
  • [x] Tell Report with draggable before/after seam
  • [x] Voice art-direction with text-preset fallback
  • [x] Cursor MCP server (capture / diagnose / redesign / apply)
  • [x] Live URL capture (any reachable HTTP URL, with offline artifact fallback)
  • [x] GitHub repo setup — clone, install, run, and capture localhost
  • [x] Token-grounded live reconciliation in the before/after seam
  • [x] Contrast-grounded reconciliation with readable text/control pairings
  • [x] Multi-page route discovery and per-page scanning
  • [ ] Per-state evidence thumbnails (hover, focus, error) in the report
  • [ ] A shareable, hosted report link

Contributing

Contributions are welcome. The short version:

  1. Fork the repo and create a branch: git checkout -b feature/your-idea
  2. Make your change and keep it green: pnpm typecheck && pnpm test
  3. Open a pull request describing the why, not just the what.

New detectors are the highest-leverage contribution — they live in packages/core/src/detectors and each ships with a golden test.


License

Released under the MIT License.

The sample app under fixtures/generic-app/ is a deliberately bland demo target used only as input for Tell — it is not part of the product. See CONTRIBUTIONS.md for the full breakdown of what's original work versus demo input.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured