UI Perception Engine

A perception layer for AI agents — human-like understanding of web interfaces.

Fuses structural (DOM + a11y tree + CSS), visual (three-tier vision pipeline), and temporal data into a unified UI Scene Graph that an LLM can reason over fluently. Exposed as an MCP server so Claude can navigate and inspect any live URL.

What It Does

• Navigates to any URL in a real Playwright browser
• Extracts a compact, LLM-readable scene graph from the live DOM
• Detects UI elements visually via OmniParser V2, understands layout via Qwen3-VL, and performs deep UX analysis via Claude Vision
• Tracks UI transitions and diffs between states
• Predicts affordances (what you can interact with, and what happens when you do)
• Exposes everything as MCP tools that Claude can call directly

MCP Tools

Tool	Description
navigate	Navigate to a URL and return the initial scene graph
get_scene	Re-capture the current scene (compact text or full JSON)
get_affordances	List interactive elements ranked by priority
act	Execute browser actions: click, type, scroll, hover, keypress, navigate, wait
get_console_logs	Return captured browser console messages
get_network_errors	Return failed network requests
get_screenshot	Capture a screenshot and return it as an image
detect_elements	Run element detection (OmniParser or Claude Vision fallback)
analyze_visual	Visual understanding via Qwen3-VL: hierarchy, contrast, spacing, UX
compare_states	Diff current vs previous scene graph — shows what changed
watch	Start real-time keyframe capture (CDP screencast + perceptual hashing)
stop_watch	Stop keyframe capture and return summary of changes

Three-tier vision pipeline (visual=true)

Both navigate and get_scene accept an optional visual: true parameter. When enabled, the engine runs a three-tier vision pipeline:

• Tier A: OmniParser V2 — Fast element detection (~0.8s). A YOLOv8 + Florence-2 model running as a Python sidecar on port 8100. Detects buttons, inputs, images, icons, and other UI elements with bounding boxes and labels.
• Tier B: Qwen3-VL via Ollama — Visual understanding (~2-4s). Analyzes the screenshot with detected element context to assess visual hierarchy, contrast issues, spacing problems, affordance clarity, and state indicators.
• Tier C: Claude Vision API — Deep UX analysis (~3-5s, on-demand only). Provides detailed qualitative analysis when requested via analyze_visual or depth: 'deep'.

Graceful degradation: Each tier skips silently if its backing service is unavailable. The system works with any combination of services running — from all three tiers down to structural-only analysis with no vision services at all.

Use detect_elements for fast Tier A detection only, or analyze_visual for the full Tier A + B pipeline.

act action types

Type	Parameters
click	x, y
clickSelector	selector
type	text, selector (optional)
scroll	direction (up/down), amount (optional)
hover	x, y
wait	ms
navigate	url
back	—
pressKey	key

Installation

From npm

npm install -g ui-perception-engine

Or use without installing:

npx ui-perception-engine

Install Playwright browsers (first time only):

npx playwright install chromium

From source

git clone https://github.com/dirkknibbe/uipe.git
cd uipe
pnpm install
pnpm build

Claude Code MCP Configuration

Add to your Claude Code MCP config (~/.claude/mcp.json or project .mcp.json):

{
  "mcpServers": {
    "ui-perception-engine": {
      "command": "npx",
      "args": ["ui-perception-engine"]
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "ui-perception-engine": {
      "command": "uipe"
    }
  }
}

Or from a local clone:

{
  "mcpServers": {
    "ui-perception-engine": {
      "command": "node",
      "args": ["/path/to/uipe/ui-perception-engine/dist/src/mcp/index.js"],
      "env": {
        "OLLAMA_URL": "http://localhost:11434",
        "OLLAMA_MODEL": "qwen3-vl:8b",
        "OMNIPARSER_URL": "http://localhost:8100",
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Environment variables

Variable	Purpose
ANTHROPIC_API_KEY	Claude Vision API — detection fallback + deep analysis (Tier C)
OLLAMA_URL	Ollama server URL (default: http://localhost:11434)
OLLAMA_MODEL	Vision model name (default: qwen3-vl:8b)
OMNIPARSER_URL	OmniParser V2 sidecar URL (default: http://localhost:8100)

See .env.example for the full list of configurable variables including frame capture, browser, and temporal settings.

