Back to search
frontend-toolkit
MarketplaceWorks with Claude CodeGitHub
Claude Code plugin: frontend code-review and browser UI-verification agents, plus a Stop hook that asks for browser evidence after UI edits.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/wiktorsmura--claude-frontend-toolkit)HTML
<a href="/listings/wiktorsmura--claude-frontend-toolkit"><img src="/api/listings/wiktorsmura--claude-frontend-toolkit/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
/plugin marketplace add WiktorSmura/claude-frontend-toolkit /plugin install frontend-toolkit@frontend-toolkit
Restart Claude Code, then check `/hooks` for a `Stop` entry from `frontend-toolkit`.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 13 lines
- {
- "name": "frontend-toolkit",
- "owner": {
- "name": "Wiktor Smura"
- },
- "plugins": [
- {
- "name": "frontend-toolkit",
- "source": "./",
- "description": "Frontend code-review and browser UI-verification agents, a Stop hook that asks for browser evidence after UI edits, and React frontend skills."
- }
- ]
- }
README.md · README · 171 lines
- <p align="center"><img src=".claude-plugin/icon.png" width="96" alt=""></p>
- # frontend-toolkit
- A [Claude Code](https://code.claude.com) plugin for React frontends. It gives Claude two read-only
- specialists, one for code review and one for UI verification in a real browser, plus a Stop hook
- that won't let a UI change end with "it compiles" as the only proof.
- | | What | Needs |
- | --- | --- | --- |
- | **Agent** | `frontend-reviewer`: severity-graded review of a frontend diff | nothing |
- | **Agent** | `ui-verifier`: checks the running app against a Figma node or named criteria | a browser MCP; Figma MCP for design-diff |
- | **Hook** | `Stop` gate: asks for a browser screenshot after UI files change | bash |
- | **Hook** | PR gate (opt-in): blocks `gh pr create` until `frontend-reviewer` has run on the branch | bash, git |
- | **Skills** | implementation, architecture, review, a11y, performance and verification guides | nothing |
- ## Install
- ```
- /plugin marketplace add WiktorSmura/claude-frontend-toolkit
- /plugin install frontend-toolkit@frontend-toolkit
- ```
- Restart Claude Code, then check `/hooks` for a `Stop` entry from `frontend-toolkit`.
- ## First use (30 seconds)
- On a branch with frontend changes, ask:
- > Use the frontend-toolkit:frontend-reviewer agent to review my changes against main.
- You get findings graded critical / warning / suggestion, each with `file:line` and a concrete fix,
- then APPROVE or REQUEST CHANGES. The reviewer runs your repo's typecheck, lint and test commands
- instead of re-reviewing what they already catch. It spends its attention on what tools can't see:
- logic in the wrong layer, server data copied into client state, missing loading and error states,
- effect misuse, accessibility, and drift from your documented conventions.
- ## The agents
- **`frontend-reviewer`** reads your `CLAUDE.md`, in-repo skills and neighbouring code before judging
- anything, and treats violations of *your* conventions as findings. It only reports problems it can
- tie to user impact or real maintenance cost, so it won't pad the list with taste.
- **`ui-verifier`** has two modes:
- - **design-diff**: exports the Figma node, captures the same state in the app, and compares
- screenshots *and* computed styles (±2px tolerance, exact token colours). Verdict PASS or FAIL.
- - **visual-check**: no design needed. It states the criteria first ("the dialog title doesn't wrap
- at 1280px"), then reports each one as MET or NOT MET with evidence.
- Every report ends with `Covered:` and `Not verified:` lines, so gaps are stated rather than hidden.
- The verifier never starts or kills dev servers and never handles credentials: if it hits a login
- wall, it asks you to log in. If your repo has its own browser or dev-server skill, the verifier
- lets that skill handle ports, login and feature flags.
- ```
- Use the frontend-toolkit:ui-verifier agent: design-diff of the settings dialog at
- http://localhost:5173 against <figma-node-url>.
- ```
- ## The Stop hook
- When Claude is about to finish a turn, the hook checks the session transcript, subagents included.
- If `.tsx/.jsx/.vue/.svelte/.astro/.css/.scss` files were edited after the last browser screenshot,
- it blocks the stop once with a request: verify in the browser, or say in one line why no user
- could see the change.
- - Test, spec and story files don't count, and neither do files under `__tests__/`.
- - Any tool with "screenshot" in its name or action counts as evidence: Claude in Chrome,
- Playwright, chrome-devtools, Puppeteer, preview tools, or a `screenshot` command run from the
- shell. Figma exports don't count, since they show the design, not the app.
- - It fires at most once per edit, so it can't trap Claude in a loop.
- - It stays off in React Native, Expo and Electron projects.
- - Turn it off with `FRONTEND_TOOLKIT_UI_GATE=off`, globally in `~/.claude/settings.json`
- (`"env": {"FRONTEND_TOOLKIT_UI_GATE": "off"}`) or for one repo in `.claude/settings.local.json`.
