← MCP server and skill grades

ui-ux-pro-max-skill

nextlevelbuilder/ui-ux-pro-max-skill

C
Type
Plugin
Scanned
Nov 13, 2026
Version
6ae86f61f520
Files read
647

This project is a plugin and command-line tool for AI coding assistants that provides searchable guides for UI styles, colors, and fonts, along with helpers for creating design tokens, slides, banners, and logos. The biggest risk is that an untrusted design-token file can sneak runnable code into your generated Tailwind configuration, and despite claiming to work strictly offline, the default install bundles extra tools that call outside AI services, install unpinned packages, and edit their own scripts.

What it can reach

Files
Reads project files and brand guidelines, writes generated design docs, token files, Tailwind configs, HTML slides, and images into your workspace, and adds, removes, or edits AI skill files in your project or home folder.
Network
Although basic style search runs locally, the command-line tool checks GitHub for updates, and bundled design tools and starter configs connect to Google Gemini, Atlas Cloud, MuAPI, Pexels, Unsplash, npm, and web CDNs.
Shell commands
Runs bundled Python and Node.js scripts, uses npm and npx to update the tool, install UI components, or launch unpinned helper servers, and runs commands to start headless browsers.
Secrets
Reads environment variables and local .env files (including in ~/.claude) for API keys and tokens such as GEMINI_API_KEY, ATLASCLOUD_API_KEY, MUAPI_API_KEY, GITHUB_TOKEN, and UI_PRO_MAX_GITHUB_TOKEN.
Browser
Launches headless Chrome or Chromium (sometimes with browser sandboxing turned off), Playwright, Puppeteer, or Chrome DevTools tools to render pages, inspect layouts, and take screenshots.

Findings

1 high2 medium6 low
  1. HighJavaScript code injection in Tailwind token config generator

    .claude/skills/design-system/scripts/generate-tokens.cjs:159

    When run with --format tailwind, generate-tokens.cjs serializes the color map with JSON.stringify and then replaces every double quote with a single quote without escaping existing single quotes in token names. If a user or agent converts an untrusted tokens.json file containing a single quote and line comment in a color key, the generated JavaScript config contains live code outside string quotes that executes when Tailwind or Node loads the file.

    Fix: Keep the JSON.stringify output directly instead of replacing double quotes with single quotes, and validate token keys against a strict identifier pattern before generating code.

  2. MediumUnvalidated package specifier and component URLs passed to npx shadcn

    .claude/skills/ui-styling/scripts/shadcn_add.py:116

    shadcn_add.py reads the shadcn version field from the project package.json, strips only leading range symbols, and interpolates it into npx shadcn@{shadcn_version}. In an untrusted repository or under prompt injection, a crafted package specifier (such as an npm alias or tarball URL) or a remote URL passed as a component argument can cause npx or the shadcn CLI to fetch and execute attacker-chosen packages and code.

    Fix: Validate the version read from package.json against a strict semver regular expression and restrict component arguments to lowercase alphanumeric slugs.

  3. MediumUnescaped HTML interpolation in CIP presentation generator

    .claude/skills/design/scripts/cip/render-html.py:312

    Unlike generate-slide.py, render-html.py inserts the brand name, industry, and fallback image filenames directly into the generated HTML document without HTML escaping. If prompt injection supplies a crafted brand string or an untrusted image folder contains a crafted PNG filename, opening the generated presentation executes arbitrary JavaScript in the user's browser under the file:// origin.

    Fix: Escape brand_name, industry, deliverable metadata, and image paths with html.escape(..., quote=True) before embedding them in HTML.

  4. LowBundled design skill instructs the agent to edit failing skill scripts

    .claude/skills/design/SKILL.md:75

    The bundled design skill tells the AI agent to directly modify the skill's own Python scripts whenever a script fails. If triggered during a session involving untrusted content or in a global ~/.claude/skills installation, this instruction lowers the agent's barrier to rewriting executable scripts that persist across future sessions and projects.

    Fix: Remove the instruction to edit skill scripts at runtime and tell the agent to report script errors to the user instead.

  5. LowInstaller deploys network-capable sub-skills despite offline-only claim in README

    cli/src/utils/template.ts:402

    The main README states that the scripts use only the Python standard library, install nothing, and make no network calls, and describes brand, logo, CIP, slide, and banner tools as a separate tier. In practice, both the Claude plugin and uipro init install six additional sub-skills whose scripts read ~/.claude/.env secrets, call external AI APIs (Gemini, Atlas Cloud, MuAPI), and run npx shadcn.

    Fix: Document all seven bundled skills and their network, .env, and npx behaviors in the main README, or make installing the extra sub-skills opt-in via a CLI flag.

  6. LowOptional stack starter pre-approves reading /home/** and unprompted browser/audit execution

    stack/.claude/settings.json:9

    This starter directory is not active on normal plugin or CLI installs, but if a user clones the repo and runs Claude Code inside stack/, settings.json pre-approves Read(//home/**) (every file under /home on Linux, including SSH keys and cloud credentials), mcp__playwright, mcp__chrome-devtools, Bash(npx playwright:*), and Bash(node scripts/design-audit.mjs:*), which accepts --executable to launch any binary. Prompt injection from an audited webpage could read secrets from /home and exfiltrate them through the browser tools without confirmation, despite stack/README.md describing these as safe permissions.

    Fix: Remove Read(//home/**) so reads stay scoped to the project directory, require confirmation for browser navigation and Playwright CLI commands, and remove or restrict the --executable flag in design-audit.mjs.

  7. LowOptional stack starter auto-runs unpinned @latest MCP packages

    stack/.mcp.json:6

    Not used in a standard plugin or uipro init install, but when opening Claude Code in the optional stack/ starter directory, enableAllProjectMcpServers is paired with npx -y and @latest for @playwright/mcp, chrome-devtools-mcp, and shadcn, automatically downloading and executing the newest unpinned npm releases on startup.

    Fix: Pin exact verified package versions in stack/.mcp.json instead of @latest and let the user approve project MCP servers explicitly.

  8. LowSocial photos skill recommends running headless Chrome with --no-sandbox

    .claude/skills/design/references/social-photos-design.md:135

    In its recommended screenshot export command ('Option A: Chrome Headless CLI'), the skill instructs the agent to run Google Chrome with --no-sandbox on local file:// HTML documents while waiting 5 seconds for external assets such as Google Fonts or images to load. Disabling Chrome's OS process sandbox removes browser isolation protections if untrusted content or a compromised external resource triggers a renderer vulnerability, and is unnecessary on standard desktop macOS and Linux environments.

    Fix: Remove the --no-sandbox flag from the Chrome headless CLI example.

  9. LowSocial photo export guide runs headless Chrome with --no-sandbox

    cli/assets/skills/design/references/social-photos-design.md:135

    The recommended Chrome headless command in the social-photos workflow passes --no-sandbox when rendering local file:// HTML pages that load external fonts and images. On standard desktop systems Chrome's sandbox works in headless mode, and disabling it removes process isolation if rendered HTML or remote assets trigger a browser vulnerability.

    Fix: Remove --no-sandbox from the Chrome headless command example so browser sandboxing stays enabled.