Local Vision Services

The three-tier vision pipeline uses two local services. Both are optional — the system degrades gracefully without them.

Ollama (Tier B — visual understanding):

# Install Ollama: https://ollama.com
ollama pull qwen3-vl:8b
ollama list                    # verify model is available
# Ollama serves on http://localhost:11434 by default

OmniParser V2 (Tier A — element detection):

OmniParser runs as a Python FastAPI sidecar on port 8100. See the Local Vision Handoff doc section 5 for full setup instructions.

# Quick check if OmniParser is running:
curl -s http://localhost:8100/health

Without local services: If neither Ollama nor OmniParser is running, visual=true falls back to Claude Vision API (requires ANTHROPIC_API_KEY). If no vision service is available at all, the engine uses structural-only analysis (DOM + a11y tree).

Using with the live-deployment-check Skill

The live-deployment-check skill pairs directly with this MCP server to visually verify a deployed site or app — catching broken images, empty routes, stuck spinners, and placeholder text that only surface in a real browser.

Workflow

1. navigate(url)            → load the page, get initial scene
2. get_scene()              → re-capture after JS hydrates (critical for SPAs)
3. get_console_logs()       → check for JS errors (type="error")
4. get_network_errors()     → check for failed API/resource requests
5. Scan scene output        → look for broken signals (see below)
6. act() on nav links       → walk routes, verify each one loads
7. Report findings          → list what's working and what's broken

Common Signals in Scene Output

# Broken image:
img[img]:"broken"

# Empty SPA route (component failed to load):
router-outlet[element]          ← no children = problem

# Stuck loading spinner:
progressbar[progressbar]        ← present after JS settles = API error

# Route loaded correctly:
router-outlet[element]
  app-order-list[element]:"Order Management..."   ← has content = good

Example

// After deploying an Angular app
navigate("http://your-app.vercel.app")
get_scene()                     // wait for hydration
→ check router-outlet has content, no broken img nodes

// Walk routes
act({ type: "clickSelector", selector: "a[href='/orders']" })
get_scene()
→ verify orders page loaded

act({ type: "clickSelector", selector: "a[href='/customers']" })
get_scene()
→ verify customers page loaded

What to Check

• Broken images — img nodes with "broken" content
• Empty routes — router-outlet with no child elements
• Stuck spinners — progressbar still present after get_scene()
• Placeholder text — undefined, null, TODO, <repo-url> in visible text
• Error pages — 404 or error component rendered instead of expected content

Development

pnpm test          # run tests
pnpm test:watch    # watch mode
pnpm lint          # lint
pnpm build         # compile TypeScript
pnpm mcp           # start MCP server (after build)
pnpm start:dev     # check services + start MCP server

Architecture

src/
├── config.ts           ← centralized config (dotenv)
├── types/              ← shared types (contracts between pipelines)
├── browser/            ← BrowserRuntime (Playwright)
├── pipelines/
│   ├── structural/     ← DOM + a11y tree extraction
│   ├── visual/
│   │   ├── index.ts        ← Three-tier orchestrator (detect/understand/deep)
│   │   ├── omniparser.ts   ← OmniParser V2 client (Tier A)
│   │   ├── claude-vision.ts ← Claude Vision API (Tier C)
│   │   ├── ollama-vision.ts ← Qwen3-VL via Ollama (Tier B)
│   │   └── frame-capture.ts ← CDP screencast + perceptual hashing
│   ├── fusion/         ← merge visual + structural → SceneGraph
│   ├── temporal/       ← change detection + state tracking
│   └── affordance/     ← predict interaction outcomes
├── mcp/                ← MCP server (12 tools)
└── utils/

Viewport default: 1280x720 (configurable via env) Screenshot format: PNG (lossless, required for vision models)