- ## The PR gate (opt-in)
- Set `FRONTEND_TOOLKIT_PR_GATE=on` (in `~/.claude/settings.json` under `"env"`, or per repo in
- `.claude/settings.local.json`) and `gh pr create` is blocked on any branch whose commits change UI
- files, until `frontend-reviewer` has run on that branch.
- - When the reviewer finishes, a `SubagentStop` hook writes a marker for the branch inside the git
- directory. The marker is per worktree, never committed, and survives later commits, so fixing
- the findings doesn't need a second review. It records that a review ran, not that it passed.
- - It follows a leading `cd <dir> &&` in the command. It's a reminder, not a lock: a
- `gh pr create` hidden inside a script isn't seen, and anything it can't work out (no git, no
- base branch) lets the command through.
- - To skip it once, put `FRONTEND_TOOLKIT_PR_GATE=off` in front of the command.
- ## The skills
- | Skill | Load it for |
- | --- | --- |
- | `using-frontend-toolkit` | Starting frontend work from a plan or orchestrator: stack detection, gates, what to put in subagent prompts |
- | `implementing-frontend` | Components, design tokens, Figma-to-code, state placement; `references/overlays-and-layering.md` for dialogs, popovers and overflow |
- | `frontend-architecture` | Where logic lives: components, hooks, commands, services, stores; the validated API boundary |
- | `reviewing-frontend` | Review criteria (preloaded by `frontend-reviewer`) |
- | `ensuring-accessibility` | WCAG 2.1 AA: keyboard, focus, ARIA, contrast |
- | `optimizing-frontend` | Slow renders, bundle size, Core Web Vitals, memory leaks |
- | `ui-verification` | Evidence contract and report format (preloaded by `ui-verifier`) |
- The skills are written for React 19 with Vite, Tailwind 4, shadcn/Radix, TanStack Query, Zustand
- and react-hook-form + Zod. They apply library-specific advice only when `package.json` shows the
- library is there; on another stack, they keep to the parts that don't depend on it.
- Name a skill in a prompt or plan task to load it reliably ("use
- `frontend-toolkit:implementing-frontend`"). Claude also picks them up from their descriptions, but
- much less reliably: see [Design notes](#design-notes).
- ## Is it working? Measure it
- ```
- python3 scripts/usage-report.py --since 2026-01-01
- ```
- Reads your Claude Code transcripts (`~/.claude/projects`), locally and read-only, and reports:
- - how many sessions edited UI files, and how many of those ended with a browser screenshot after
- the last edit (the same test the Stop gate applies);
- - how often each gate blocked;
- - which toolkit agents ran and which skills loaded, per project.
- Run it before installing to get a baseline, then a few weeks later. If the evidence rate and the
- agent runs don't move, the toolkit isn't earning its keep in your setup. Use `--json` for the raw
- numbers.
- ## Your repo wins
- Your `CLAUDE.md`, `AGENTS.md`, in-repo skills and existing code override every default here. The
- skills and agents look for them first and never "correct" a codebase toward the toolkit's
- preferences.
- ## Design notes
- This plugin used to route work to its skills with `SessionStart` and `SubagentStart` notes ("load
- skill X when doing Y"). Over a month of real use in a production React codebase, 86 sessions and
- 203 subagents received those notes, and **none** loaded a skill because of them. What did work was
- naming a skill in a prompt, dispatching an agent, and hooks that check results. So 1.0 drops the
- routing and keeps those three. Details are in [CHANGELOG.md](CHANGELOG.md).
- ## Development
- ```
- bash hooks/test-ui-evidence-gate.sh # Stop gate self-check
- bash hooks/test-pr-review-gate.sh # PR gate self-check
- python3 -m unittest discover -s scripts # usage report self-check
- claude plugin validate . # manifest and structure
- claude --plugin-dir . # try it without installing
- ```
- **Evals** (`evals/`, run with `claude plugin eval`) check that review requests reach
- `frontend-reviewer`, Figma questions reach `ui-verifier`, backend tasks touch neither, and that
- the Stop gate fires on a real edit. The whole suite costs about USD 0.25 per run and runs on your
- own Claude credential:
- ```
- claude plugin eval . --trust-plugin --no-publish --ablation none --scaffold \
- --allow-tools Edit --model claude-sonnet-5-5
- ```
- CI runs the self-checks and validation on every push, which is free on a public repo. It runs the
- evals only when the repo has an `ANTHROPIC_API_KEY` or `CLAUDE_CODE_OAUTH_TOKEN`
- (`claude setup-token`) secret.
- Works on macOS and Linux. The hook needs only bash, grep, sed and sort.
- ## License
- [MIT](LICENSE). Parts of the skills are adapted from
- [TheSoftwareHouse/copilot-collections](https://github.com/TheSoftwareHouse/copilot-collections) (MIT).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.